The Journal on Information Technology in Healthcare

The Journal on Information Technology in Healthcare Editor Clyde Saldanha

Editorial Board Lodewijk Bos (Netherlands) Jimmy Chan (Hong Kong) Stephen Chu (New Zealand) Charles Doarn (USA) Li Felländer-Tsai (Sweden) Syed Haque (USA) Robert Istepanian (UK) Chien-Tsai Liu (Taiwan) Valentin Masero (Spain) Jeannette Murphy (UK) Dean Sittig (USA) Roger Tackley (UK)

HJ I T

Volume 2 Issue 6 2004 ISSN 1479-649X

The Journal on Information Technology in Healthcare Aims and Scope The Journal on Information Technology in Healthcare aims to improve the quality and safety of patient care, by encouraging and promoting the use of information technology (IT) in healthcare. The journal acts as a medium for the international exchange of knowledge and experience of the benefits of IT in healthcare. It publishes papers that educate healthcare professionals on the use of IT in clinical practice, and particularly papers that provide objective evidence of the benefits of IT in healthcare. Subscriptions The Journal on Information Technology in Healthcare is published 6 times a year. Subscription prices for online access only to 6 issues for 2005 and back issues of the Journal are: Sterling Euros US Dollars

Institutional £100 €180 US$180

Individual £70 €120 US$120

Organizations Up to 100 members 101–500 members 501–1,000 members 1,000+ members

£200 £300 £400 £500

or US$350 or US$500 or US$700 or US$1,000

Payment can be made by cheque or banker’s draft. Cheques should be made payable to The Journal on Information Technology in Healthcare. Subscription orders and requests for sample copies should be sent to: JITH, 72 Churston Drive, Morden, Surrey, SM4 4JQ, UK. E-mail: [email protected]. Fax : +44 (0)870 130 1572. Copyright © The Journal on Information Technology in Healthcare. All rights reserved. The Journal on Information Technology in Healthcare is protected by copyright. Apart from fair dealing for the purposes of research, private study, criticism or review, no part of this publication may be reproduced, stored or transmitted in any form or by any means without the prior written permission of the Editor. Disclaimer All papers are published in good faith. Authors are responsible for the scientific content and accuracy of their papers. Although every effort is made to ensure accuracy and avoid mistakes, no liability on the part of the Editor, publisher or their agents is accepted for the consequences of any inaccurate or misleading information. The opinions, data and statements that appear in articles published in the journal are those of the authors, and not necessarily those of the Editor or publisher. The Editor and publisher disclaim any responsibility or liability for such material and do not guarantee, warrant or endorse any product or service described in this publication. Publisher: Health Technology Press, 72 Churston Drive, Morden, Surrey, SM4 4JQ, UK. Advertising: To advertise in this journal, please contact William D’Sa, 25 Oxford Avenue, London, SW20 8LS, UK. Tel +44 (0)20 8543 1230. Fax +44 (0)870 130 1572. E-mail: [email protected] Typeset by Toby Matthews, Oxford, UK.

HTI J

The Journal on Information Technology in Healthcare Volume 2 Issue 6

CONTENTS Contents Electronic Prescription: Standards and Decision Support Issues Stephen Chu

385

The JUST VR Tool: An Innovative Approach to Training Personnel for Emergency Situations Using Virtual Reality Techniques Andreas Manganas, Manolis Tsiknakis, Erich Leisch, Michal Ponder*, Tom Molet†, Bruno Herbelin*, Nadia Magnenat-Thalmann†, Daniel Thalmann*, Marco Fato‡, Andrea Schenone‡

399

A Multiple Decision Tree-based Method for Differentiation of a Split First Heart Sound from a Fourth Heart Sound and Ejection Click Antonis Stasis, Sotiris Pavlopoulos, Euripides Loukis

413

Web-based Asthma Collaboration Management and Public Awareness Michael Glykas, Panagiotis Chytas

427

Power to the Patient, using DI@L-log Lesley-Ann Black, Michael McTear, Norman Black

439

HTI J

© The Journal on Information Technology in Healthcare 2004; 2(6): 385–397

Electronic Prescription: Standards and Decision Support Issues Stephen Chu Department of Management Science and Information Systems, University of Auckland, New Zealand.

ABSTRACT Electronic prescription (ePrescription) has been considered a highly effective technology to help reduce the incidence of medication errors. Despite this high expectation, many ePrescription efforts in recent years have encountered problems with implementation into clinical practice. This is reflected by a survey in 2003 that found only 9% of physicians in the USA were using computerised physician order entry systems (CPOE). This paper reports on experiences gained from development and testing of an ePrescription prototype template. In addition, it explores some important issues, such as messaging and medicine terminology standards, and decision support/interface design that can affect the acceptance and effects of ePrescription.

INTRODUCTION Public and payer pressures for improved patient safety and quality of care have been escalating rapidly since the publication in 2000 of the Institute of Medicine’s report, To Err is Human1. One area highlighted was medication errors. In the USA medication errors cost up to $5.6 billion each year2 and are a major cause of increased length of stay in hospital3–5. Errors occur throughout the entire medication prescription–dispensing–administration chain but the most common error types identified are prescription errors, with inappropriate drugs or inappropriate doses being prescribed6. Other common errors include the wrong drug or wrong dose being administered or a drug being omitted7. Computerised physician order entry (CPOE) of prescriptions, or electronic prescriptions (ePrescriptions) have been considered highly effective means for reducing the incidence of prescription errors and hence improving medication safety7–11. However, although CPOE has the potential to bring significant benefits to the healthcare sector and particularly patients, few hospitals have successfully implemented such systems. In 2003 only 9% of physicians in the USA were using Correspondence and reprint requests: Dr Stephen Chu, PhD, FACS, Associate Professor of Health Informatics, Department of Management Science and Information Sytems, University of Auckland, Private Bag 92-019, Auckland, New Zealand. E-mail: [email protected].

385

Chu CPOE12. Data from the (US) National Association of Chain Drug Stores revealed that 80% of patient visits to a doctor would have been given at least one prescription13. In New Zealand every day approximately 1,650 patients are admitted to hospital, 2,000 patients attend an emergency department, 50,000 patients see their general practitioner, and 105,000 prescriptions are filled14. Although most of the prescriptions from general practitioners are given to patients as computer printouts, none are transmitted electronically to the community pharmacies. This absence of electronic transmission also exists in acute care hospitals. Quality and economic pressures have forced many countries to begin piloting or implementing some form of electronic prescription systems. Examples include Australia15, the USA16, the United Kingdom10 and Taiwan7. Widespread adoption has been impeded by significant barriers17. These include the need for standards in medication terminology, electronic prescription messaging, robustness and usefulness of mobile connectivity, interface design, implementation and running costs, and electronic signatures. Addressing these issues adequately and promptly is essential to ensure acceptance of ePrescription and its benefits. This paper reports experiences gained from the design and trial of a prototype electronic prescription template. It discusses issues such as messaging and medicine terminology standards, interface and design. It also describes the provision of decision support, how this can be delivered, and how ePrescription can facilitate clinician workflow. ELECTRONIC PRESCRIPTION Electronic prescription (ePrescription) is the use of information and communication technologies to capture, review, edit and electronically transmit prescription information about pharmaceutical products by legally and professionally qualified and registered healthcare practitioners to registered pharmacies (or dispensing systems). Information transmitted includes demographic information about the patient, their health problems and/or diagnoses, prescription information (drug, dose, frequency, route of administration, duration of treatment, special instructions), and the prescriber details. The ePrescription legally authorises the receiving pharmacy (or dispensing system) to dispense the prescribed products identified in the prescription. The scope of prescribed products varies from country to country as permitted by the government authorities or health insurance carriers. It may encompass medicinal products, biologics, medical devices, medical consumables, (e.g. dressings or test strips), materials and/or services. The ePrescription chain starts with the authoring, authentication and transmission of the ePrescription and ends with the patient receiving the prescribed pharmaceutical products. For ePrescription systems to be readily accepted by clinical staff, they should provide additional benefits to the clinical decision-making and workflow processes. The ability of such systems to support point-of-care

386

The Journal on Information Technology in Healthcare 2004; 2(6): 385–397

Electronic Prescription: Standards and Decision Support Issues ordering by providing real-time feedback on appropriateness of an order based on accurate patient clinical information is fundamental. Provision of knowledge about the prescribed products, precautions and warning information such as drug–drug and drug–disease interactions, are also considered important value-added support services to ePrescription. THE ePRESCRIPTION SERVICE SCOPE MODEL Electronic prescriptions can vary significantly in scope and in technical complexity. In its simplest form it is the electronic transmission of electronic prescription messages from the clinician’s prescribing application to the pharmacy information system. More complex systems will include decision support and reporting functions. Even with the simplest functional design, an ePrescription system is expected to be able to effectively enable the following scenarios: • The prescriber may alter or cancel a prescription and specify special instructions for a particular medicinal product. • The dispenser may only dispense part of the prescription, and may replace the prescribed medicinal product with a suitable substitute (e.g. a branded medicine with a generic substitute), and can electronically inform the prescriber. • The prescription may allow limited repeat dispensing, enabling patients to request a refill of the prescribed products at a later date. In the ePrescription environment, the patient can initiate a repeat prescription in person, by telephone or via e-mail. In the case of a prescription with a repeat authorisation, provided that certain parameters such as the time since the last fill and the life span of the prescription (e.g. 3–6 months from first issue) have been met, the dispenser may provide the medication without the need to contact the prescribing physician. With the use of ePrescription systems, it would be possible for a repeat notification to go to the prescriber for their records. In the case of a repeat request that does not have a repeat authorisation, the dispenser could use electronic methods to communicate the request for additional refill authorisations to the original prescriber or the patient’s general practitioner. During 2002–3, one of the tertiary teaching hospitals in the Auckland region piloted an ePrescription system. The project tested electronic transmissions of prescription information from wired workstations to the hospital pharmacy. Guideline-based decision support and reporting functionality were not evaluated in the pilot. The transmissions were successful and more acute care hospitals are currently working on similar projects. There is also a plan to extend the project to support ePrescription transmission between general practices and community pharmacies. To help guide the discussion and scope of the project appropriate The Journal on Information Technology in Healthcare 2004; 2(6): 385–397

387

Chu

Figure 1. The ePrescription scope model

to stakeholder requirements, an ePrescription system scope model has been developed (Figure 1). This model is based on the lowest level of technical capability of all stakeholders likely to be involved in the ePrescription chain. It is currently modelled on face-toface interactions between the patient and the healthcare provider and also the use of wired workstations. The design can be extended to support Internet-based consultation and the use of wireless devices by prescribers to interact with a local ePrescription server. It also does not preclude the adoption of smart card and electronic kiosk technologies for authorisation to pharmacists by patients, and patient access of the ePrescription repository. The design mandates the use of a secured health intranet through the implementation of IPSec (secured Internet communication) protocol. One of the strong drivers for ePrescription implementation is the provision of relevant reports on resource (medication) utilisation, provider prescription behaviours and cost effectiveness. Such information is of particular importance to fundholders and payers. The hospital-based and central repositories and the ePrescription data warehouse will provide accurate source data for such reporting and knowledge discovery purposes. New Zealand is one of the early adopters of Health Level Seven (HL7) messaging standards. For electronic transmission of prescription information between prescribers, the ePrescription repository, pharmacies, and the Ministry of Health (or other Public health organisations), use of Version 2.x HL7 electronic prescription messaging standard is recommended. However, V3 and in particular clinical document architecture (CDA) are also possible candidates for supporting standard electronic transmissions. The following sections will present discussions on evolving ePrescription pilots and explore a number of key standard and technical issues related to the design

388


Electronic Prescription: Standards and Decision Support Issues and implementation of ePrescription applications, both in wired and wireless environments. THE ePRESCRIPTION TEMPLATE PROTOTYPE The ePrescription template is the clinicians’ interface to the electronic prescription system. The template design was based on an iterative design methodology and a development framework defined by Health Level Seven (HL7). The methodology involved the analysis of prescription workflow use cases (clinical scenarios) defined by domain experts. The iterative analysis by designer and review with clinicians allowed the identification of clinical concepts and medication data set required to be captured and presented to the clinicians during the prescription procedure. Of particular importance is the identification, through the use of case analysis, of problems related to representation and display of drug names. For example, compound drugs such as “Panadiene Forte” which contains “paracetamol” 500 mg and “codeine” 30 mg, or “Augmentin” which contains “Amoxicillin” 500 mg and “Clavulanic acid” 125 mg. In the paper system, it is not uncommon for the clinician to write a compound drug prescription as “Amoxicillin 500 mg” + “Clauvacinic Acid 125 mg”. In the electronic system, such practice would result in the prescription being treated by the system as two separate prescriptions, thus causing problems in recording and auditing of drug dispensing or administration. Representing the drug as a compound, e.g. “Augmentin 625 mg” does not satisfy all clinicians, especially when the dose needs to be adjusted as in the case of paediatric patients. As a quick fix for the pilot, it was decided that compound drugs were to be displayed to clinicians by the product names (as compounds). A commentary box would be provided to allow the prescribing clinician to document special requirements. Further work will be required to satisfactorily resolve this issue. The prototype template was designed as an independent add-on module to existing clinical information systems. Any future changes to the template will not require redesign of other clinical system components. Contextual linkage between the template and home applications could be maintained by the context manager. The context management technology was based on the context management standard defined by Health Level Seven’s Clinical Context Object Workgroup18. The context manager retrieves all the patient information required by the user and groups the information together for the patient (context). Users can switch between different patients, and switch between applications within the same patient without losing the context or the need to perform multiple logons. The clinician can launch the ePrescription template while reviewing a patient’s clinical note from within a ‘home’ application, such as a clinical information system or electronic healthcare record system. The context manager will automatically load the template with the patient’s clinical data which includes identification data and medication history. Depending on the ‘home’ application from which The Journal on Information Technology in Healthcare 2004; 2(6): 385–397

389

Chu

Figure 2. ePrescription application launched from acute myocardial infarction management clinical pathway Medication list contextually ‘constrained’ to those defined for limiting the infarct, reperfusing the myocardium, stabilising haemodynamics and treating associated symptoms.

the ePrescription template is launched, the medication list in the drop down box can be specifically tailored to suit the clinical context of the patient. For example, if the ‘home’ application that launches the ePrescription request is an electronic clinical pathway system, the medication list in the drop down box can be contextually ‘constrained’ to drugs relevant to the problem(s) on a specific clinical pathway. Figure 2 shows an ePrescription template launched from a myocardial infarction management clinical pathway. The drug list has been contextually constrained to contain medications that are judged by the rules set up for the context manager to be best suited for this clinical condition. The template also provides the mechanism for clinicians to override the system constraint. Selecting the ‘Show All Drugs” menu will result in a complete list of drugs commonly used by the clinical unit or hospital to be included in the drop down box. The clinician can also retrieve other contextually related information. For example, selecting the ‘Show Associated Order’ menu option will bring up related diagnostic tests such as serum CK-MB (Creatine Kinase-Myocardial Band) and serum troponin for patients suspected of having suffered a myocardial infarct. DECISION SUPPORT Some of the key user (clinician) requirements of ePrescription are the availability of decision support functions such as real-time access to contextually relevant and

390


Electronic Prescription: Standards and Decision Support Issues

Figure 3. Drug interaction alert screen with provision for clinician override The ‘Data Source’ link at the bottom of the screen provides access to a drug database for more detailed information about drug side effects and interactions.

evidence-based drug knowledge, and drug interaction alerts. Drug interaction alerts are perhaps one of the more difficult requirements to meet. Poor alert sensitivity and too many false alerts have been proven to be the root cause of clinician ‘alert fatigue’ resulting in computer-generated alerts being ignored and by-passed by the clinicians19. The majority of patients seeking medical and hospital care today suffer from chronic illnesses20,21 and take multiple medications. As a consequence the possibility of triggering a drug interaction alert with each newly prescribed drug is high. Integration of the ePrescription prototype with a drug knowledge-base was attempted with the aim of providing relevant triggers to generate important drug interaction alerts. To minimise the impact of ‘alert fatigue’, clinicians involved in developing the prototype were required to rank potential interactions. Four categories of alert were identified. • Trivial alerts were considered to be of little clinical significance. These did not produce a real-time alert, but batch reports were generated and sent to the prescribing clinicians and auditing system at the end of each day. • Minor alerts could be over-ridden by the prescribing clinician (Figure 3). • Moderately serious alerts could only be over-ridden by senior clinicians. Reasons for over-riding the alert had to be provided. • Alerts categorised as serious could not be over-ridden. RESULTS To determine the effects of the ePrescription prototype, the electronic template was evaluated against a set of criteria including: • The number of clarification checks required • Its potential to reduce drug interactions and allergies • User satisfaction The prototype was rated quite highly by clinical users on all three criteria. Clarification checks for drug names and ambiguous doses (e.g. 5 mg versus 8 mg; 10 ml versus 16 ml) due to poor handwriting were completely eliminated. Potential drug The Journal on Information Technology in Healthcare 2004; 2(6): 385–397

391

Chu allergy and interaction clarifications by pharmacists were also significantly reduced (by up to 50%). This result was comparable to international studies and reports. For example, a recent ePrescription benefits report by the Advanced Concepts USP22 and two studies at US and Canadian children hospitals6,23 showed that 79% of prescription errors were identified and prevented with CPOE, and there was an overall reduction in medication error rates of 40%. It also showed effectiveness in preventing known drug allergies (90% reduction) and drug interaction problems (60–75% reduction). The electronic application was considered by clinicians as functionally ‘satisfactory’ as it provided useful and ‘adjustable’ alerts, and useful links to a drug knowledge-base. These features increased the usefulness of the electronic system. Average satisfaction level on a 10 point scale was 7.35 (n = 20, standard deviation = 1.18, 95% confidence interval = ±0.52). As far as possible data fields of the electronic template were pre-populated with contextually related and relevant data at it was launched. When a drug was selected, the dose, frequency and unit of measurement data fields were automatically filled with data appropriate for the type and form of drug chosen. Users were able to override system default values provided that the override was within legitimate limits as determined by the drug knowledge-base or clinical unit rules. System usability rating by clinicians was also very high. Average usability rating on a 10 point scale was 7.05 (n = 20, standard deviation = 1.28, 95% confidence interval = ±0.60). The electronic template was tested in a small-scale pilot. Evaluations conducted during the trial were relatively simplistic. The impact of ePrescription on patient clinical outcomes were not assessed. Full technical capability of the prototype, such as system performance (response time), scalability, and reliability/fault tolerance were also not tested. System technical merit ratings are therefore not available. Medication errors can and often do result in costly adverse medical events. Adverse drug event studies have demonstrated that 56–79% of medication errors can be attributed to errors of prescription8. Illegible handwriting is well documented as a major cause of errors due to misinterpretation of drug names, doses and formulations. It is anticipated that the implementation of ePrescription applications should result in significant reductions in medication errors. Prescription errors also trigger clarification checks by nurses and pharmacists. A more extensive study on CPOE outcomes conducted over four 2-month periods in 1992, 1993, 1995 and 1997 at Brigham and Women’s Hospital indicated that although CPOE could decrease the potential for adverse drug events to occur (up to 86%), it increased the incidence of actual serious, preventable errors during the early phase of implementing such systems24. The increase in errors was attributed to “bugs” in early versions of the application and the errors were intercepted by humans and prevented. Such problems can cause loss of clinician confidence and lead to implementation failure. Mindful about the potential pitfalls of CPOE, the electronic ePrescription prototype will be subject to thorough analysis of functional design by

392


Electronic Prescription: Standards and Decision Support Issues the clinical stakeholders. Lessons learnt from international studies will be used to identify potential points of failure, improve the system design and to develop a set of comprehensive and reliability evaluation criteria and methodology. ePRESCRIPTION MESSAGING STANDARDS The ePrescription prototype was designed without adhering to any electronic messaging standards. The data stream consisted of text strings with the data fields separated by comma delimiters. Currently, work has begun to develop an ePrescription standard messaging based on the Health Level Seven V2.3.1 ePrescription messaging standard that was developed in 2001 by Standards Australia. This HL7 ePrescription messaging standard defined the message structure (data components and syntax) for a medication order (prescription) from the prescriber application and the acknowledgement/response message from the filler (pharmacy) application. The train diagram in Figure 4 shows the message segments for a complete electronic prescription order message. It has provision for carrying patient, provider, visit, and insurance, financial/billing information relevant to the prescription. It also allows the transmission of relevant clinical information (within the optional observation information and notes segments) in support of the prescription. This ePrescription messaging standard has been tested in the Australian MediConnect ePrescription pilot project implemented in Tasmania from September 2003. In this field test, an electronic prescription application was developed and integrated with general practitioner’s (GP) practice software. HL7 standard

Legend: MSH = Message header PID = Patient identification data PD1 = Provider identification data PV1 = Patient visit information PV2 = Additional patient visit information IN1 = Insurance data IN2; IN3 = Additional insurance data GT1 = Guarantor information

AL1 = Allergy information ORC = Common order information (grouper) RXO = Prescription order data RXR = Additional route information RXC = Drug component information (for compound drugs or intravenous fluids with additives)

OBX = Optional observation information (e.g. body weight, body surface area, etc) BLG = Billing information NTE = Notes

Figure 4. HL7 V2.3.1 prescription order message segments The Journal on Information Technology in Healthcare 2004; 2(6): 385–397

393

Chu compliant ePrescription messages were transmitted to dispensing software at the community pharmacy. The result has been reported to be highly satisfactory. The ePrescription messaging standard so far is only designed for direct pointto-point transmission, i.e. from GP directly to the pharmacy software. It lacks a query message standard for the pharmacy system to query the central electronic prescription repository. Until a query message standard has been developed and tested, the design illustrated in Figure 1 cannot be implemented. A query message standard is a key component of the ePrescription system as it also needs to be able to support queries on the central repository by doctors, pharmacists and even patients on their medication histories. MEDICINE TERMINOLOGY AND CENTRAL MEDICINE REPOSITORY For ePrescription and other electronic healthcare information systems to work properly and effectively, it is essential that clinicians have access to a trusted, authenticated central source of medicinal information. This is necessary to ensure access to and interchange of consistently reliable and standard information on medicinal product details among providers (prescriber, dispenser and administrator of medicine) and patients. Each medicine/product will have its own unique identification code (which is likely to be the GTIN (Global Trade Item Number) developed by EAN.UCC (European Article Numbering system International and the Uniform Code Council). It is desirable that each medicine can be identified (barcoded) to the item/dose level. Small, space-constrained items such as unit dose packages, tablets and ampoules will be identified by a new barcode technology, the Reduced Space Symbology (RSS)25. In addition, related products will be grouped under a virtual medicine which represents a unique combination of active ingredient(s), dosage, form(s) and strength(s). For example, all brands of all paracetamol 500 mg tablets will be grouped under the same virtual medicine. Each virtual medicine will be allocated its own unique internal GTIN. Currently, the central medicines data repository is not available in Australia and New Zealand. The two countries are collaborating in jointly developing a medicine terminology model and the central repository. The terminology model is likely to be based on a model developed by the UK National Health Services Information Authority and Blue Wave Informatics, a UK-based health informatics consultancy. CONCLUSION Electronic prescription systems have tremendous potential to significantly improve the quality of medication prescriptions and reduce medication errors. A substantial amount of work on the topic has been done in Australia and New

394


Electronic Prescription: Standards and Decision Support Issues Zealand. The ePrescription template and its small-scale testing has proved that such a concept and system can bring significant benefits and is well accepted by the user (clinician) community. However, despite the favourable user satisfaction ratings on system functionality and (interface) usability, the author acknowledges that the evaluations were simplistic and did not test the prototype against a number of important parameters such as impact on patient outcomes, reduction in medication errors and cost savings. Lessons learnt from international studies such as the Harvard project24 can provide extremely useful inputs to redesign the ePrescription system, and guide the development of thorough evaluation criteria and methodology. HL7 V2.3.1 ePrescription standard was used successfully in the Australian MediConnect trial in Tasmania. This standard will be used in New Zealand for ePrescription data transmission. However, an HL7-based standard query message needs to be developed to support queries of the central ePrescription repository by authorised pharmacists. The standard query message is necessary if the ePrescription system is to avoid the restrictive point-to-point transmission (directly between ePrescription application and pharmacy system) model and allow the effective implementation of an ePrescription repository model. International collaborative efforts are already underway to develop a standard drug terminology model and standards for a central medicine repository. These critical public health information infrastructures are important to ensure the quality use of medicines and to reduce the risk of medication errors. For ePrescription to provide the desired positive impacts on clinical workflow, mobile connectivity and use of hand-held devices such as personal digital assistants (PDAs) are considered critical success factors. However, the risks to security are serious with current wireless network technology26. Acceptance of this enabling technology requires the security and ePrescription interface design for PDAs to be adequately addressed. These issues will be explored in detail in another paper. REFERENCES 1 Kohn LT, Corrigan JM, Donaldson MS, eds. To Err is Human: Building a Safer Health System. Institute of Medicine, Washington DC: National Academy Press, 2000. 2 Denogean AT. System could prevent Rx mistakes. Tucson Citizen 2003(7 July). http://www. tucsoncitizen.com/local/7_7_03ua_hospital.html. 3 Audit Commission. A Spoonful of Sugar: Medicines Management in NHS Hospitals. Audit Commission, 2001. 4. Brennan TA, Leape LL, Lair NM. Incidents of adverse events and negligence in hospitalised patients: result of Harvard Medical practical study. New England Journal of Medicine 1991; 324: 370–76. 5. Malpass A. An analysis of Australian adverse drug events. Journal of Quality Clinical Practice 1999; 19: 27–20. The Journal on Information Technology in Healthcare 2004; 2(6): 385–397

395

Chu 6. CPOE Study. Vanderbilt Children’s Hospital in Nashville, Tennessee, 2003. www.iHealthbeat. org; 5 January 2004. 7. Chen CI, Liu CT, Chen CF, Li YC, Chao C. Medication errors in a hospital in Taiwan: incidence, aetiology and proposed solutions, The Journal on Information Technology in Healthcare, 2004; 2: 11–18. 8. Bates DW, Leape LL, Cullen DJ. Effects of computerised physician order entry and a team intervention on prevention of serious medication errors. Journal of American Medical Association 1998; 208: 1311–16. 9 Kaushal R, Shojana KG, Bates DW. Effects of computerized physician order entry and clinical decision support systems on medication safety. Archive of Internal Medicine 2003; 163: 1409–16. 10 Mitchell D, Usher J, Gray S et al. Evaluation and audit of a pilot prescribing and drug administration. The Journal on Information Technology in Healthcare 2004; 2: 19–29. 11 Walton RT, Harvey E, Dovey S, Freemantle N. Computerised advise on drug dosage to improve prescribing practice. Cochrane Database of Systematic Review 2001; 1: CD002894. 12 American Medical News. http://www.ihealthbeat.org. 13 National Association of Chain Drug Store (NACDS). The Chain Pharmacy Industry Profile, 2001. 14 Ministry of Health. Overview of Health and Disability Services. Wellington, New Zealand, 2001, p. 23. 15 The Australian Government Department of Health and Aging. The MediConnect Project. http://www.mediconnect.gov.au/. 16 ISO/TC215 – WG6 ePharmacy Final Report V1.9. International Standards Organization, December 2002. 17 Briggs B. CPOE: vendors see systems taking off. Health Data Management 2004(8 April): http://www.healthdatamanagement.com/html/current/CurrentIssueStory.cfm?Post ID=17310. 18 The CCOW Standard – Clinical Context Object Workgroup, Health Level Seven; http:// www.hl7.org/Special/committees/visual/visual.cfm; http://www.ccow-info.com/ 19 Glassman PA, Simon B, Belperio P, Lanto A. Improving recognition of drug interactions: benefits and barriers to using automated drug alerts. Medical Care 2002; 40: 1161–71. 20 Fries J, Friday GA, Gira C et al . Health project consortium. Reducing healthcare costs by reducing the need and demand for medical services. New England Journal of Medicine 1993; 329: 5–9. 21 Lorig K. Cost-effective self-management for chronic disease, Strategic Medicine, 1998; 2: 37–40. 22 Effects of new technology on formulary compliance and generic utilization. Report conducted for Aetna by Advanced Concepts USP. http://www.advancedconceptsusp.com/request_studyone.html. 23 King WJ, Paice N, Rangrej J, Swartz R. The effect of computerized physician order entry on medication errors and adverse drug events in pediatric inpatients. Pediatrics 2003; 112: 506–9. 24 Bates DW, Teich JM, Lee JM, Seger DR. The impact of computerized physician order entry on medication error prevention. Journal of American Medical Informatics Association 1999; 6: 313–21 25 Fitzgerald M. Using barcode technology to cut cost in healthcare. Proceedings, Combined Australian and New Zealand Health Informatics Conference. Sydney and Auckland, 2004.

396


Electronic Prescription: Standards and Decision Support Issues 26 Johnson BC. Wireless 802.11 LAN Security: Understanding the Key Issues. 2002: www.SystemExperts.com.

HTI J


397

HTI J


The JUST VR Tool: An Innovative Approach to Training Personnel for Emergency Situations Using Virtual Reality Techniques Andreas Manganas, Manolis Tsiknakis, Erich Leisch, Michal Ponder*, Tom Molet†, Bruno Herbelin*, Nadia Magnenat-Thalmann†, Daniel Thalmann*, Marco Fato‡, Andrea Schenone‡ Institute of Computer Science (ICS), Foundation for Research and Technology-Hellas (FORTH), Greece; * Virtual Reality Lab (VRlab), Swiss Federal Institute of Technology (EPFL); † MIRALab, University of Geneva, Switzerland; ‡ Department of Communication, Computer and System Sciences, University of Genoa, Italy.

ABSTRACT Objective: To develop a virtual reality (VR) tool to aid personnel in dealing with real medical emergencies. The tool is designed to help trainees improve their knowledge and skills in dealing with health emergency situations, and particularly to help them overcome their psychological barriers in a real emergency. Design: The JUST VR system consists of a hardware platform and a dedicated software application. The hardware platform comprises two separate but networked PCs, a set of high-end virtual reality devices and a stereo projection display. Each interactive medical emergency scenario simulation involves four main actors: the trainee, the virtual assistant, the virtual victim and the simulation supervisor. The simulation supervisor monitors and adjusts the simulation course whilst the trainee observes and makes decisions that are executed by the virtual assistant. Setting: The VR tool can be used in any suitably equipped room. Methods: The VR simulator was evaluated by twenty non-medical personnel in a specially adapted room in a hospital in Italy. Each participant dealt with one of two different medical scenarios and then completed a questionnaire. This was designed to assess their impression of their actual presence in the scenario and the degree of reality of the simulation. Results: Results from the questionnaire revealed that in general participants were convinced of their ‘feeling of presence’ in the scenario and the degree of reality of the simulation. Conclusion: The JUST VR Tool can realistically simulate medical scenarios that trainees can interact with. The system should help to overcome some of the weaknesses of present training methods by helping personnel overcome psychological barriers that may impair their decision-making process in real emergencies. The system also supports the traditional learning phase and should improve knowledge retention. The JUST VR Tool is a valuable tool for complementary training of personnel dealing with emergency situations. Its clinical benefits, however, remain to be established.

Correspondence and reprint requests: Andreas Manganas, Institute of Computer Science (ICS), Foundation for Research and Technology-Hellas (FORTH), STEP-C, Heraklion, Crete, Greece. E-mail: [email protected].

399

Manganas, Tsiknakis, Leisch et al INTRODUCTION It is a well known fact that in emergency medical situations, effective care and timely intervention by ‘on-the-field’ operators saves lives and greatly reduces the harmful effects of injury. In a number of cases these injuries can be severe and lead to permanent disabilities resulting in high social costs. Direct social costs stem from expenditure for medical treatment, rehabilitation and in some cases the need for home assistance. Indirect social costs can range from a few days absence from work to permanent loss of employment as a result of inability to perform previous duties. Emergencies are frequently initially dealt with by non-medically qualified people. It is important that these personnel are provided with proper training to enable them to recognise an emergency situation and deal with it in the correct manner. In most European countries, proper training of first aid providers is still a critical issue. In general, the training mainly addresses three issues: • Content (i.e. what to do in the case of an emergency situation) • Methodology (i.e. how to do it) • Improvement of attitude (i.e. ‘the lowering of psychological barriers’, in the presence of an emergency). With traditional learning methods, the psychological impact of accident scenarios (e.g. the distress of the victim or the presence of blood) on first aid providers is not tested in advance. As a consequence the performance of first aid providers may be impaired when confronted with a real emergency. To help improve the performance of non-medical and medical personnel in dealing with real emergencies we have designed and developed an advanced information technology (IT) verification tool based on virtual reality (VR). The system can be categorised as a situation-training tool with VR simulation scenarios built around an interactive and changing VR environment. This imposes on trainees a number of requirements such as the need for fast, precise assessment of the situation and the need for rapid decision-making. The coupling of simulation and VR has already proved to be extremely helpful in other domains such as aviation and also in other branches of medicine1. The system was tested in a pilot study carried out at the San Martino Hospital in Genoa. The primary aim of the study was to evaluate participants’ views on the reality of the experience with the VR Tool, i.e. to assess their ‘sense of presence’ in the scenario, and the reality of the simulation. THE JUST VR TOOL SYSTEM The JUST VR system has two main components: a hardware platform and a dedicated software application. The hardware platform consists of two separate but networked PCs plus a set of high-end virtual reality devices (such as a digital glove, a magnetic-based tracking system and a stereo projection display). One PC,

400


The JUST VR Tool known as the Simulation PC, is a high-end graphics computer that takes care of the communication with the VR devices and the execution of the simulation kernel. The other is a Simulation Controller PC that is dedicated to simulation control and provides all the necessary graphical user interfaces to the simulation supervisor. From the architectural point of view, the JUST VR system is built around the distributed application model. The software component of the VR Tool manages the real-time VR simulation, by using the scenario and the database. These were created by expert VR designers with relevant input from medical experts. Selection of the most appropriate components for the VR surrounding environment was based primarily on the following two considerations: • From a medical point of view, compliance with currently acceptable practices in emergency situations and adherence to current medical emergency guidelines. • From a commercial point of view, cost effectiveness and future potential for market penetration. With respect to the software, the main design principle has been of maximal modularity. This has been adopted to ensure flexibility and the potential for modifications and extensions in response to expected feedback from system testing. In effect the JUST VR system features a component based software architecture, where responsibilities and mutual interaction of software modules are clearly specified. The JUST VR system components (see Figure 4) may be grouped into the following two main categories: • The system kernel components responsible for the interactive real-time simulation initialisation and execution. • The interaction components driving external VR devices and providing various graphical user interfaces (GUIs) that allow interactive scenario authouring, triggering and control. The content to be created and used by the JUST VR system was specified and classified into main categories. These were static and dynamic content building blocks like models of three-dimensional (3D) scenes, virtual humans, objects, animations, behaviours, speech, sounds, etc. The interactive medical emergency scenarios were built from the content building blocks. The overall spatial configuration of the JUST VR system is depicted schematically in Figure 1. The simulation supervisor is situated behind the trainee so that he can observe both his simulation controlling GUI as well as the trainee and the simulation image projected on the big screen. The back projection option is employed so that in the course of the simulation, the trainee can approach the screen closely, without causing image projection occlusions as would occur with a front projection option. During the simulation, the trainee wears a right hand instrumented digital glove with the magnetic hand tracker attached to the wrist. This enables the exploration of the virtual scenes and the communication of simple and meaningful gestures. In addition the trainee wears wireless shutter glasses to enable stereThe Journal on Information Technology in Healthcare 2004; 2(6): 399–412

401

Manganas, Tsiknakis, Leisch et al

wires simulation supervisor

Simulation Controller PC

magnetic tracking space

3 m

magnetic field emitter

projection screen

wall projector

IR shutter glasses emitter

trainee head tracker

3m

cyber touch hand tracker

FOB Units

3 m

Cyber Glove Unit

Simulation PC

top view 3m

side vi ew

3m

IR shutter glasses emitter

projection screen

magnetic tracking space

simulation supervisor

Simulation Controller PC

head tracker

trainee

3m

cyber touch hand tracker

wall projector

magnetic field emitter

Figure 1. Schematic spatial configuration of JUST VR system

402

Figure 1


The JUST VR Tool

1 – Magnetic Field Emitter, 2 – Head Tracker, 3 – Navigation Ring , 4 – Sliding Margin, 5 – No Navigation Zone

Figure 2: VR System navigation paradigm oscopic viewing of the simulation images and a head tracker to allow navigation inside the virtual scenes. A principal objective of the JUST VR system is to create life-like stressful conditions that the participant experiences, i.e. to give the participant a ‘sense of presence’. To enhance the reality of the scenario, the VR simulation may generate stressful conditions by providing: • Surplus data, which the participant must quickly filter to select the relevant information. • Limited data, which forces the participant to quickly make extrapolations and working assumptions in order to select proper strategies. • Misleading information that can confuse the participant and cause him to lose his self-confidence in assessing the situation and making the right decisions. • Tight time constraints, which can vary from case to case. • Unexpected events, which can confuse the situation or distract the participant from following his plan of action. The JUST VR system is designed to allow both the trainee and the simulation supervisor to interact with it simultaneously. The trainee focuses on the simulation The Journal on Information Technology in Healthcare 2004; 2(6): 399–412

403

Manganas, Tsiknakis, Leisch et al course and the content itself. He interacts with the system in the most natural way by using high-end VR input and output devices, which include stereoscopic visualisation, motion tracking and tactile feedback. The trainee aims to: • Recognise and correctly assess a medical emergency situation • Show a certain level of competence and appropriate knowledge • Select and apply the appropriate medical procedures • Select and correctly use the specific medical equipment • Act within tight time constraints The simulation supervisor monitors and adjusts the simulation course by interacting with the system using a graphical user interface, which runs on the Simulation Controller PC. The simulation supervisor: • Selects the medical emergency scenario • Configures the scenario (if needed), e.g. by selecting the user profile, changing the difficulty level, introducing extra events, etc. • Initialises the scenario • Executes the scenario by having control over particular elements, such as the order of extra events • Monitors the simulation state through reception of feedback from the Simulation Engine • Reacts to voice commands of the trainee and converts them to the respective actions available in the course of the simulation. VR TOOL PROTOTYPE TRIALS AND VALIDATION The evaluation trials took place in Italy in a specially adapted VR environment as shown in Figures 3 and 4.

Figure 3. The virtual reality room

404


The JUST VR Tool

Figure 4. The VR Tool schema In order to evaluate the ‘sense of presence’ of the VR Tool, two scenarios were utilised – an office scene and a city scene. Both scenarios were based on the Basic Life Support (BLS) algorithm which in turn is based on guidelines, material and evaluation methodologies provided by the European Resuscitation Council. The BLS algorithm is graphically depicted in Figure 5. The algorithmic procedure has three basic objectives for trainees coping with an emergency situation. These are to recognise an emergency situation; to call for appropriate help; and to safely perform efficient cardiopulmonary resuscitation (CPR). The trials for the VR Tool were performed with twenty randomly selected trainees who attended the JUST Web/CD course on BLS2. The trainees were divided into two equal groups – one group assessed the office scenario and the other group evaluated the city scenario. It should be noted that all the students evaluated the JUST VR Tool just after completing their BLS course. In the office scenario the trainee is in a building one evening when he hears a noise (caused by a person falling to the ground) followed by a female voice calling for help. The trainee needs to navigate through the virtual environment to enter the office and locate the female. The trainee discovers a middle-aged gentleman, dressed in a business suit lying on the floor with a distressed young woman (Virtual Assistant – VA) leaning over him. The situation suggests that the man lying on the floor has experienced some kind of a sudden health problem, but the situation is not clear and needs to be assessed by the trainee. From this point onwards, the immersive VR decision training simulation begins following the interactive scenario’s multi-path tree as outlined in Figure 5. This continues until the arrival of the ambulance. The VA asks questions about what to do and the trainee needs to make correct decisions in a timely manner. If the trainee hesitates or delays making a decision, the VA prompts him to hurry up. In the event that the trainee fails to make a correct decision, the VA performs the relevant correct actions. This usually increases the trainee’s stress and embarrassment. The Journal on Information Technology in Healthcare 2004; 2(6): 399–412

405

Manganas, Tsiknakis, Leisch et al

Figure 5. Adult basic life support (BLS) procedural algorithm

Figure 6. Office Scenario The trainee and his virtual assistant have to assess and manage a victim who collapses in an office.

Figure 7. City Scenario The trainee and his virtual assistant have to assess and manage a victim they find collapsed in a park.

406


The JUST VR Tool The city scenario also takes place in the evening but is located in a municipal park. The trainee observes people passing by him and gathering around a middle-aged man lying on the ground. It is not clear from immediate inspection what is wrong with the man. It is possible that he is drunk, has taken a drug overdose, been attacked and beaten unconscious or even been the victim of a road traffic accident as the park is adjacent to a busy road. The victim is very similar to the office scenario, but in comparison to the office scenario, the city scenario is more challenging for the trainee because of the greater number of options that need to be considered. In addition the presence of people gathering around the victim

JUST Questionnaire Q1. Please rate your sense of being in the office space / public park, on the following scale from 1 to 7, where 7 represents your normal experience of being in a place. I had a sense of “being there” in the office space / public park Very much (7) … Not at all (1) Q2. To what extent were there times during the experience when the office space / public park was a reality for you? There were times during the experience when the office space / public park was a reality for me Almost all the time (7) … At no time (1) Q3. When you think back about your experience, do you think of the office space / public park more as images that you saw, or more as somewhere that you visited? The office space / public park seems to me to be more like … Somewhere that I visited (7) … Images that I saw (1) Q4. During the time of the experience, which was strongest on the whole, your sense of being in the office space / public park, or of being elsewhere? I had a stronger sense of … Being in the office space / public park (7) … Being elsewhere (1) Q5. Consider your memory of being in the office space / public park. How similar in terms of the structure of the memory is this to the structure of the memory of other places you have been today? By ‘structure of the memory’ consider things like the extent to which you have a visual memory of the office space / public park, whether that memory is in colour, the extent to which the memory seems vivid or realistic, its size, location in your imagination, the extent to which it is panoramic in your imagination, and other such structural elements. I think of the office space / public park as a place in a way similar to other place that I’ve been today Very much so (7) … Not at all (1) Q6. During the time of the experience, did you often think to yourself that you were actually in the office space / public park? During the experience I often thought that I was really standing in the office space / public park Very much so F(7) … 1. Not very often (1) i

Figure 8. JUST Questionnaire The Journal on Information Technology in Healthcare 2004; 2(6): 399–412

407

Manganas, Tsiknakis, Leisch et al tend to both distract the trainee’s attention and increase his stress. From this point onwards, the immersive VR decision training simulation begins following the interactive scenario’s multi-path tree as outlined in Figure 5. As with the office scenario, the VA asks the trainee what to do, prompts the trainee if he hesitates or delays making a decision, and performs actions in response to the trainees instructions or the correct action if the trainee makes a wrong decision. The scenario again ceases with the arrival of the ambulance. After they had completed the scenarios, the trainees were asked to answer a questionnaire (Figure 8). This was based on current virtual reality literature originally developed by Slater, Usoh and Steed (SUS)3, and was designed to assess the trainees impression of: (i) Their ‘sense of presence’ in the scenario (ii) The degree of reality of the simulation.

Figure 9. Rating by each trainee for each of the six questions (Office Scenario)

Figure 10. Rating by each trainee for each of the six questions (City Scenario)

408


The JUST VR Tool

Figure 11. Comparison for each one of the six questions) for the Office Scenario (JOF), the City Scenario (JCT) and the SUS reference results RESULTS The results of the questionnaires for each of the two scenarios are presented for each individual student in Figures 9 and 10. Figure 11 presents a summary of the results for both scenarios. As a comparison and for reference results are also given from the SUS paper3. The results demonstrated that in general participants did feel they were participating in real-life scenarios and were convinced of the reality of the simulation. The results were in fact better than those reported in the SUS paper3. This confirms that the VR Tool can create a “sense of presence”. DISCUSSION Currently the acquisition of knowledge in emergency medicine techniques is obtained via books, videos, multimedia CDs, the Internet and traditional classes. Training in psychomotor skills is currently provided through manikins, human actors, medical equipment training, and on-the-job training. However, although these training methods are undoubtedly valuable they do not necessarily enable people to cope with the decision-making process in real emergencies. In an emergency situation people operate under conditions of uncertainty, distractions, confusion and tight time constraints. They are expected to assess the situation, manage stress, use their knowledge under duress, take responsibility, and make decisions confidently and correctly. A key factor in performance is previous experience to overcome doubts and build self-confidence. The vision of the JUST project was to create a system that would increase the ability of non-medical personnel to deal with medical emergencies through improved training. The aims of the developed VR Tool are dual. It is designed to The Journal on Information Technology in Healthcare 2004; 2(6): 399–412

409

Manganas, Tsiknakis, Leisch et al help trainees improve their knowledge and skills in dealing with health emergency situations, and also to help them overcome their psychological barriers in a real emergency. It does this by using state-of-the art VR technology to enable a person to experience the same feeling as if they were physically present, i.e. to give them a ‘sense of presence’. The superiority of VR as a training method is dependent on this property, as the association with a life-like experience improves retention of knowledge. When designing and implementing the VR Tool, special care was taken to ensure that no special knowledge or expertise in virtual reality or information technology in general, was required of its users. Simulation supervisors are required to gain some knowledge for properly handling the VR devices of the system. They are also required to familiarise themselves with the system so that they can give appropriate instructions to the trainee and explain the proper usage of certain instruments, which are needed during the simulation. Finally they need to understand the overall concept and objectives of the simulation, so that they can explain them in advance to the trainee. Trainees only use voice and advanced VR devices when interacting with the system. The system can be used for training all personnel ranging from citizens with no medical background to physicians specialising in emergency medicine. Each interactive medical emergency scenario simulation involves four main actors: the trainee, the virtual assistant (VA), the virtual victim and the simulation supervisor. The role of the trainee is to observe and make decisions, while the role of the virtual assistant is to execute those decisions. The virtual assistant can also ask questions and act autonomously. The VA can: • Refuse to execute orders that would push the VR scenario into directions that are not desirable from a training point of view. • Encourage, suggest and prompt the user to take certain actions. This behaviour allows the trainer to guide the participant through the interactive scenario. • Make and execute decisions by themselves in the event of lack of interaction or cooperation from the trainee. The virtual assistant is necessary to interact with the virtual victim as with current technology it is not possible for the trainee to perform appropriate medical procedures on the virtual victim or operate medical equipment. The VR system only checks the ability of the trainees to adopt correct decision-making procedures within time critical constraints, and aids them to overcome the psychological barriers that occur in real-life situations. However, it should be appreciated that competency in performing relevant manoeuvres, e.g. opening the airway, is critical to successful CPR. These skills must be practised on manikins and trainees must be competent in performing them. The VR tool does, however, also assist in this process by enabling trainees to observe correctly executed procedures in semi-real conditions.

410


The JUST VR Tool The success of the VR Tool in creating a ‘sense of presence’ was confirmed by our small pilot trial with two different scenarios. Based on experience in other industries it is expected that it will greatly improve the training and ability of trainees to cope with real-life emergencies. The cost of the system will depend on the equipment selected and we estimate that it will vary between $20,000 and $80,000 depending on the degree of sophistication of the VR simulation and the audio equipment. CONCLUSIONS The JUST VR Tool system has a number of innovative characteristics that make it an important addition to current emergency training techniques. These can be summarised as follows: • This tool provides a new application area for VR technology. The benefits of this type of technology have already been established in other fields of medicine such as surgery1,4,5. • It provides a verification tool which concentrates mainly on the decisionmaking content and the appearance of injured virtual humans. • It provides a realistic alternative to training people by using synthesised movements of emergency procedures. The generic and strongly component-based character of the system gives it flexibility and extendibility. The component-based approach allows the system to deal with different requirements, without the need for changes to the underlying system architecture and implementation. The technology should soon be an affordable option for institutions and training centres. ACKNOWLEDGEMENTS JUST was a three-year European project that started in January 2000. Its total investment amounted to 3.6 million Euros and was co-funded by the Commission of the European Community within the framework of the Information Society Technologies (IST) programme (EU Contract number: IST-1999-12581). The JUST project involved fourteen partners from seven European countries, in a multidisciplinary partnership with participants from industrial, academic, research and health sectors. Thanks are due to all the people from the following organisations that contributed to the JUST project during its three year period, namely the Foundation for Research and Technology – Hellas (FORTH), Greece; TSD Projects s.r.l. (TSD), Italy; the Dipartimento di Informatica, Sistemistica e Telematica dell’Universita degli Studi di Genova (DIST), Italy; Interactive Labs s.r.l. (GIUNTI Ilabs), the Danish Centre for Health Telematics, County of Funen (FUNEN), Denmark; the Assistance Publique Hopitaux De Paris SAMU 92 (SAMU), France; the Sociedad Andaluza para el desarrollo de la informática y la electrónica SADIEL S.A. (SADIEL), Spain; University of Geneva (UNIGE), Switzerland; the Universidad Politécnica de Madrid (UPM), Spain; the Ecole Polytechnique Fédérale de Lausanne (EPFL), Switzerland; the European Resuscitation Council (ERC), Belgium; the University of Crete (UCH), Greece; the Azienda The Journal on Information Technology in Healthcare 2004; 2(6): 399–412

411

Manganas, Tsiknakis, Leisch et al Ospedaliera “Ospedale S. Martino” e Cliniche Universitarie Convenzionate (S.Martino), Italy; the Assistance Publique – Hopitaux de Paris (AP-HP), France; and the Empresa Publica de Emergencias Sanitarias (EPES), Spain.

REFERENCES 1 Saldanha C. Use of simulators and virtual reality to improve patient outcomes. The Journal on Information Technology in Healthcare 2004; 2: 149–54. 2 Manganas A, Tsiknakis M, Leisch E, et al. “JUST in time health emergency interventions: an innovative approach to training the citizen for emergency situations using virtual reality techniques and advanced IT tools (The Web-CD)”. In: Medical and Care Compunetics 1. L Bos, S Laxminarayan, A Marsh (eds). Studies in Technology and Informatics, 2004; 103: 315–26, IOS Press 3 Usoh, M, Catena E, Arman S, Slater M. Using presence questionnaires in reality. Presence Journal 2000; 9: 497–503. 4 Mabrey JD, Gillogly SD, Kasser JR , et al. Virtual reality simulation of arthroscopy of the knee. Arthroscopy 2002; 18: E28. 5 Jacomoides L, Organ K , Cadeddu JA, Pearle MS. Use of a virtual reality simulator for ureteroscopy training. J Urol 2004; 171: 320–23.

412

HTI J


HTI J


A Multiple Decision Tree-based Method for Differentiation of a Split First Heart Sound from a Fourth Heart Sound and Ejection Click Antonis Stasis, Sotiris Pavlopoulos, Euripides Loukis* School of Electrical and Computer Engineering, National Technical University of Athens, * University of Aegean, Dept. of Information and Communication Systems Engineering, Samos, Greece.

ABSTRACT Objective: Differentiating a fourth heart sound (S4), from a split first heart sound (SP1), or ejection click (EC), is often difficult particularly for inexperienced clinicians. The objective of this study was to develop and evaluate a computer-assisted classification tool to aid in this difficult differentiation problem, and in general for heart sound differentiation and diagnosis. Design: Developmental study. Methods: Emphasis was given to the selection of appropriate features that are adequately independent from the heart sound signal acquisition method. Relevance analysis was initially performed to identify the features of the heart sound most relevant to aiding diagnosis of S4, SP1 and EC. To detect and differentiate S4, SP1 and EC, a detection decision tree (DeDT) and a differentiation decision tree (DiDT) were used independently and also together in a multiple decision tree architecture. The DeDT provides three suggestions for each heart sound pattern, whereas the DiDT provides one. The MuDT analyses the suggestions of both decision trees to provide one final suggestion for each sound pattern. Results: Relevance analysis on the different heart sound features demonstrated that the most relevant features for aiding diagnosis of S4, SP1 and EC are the frequency features and the morphological features that describe S1. The DeDT architecture demonstrated an average classification accuracy of 80.56%, sensitivity of 70.93%, and specificity of 83.42%, but provided more than one suggestion for many cases. The DiDT architecture demonstrated an average classification accuracy of 66.46%, a sensitivity of 66.15% and a specificity of 82.15%, and only provided one suggestion for each case. The MuDT architecture slightly improved performance compared to the DiDT architecture. Average classification accuracy was improved by 2.79%, classification sensitivity by 2.73% and classification specificity by 1.26% Conclusions: The present work has demonstrated that decision tree algorithms can be successfully used as the basis for a decision support system to assist inexperienced clinicians in heart sound diagnosis. Further work is currently in progress to improve the accuracy, specificity and sensitivity of the system.

Correspondence and reprint requests: Dr. Sotiris Pavlopoulos, Biomedical Engineering Laboratory, School of Electrical and Computer Engineering, National Technical University of Athens, Zografou Campus, Athens 15773, Greece. E-mail: [email protected].

413

Stasis, Pavlopoulos & Loukis INTRODUCTION Auscultation of the heart is a cheap screening method for cardiac pathology and is performed as part of the routine clinical examination. Experienced physicians can often diagnose cardiac pathology on the basis of auscultation alone. Its importance as a diagnostic tool has, however, declined as echocardiography is routinely used today to investigate patients with suspected cardiac pathology. Echocardiography can provide both anatomical and physiological information and its value in aiding accurate diagnoses is well-established. However, it should be remembered that the initial request for an echocardiogram or a cardiologist consultation is usually based on the initial auscultatory findings. This is frequently performed by the patient’s general practitioner who may not be experienced or confident in cardiac auscultation1. For these physicians a decision support system to assist them in diagnosing different heart sounds and helping them differentiate similar heart sounds would be helpful2. Such a system would be based on acquiring and codifying the relevant knowledge of experienced cardiologists and making it available to them. In the past computer-assisted heart sound diagnosis has been treated as a classification problem. Classification algorithms were mainly based on: i) Discriminant analysis3 ii) Nearest neighbour4 iii) Bayesian networks5 iv) Neural networks6,7 v) Rule-based methods2,8,9 These different approaches have been necessary because heart sounds have more than one characteristic morphology, e.g. timing in the cardiac cycle, duration and character of murmurs, and different pathologies, e.g. aortic stenosis and mitral regurgitation, can produce similar heart sounds. A normal heart sound consists of four components. These are the first heart sound (S1), the systolic phase, the second heart sound (S2) and the diastolic phase. Additional sounds, such as murmurs or click-like sounds are heard in patients with a variety of heart diseases. Normally the mitral and tricuspid valves close simultaneously and are heard as a single first heart sound. If for any reason closure of the tricuspid valve is delayed, the two components of the first heart sound will be heard separately and this is referred to as a split first heart sound (SP1). Delayed closure of the tricuspid valve may, for example, occur with right bundle branch block (delayed contraction of the right ventricle) or an atrial septal defect (increased blood flow through the right ventricle)10. Ejection clicks are often heard shortly after S1. They are often caused by valve abnormalities, e.g. aortic or pulmonary stenosis10. The fourth heart sound (S4) is a click-like sound that is heard at the end of the diastole, just before S1. S4 is thought to be due to forceful atrial contraction and occurs in conditions when

414


A Multiple Decision Tree-based Method

Figure 1. Timing of split first heart sound (SP1), ejection click (EC), fourth heart sound (S4), diastolic murmur (DM) and systolic murmur (SM) in the heart cycle ventricular compliance is impaired, e.g. ventricular hypertrophy or fibrosis11. S4 is not heard in normal subjects. Figure 1 shows the timing of S4, SP1 and EC in relation to the cardiac cycle, and Table 1 lists some of the common condition in which they occur. It should be noted that in many of these cases the patient is asymptomatic and abnormal cardiac auscultation is an incidental finding. However, detection of this abnormal heart sound is important to ensure early diagnosis and optimal management of

Table 1. Clinical conditions in which SP1, EC and S4 can be heard Clinical conditions with a split first heart sound (SP1) Atrial septal defect Right bundle branch block Left ventricular ectopics Tricuspid stenosis Coarctation of the aorta Normal (i.e no cardiac pathology)

Clinical conditions with an ejection click (EC) Aortic Stenosis Bicuspid aortic valve Aortic regurgitation Pulmonary stenosis Eisenmenger’s syndrome Pulmonary hypertension

Clinical conditions with a 4th heart sound (S4) Aortic stenosis Severe systemic hypertension Pulmonary hypertension Hypertrophic cardiomyopathy Ventyricular hypertrophy Ventricular fibrosis Myocardial ischaemia Myocardial infarction


415

Stasis, Pavlopoulos & Loukis Heart sound feature vector classification module

Heart Sound Signal

Heart sound feature extraction module

Heart sound feature vector classification module

Heart sound feature vector classification module

Partial diagnosis Sub-system 1 Partial diagnosis Sub-system 2

Arbitration module. The partial Final diagnoses diagnosis are combined to the final diagnosis

Partial diagnosis Sub-system N

Figure 2. Integrated decision support system architecture for heart sound diagnosis9

abnormal cardiac conditions, e.g. antibiotic prophylaxis for dental procedures in a patient with a bicuspid aortic valve. From Figure 1 the difficulty in clearly differentiating sounds heard around S1 can be appreciated. In this work we propose a method that uses time-frequency features and decision tree classifiers for addressing this problem. We have attempted to develop a computer-based assisted system for analysing the morphological characteristics of heart sounds and in particular for detecting and differentiating a fourth heart sound (S4), from a split first heart sound (SP1) or ejection click (EC). The approach adopted has been to divide heart sound diagnosis into a number of simpler sub-problems, each of them dealing either with a morphological characteristic of the heart sound signal, e.g. timing of murmur, or frequency (tone) of the murmur9. Each of these sub-problems is dealt with using a method or algorithm which is most appropriate to analysing it, e.g. decision trees or neural networks. An arbitration module then processes and combines the partial diagnoses of these specialised sub-systems, to make a final diagnosis. All the above specialised sub-systems and the arbitration module incorporate and are based on expert knowledge. Their combination can lead to an integrated decision support system architecture for heart sound diagnosis, as shown in Figure 2. METHODS Preprocessing of Heart Sound Signals The characteristics of the heart sound signal are significantly affected by factors related to the signal acquisition and preprocessing method. Therefore, a heart sound diagnosis algorithm should be tested in heart sound signals from different sources and recorded with different acquisition methods for objective evaluation. For this purpose we collected heart sound signals from nine different heart sound

416


A Multiple Decision Tree-based Method sources (see Appendix) and created a “global” heart sound database. Because these sources were intended for training purposes they included heart sounds representative of all heart diseases. These heart sounds files had already been diagnosed and linked to a specific heart disease; therefore, they incorporate the knowledge of numerous experts in this area. From the available heart sound signals of this database we chose the ones containing either S4, SP1 or EC. This resulted in a total of 100 heart sound signal files. Each of these heart sound signal files was initially pre-processed and then converted to the corresponding heart sound feature vector, following a previously described method2,9. In particular, the pre-processing comprised of an initial normalisation of each signal in order to account for the amplitude variations among the signals due to different acquisition and recording methods. A set of six processing stages were then performed to identify S1 and S2 and their boundaries.

Nase Envelope

S1 & S2 boundaries

Segmentation into the structural components

Calculation of Feature Vectors Following the preprocessing tasks, the corresponding feature vector was calculated for each heart sound signal. The selection of the features was based on the technique used by experienced clinicians for analysing heart sounds to make a diagnosis or differential diagnoses. These include features such as the timing of the additional noise in the cardiac cycle (i.e. whether it occurs in diastole or systole) and other characteristics of the sound e.g. its duration, and the frequency of its tone. For these

First heart sounds

Time-warping and aligning

2nd Mean Heart Sound Cycle Calculation of Mean Division into Value 8 equal parts

Calculation of Mean square value

F1 .... F8

Dystolic phases

Division into 24 equal parts

F9 .... F32

Second heart sounds


F33 .... F40

Diastolic phases


F41 .... F88

Figure 3. Calculation of the 88 morphological features (F1–F88)

Heart Sound Signal

S1 & S2 boundaries

Segmentation into the structural components

1st Mean Heart Sound Cycle

First heart sounds Dystolic phases Second heart sounds Diastolic phases

Time-warping and aligning

Calculation of Mean Value

Bandpass filters 50–250 Hz 100–300 Hz

Calculation of Total energy

F89 F90

150–350 Hz

F91

200–400 Hz

F92

50–250 Hz

F93

100–300 Hz

F94

150–350 Hz

F95

200–400 Hz

F96

Figure 4. Calculation of the 8 frequency features (F89–F96) The Journal on Information Technology in Healthcare 2004; 2(6): 413–426

417

Stasis, Pavlopoulos & Loukis reasons we decided to use a set of time domain morphological features that cover and describe the whole heart cycle, in combination with a set of frequency domain features concerning the energy of the systolic and the diastolic phase in four significant frequency zones. These steps divide the signal into 88 morphological features and 8 frequency features as shown in Figures 3 and 4 Following the above described procedures, every heart sound signal was transformed into a feature vector (pattern) with dimensions of 1 × 96. The feature vectors of the initial 100 heart sound signals were stored in a database table with 100 records and 98 attributes-fields: one attribute named ID is the pattern identification code, one attribute named S4_SP1_EC is the characterisation (diagnosis) of the corresponding heart sound signal as having S4, SP1 or EC, while the remaining 96 attributes are the above 96 heart sound features (F1–F96). Analysing the Heart Sounds Before constructing and utilising the decision tree classifiers, we used relevance analysis12,13, to find the most suitable and relevant features for this classification problem. For this purpose, we used the value of the uncertainty coefficient12,13 of each of the above 96 features, which are the independent variables, for ranking them according to their relevance to the classifying attribute (S4_SP1_EC), which is the dependent variable. The calculation of the uncertainty coefficient of an independent variable regarding the dependent variable consists of a number of steps which gives a value between 0% and 100%. A low value (near 0%) of the uncertainty coefficient of an independent variable means that if we use this variable for partitioning the initial set of heart sounds there will be only a low increase in homogeneity regarding the dependent variable (and, therefore, low increase in classification rules accuracy), and, therefore, the relevance between this variable and the dependent variable is low. On the contrary a high value (near 100%) of the uncertainty coefficient of an independent variable indicates a high relevance with the dependent variable. In order to examine the relevance and the contribution to the differentiation of S4, SP1 and EC of each of the above mentioned 96 heart sound features, the uncertainty coefficients were calculated for each of them considering the S4_SP1_EC field as the classifying attribute – dependent variable. Decision Tree Classifiers A decision tree is a classification tree for classifying new instances (e.g. new heart sound feature vectors) into one of the categories of an important target attributedependent variable based on a number of other attributes constituting the independent variables13–15. To construct a decision tree we used a training data set of instances, for which we had the values of both the attributes that constitute the independent variables and the targeted attribute that constitutes the dependent attribute. We determined the best test (= attribute + condition) for splitting

418


A Multiple Decision Tree-based Method the training data set which creates the most homogeneous subsets regarding the dependent variable and, therefore, gives the highest classification accuracy. We used three different types of decision trees to analyse the heart sounds. Detection Decision Tree (DeDT) Architecture. The detection decision tree (DeDT) architecture treats the problem of differentiation of S4, SP1 and EC as three two-category classification sub-problems: existence of S4 or not, existence of SP1 or not and existence of EC or not. Each of these simpler twocategory sub-problems is handled by a separate decision tree, which aims to detect whether the corresponding morphological characteristic exists or not in the examined heart sound. Differentiation Decision Tree (DiDT) Architecture. A differentiation decision tree (DiDT) treats the problem of differentiation of S4, SP1 and EC as a three-category classification problem, i.e. it classifies a feature vector-pattern as having either S4, SP1 or EC. Multiple Decision Tree (MuDT) Architecture. The multiple decision tree architecture combines the DiDT and DeDT architectures to exploit the advantages of both. The suggestions made by these two decision trees are analysed by an arbitration module that makes the final decision on which of these suggestions should be accepted and which of them should be rejected (see Figures 2 and 7). In order to examine the generalisation capabilities of the constructed decision tree structures, the available feature vectors-patterns set was divided in two subsets. The first subset included 60% of the records of each class of the heart sound patterns set (S4, SP1 and EC classes), which were randomly selected and were used as the training set. The other subset consisted of the remaining patterns (40% of the records of each class) and were used as the test set. In this way the first training test (60%a–40%a) set scheme was formed. For the second scheme (60%b–40%b) the same proportions (60% training set–40% test set), were kept but random different patterns were selected for the training set. In the same way two more schemes were created (70%a–30%a and 70%b–30%b) with a different proportion (70% training set–30% test set). For the DiDT architecture, classification accuracy was calculated as the ratio of the number of the correctly classified patterns to the total number of patterns of the test data set. For the DeDT architecture, classification accuracy was calculated using the following equation: Accuracy _ X =

correctly classified (X sounds + non X sounds) tested (X sounds + non X sounds )

where X stands for either S4, SP1 or EC For each of the three morphological characteristics (S4, SP1, EC) classification sensitivity and classification specificity were also calculated. The classification The Journal on Information Technology in Healthcare 2004; 2(6): 413–426

419

Stasis, Pavlopoulos & Loukis sensitivity for a morphological characteristic is defined as the ratio of the number of the patterns correctly classified as having this morphological characteristic to the total number of patterns having this morphological characteristic of the test data set. Similarly, the classification specificity for a morphological characteristic is defined as the ratio of the patterns correctly classified as not having this morphological characteristic to the total number of patterns not having this morphological characteristic of the test data set. RESULTS For each heart sound feature we calculated the uncertainty coefficient separately from the training data set of each of the above four data schemes. Then, based on these four values, we calculated the average value and the standard deviation of the uncertainty coefficients. The average values and the standard deviations of the uncertainty coefficient for the most important features are shown in Figure 5. The graph demonstrates that the most relevant features of the classifying attribute S4_SP1_EC are the frequency features, i.e. high frequency energy and medium frequency energy in the diastolic and systolic phases, and also the morphological features that describe the first heart sound. These results are compatible with our physical understanding of the problem that S4, SP1 and EC click-like sounds appear almost simultaneously with S1. Also each of these click-like sounds is usually related to specific heart diseases that have heart sound murmurs in the systolic and the diastolic phase. The standard deviation values are generally smaller than 7%, showing that the uncertainty coefficients calculated from each scheme separately, especially the ones of the most relevant features, are similar and consistent.

STANDARD DEVIATION AVERAGE

Uncertainty Coefficient

100 90 80 70 60 50 40 30 20 10 s_ 6

s_ 8 sy

_2

s_ 4

sy

sy

7

s_ 1

sy s

sy

_5

_3

_1

s1 _

s1

s1

s1

3 di as _4 5 di as _4 7 E_ di as _h E_ f di as _m f E_ sy s_ hf E_ sy s_ m f

1

_4 as di

9

_4 as

di

7

di as _3

_3 as

as di

di

_3

5

0

Attributes-Features

Figure 5. Average values and standard deviations of the uncertainty coefficient for the most important features regarding S4_SP1_EC as the classifying attribute

420


A Multiple Decision Tree-based Method

100% 90%

Accuracy

80%

Sensitivity

70%

Specificity

60% 50% 40% 30% 20% 10% 0% S4

SP1

EC

AVER

Figure 6. Classification accuracy, sensitivity and specificity results for the detection decision tree architecture (DeDT) Table 2. Classification accuracy, sensitivity and specificity results for the differentiation decision tree (DiDT) Test data set Classification accuracy Classification sensitivity Classification specificity

40%a

40%b

30%a

30%b

Average

67.50% 65.34% 83.56%

65.00% 68.29% 79.83%

70.00% 71.54% 83.72%

63.33% 59.42% 81.49%

66.46% 66.15% 82.15%

Figure 6 gives the results obtained with the DeDT architecture. This shows an average classification accuracy of 80.56%, an average classification sensitivity of 70.93% and average classification specificity of 83.42%. The major drawback with the DeDT architecture is that for many cases it gives more than one suggestion. Table 2 shows the results for classification accuracy, sensitivity and specificity for each of the four datasets using the differentiation decision tree (DiDT) architecture. The average accuracy and sensitivity was 66% and the specificity 82%. It should be mentioned that all the calculated classification sensitivity and specificity values presented in this paper are based on a lower number of patterns than the corresponding classification accuracy values (i.e. while the classification accuracy is based on all the patterns of the test data set, the classification sensitivity is based only on the ones having the specific morphological characteristic, and the classification specificity is based only on the ones not having the morphological characteristic). It should also be mentioned that the classification accuracy for the training data set was 100% for all the examined cases with both the DiDT and the DeDT. The multiple decision tree architecture combines the DiDT and DeDT architectures and exploits the advantages of both (Figure 7). Because the DeDT architecture The Journal on Information Technology in Healthcare 2004; 2(6): 413–426

421

Stasis, Pavlopoulos & Loukis

DiDT S4 DeDT SP1 DeDT EC DeDT

Suggestion 2 Suggestion 3 Suggestion 4

Arbitration Module

Heart Sound Pattern

Candidate: Suggestion 1 Final Suggestion

Figure 7. Multiple decision tree architecture Table 3. Classification accuracy, sensitivity and specificity results for the multiple decision tree architecture Test data set

40% a

40% b

30% a

30% b

Average

Classification accuracy Classification sensitivity Classification specificity

72.50% 69.90% 81.47%

68.50% 69.42% 85.43%

70.00% 71.54% 83.72%

66.50% 64.33% 83.00%

69.25% 68.80% 83.41%

Comparison between Average Classification Accuracy of the Differentiation Decision Tree Architecture and the Multiple Decision Tree Architecture 80%

Accuracy

75% 70%

Multiple DT DiDT

65% 60% 55% 50% 60%a

60%b

Average 70%a 60% Data Schemes

70%b

Average 70%

Figure 8. Comparison between DiDT architecture and MuDT architecture

has better classification performance for this problem than the DiDT architecture, the arbitration module can be based on the following rule: “If only one DeDT detects its corresponding morphological characteristic, then the final suggestion is the one of this DeDT, otherwise the final suggestion is the one of the DiDT”.

422


A Multiple Decision Tree-based Method The results concerning the classification performance achieved using the MuDT architecture, in combination with the above arbitration rule, are shown in Table 3 for all four data schemes, while a comparison with the results of the DiDT architecture are shown in Figure 8. We can see that the MuDT architecture results in improvement of the classification performance in comparison with the DiDT with an increase of 2.79% in classification accuracy, 2.73% in classification sensitivity and 1.26% in classification specificity. DISCUSSION In this paper we investigated the use of decision trees for the differentiation of S4, SP1 and EC, which is a difficult and challenging problem in cardiac auscultation. In this direction several decision tree structures and architectures have been constructed and evaluated as to their classification accuracy, sensitivity and specificity for diagnosing these different heart sounds. We chose to use decision trees classification algorithms because the knowledge representation model that they produce is compatible with the practices followed by clinicians in making differential diagnoses. Decision trees do not work as a ‘black box’ but offer a full justification for their suggestions. Using decision trees, clinicians can trace back the model and either accept or reject the proposed suggestion, thus increasing their confidence about the final diagnosis. In contrast, neural networks and algorithms that need a lot of iterations in order to converge on a solution do not offer a justification of their suggestions and are regarded as ‘black boxes’ by clinicians. The DiDT architecture provides one final suggestion for each heart sound pattern, but its classification performance is lower in comparison with the DeDT architecture. The DeDT architecture has better classification performance, but provides three suggestions for each heart sound pattern; if these three suggestions are not consistent, the result can be confusing and probably less useful to the clinician. The multiple decision tree (MuDT) architecture achieves higher classification performance than the DiDT and also provides a single suggestion. This work has demonstrated that decision tree algorithms can be used as a basis for decision support systems to assist inexperienced clinicians with heart sound diagnosis. Decision trees can be very useful knowledge management tools in this area. They codify and effectively incorporate the knowledge of numerous highly specialised and experienced doctors, making them a valuable and useful tool for the exploitation and dissemination of knowledge. Such computer-based support can play a role in improving the quality and effectiveness of primary care, particularly in small and remote areas. In these places it may help reduce unnecessary patient travel for specialist consultations and investigations. Different decision tree structures and architectures were constructed and tested on various training and test data sets. Their performance on the training data sets The Journal on Information Technology in Healthcare 2004; 2(6): 413–426

423

Stasis, Pavlopoulos & Loukis was 100% successful, while their performance on the test data sets, which is an indicator of their generalisation capabilities, was satisfactory. For this difficult and complicated differentiation problem our decision tree structures achieved classification accuracy and sensitivity levels of almost 70% and classification specificity levels greater than 80%. These results are encouraging, taking into account the limited amount of data available for this study, and also the existing possibilities for classification performance improvements. Relevance analysis can be used to determine a small critical subset of the initial set of features that contains most of the information required for heart sound diagnosis. Further improvement in the classification performance of the examined decision tree structures and architectures is necessary. We believe that this is possible by: a) Using more heart sound signals with these morphological characteristics to give us larger training and test data sets. b) Developing more sophisticated MuDT architectures. For example, we can improve the architecture shown in Figure 7 by adding three two-categories DeDTs in the first (suggestion) stage. Each of these will be trained and become specialised (and, therefore, more efficient than the three-categories DiDT) in differentiating between two of the three targeted morphological characteristics (i.e. one DiDT for differentiating between S4 and SP1, one for differentiating between SP1 and EC, and one for differentiating between S4 and EC). If two DeDTs detect the corresponding morphological characteristics, then the arbitration module will use the output of the corresponding two-categories DiDT as the final suggestion; if all three DeDTs detect the corresponding morphological characteristics, then the arbitration module will use the output of the three-categories DiDT as the final suggestion. Along these directions, further research is already in progress. Additional research is being conducted concerning the use of neural network architectures for this differentiation problem, and the comparison of their classification performances with those of the decision trees architectures. Initial results from such studies have shown that neural network architectures can provide small improvements in classification performance (2–3% increase in classification accuracy). This improvement in performance compensates for the disadvantage of being a ‘black box’ that does not provide justification for its suggestions. Further research is required for the development of a systematic methodology for designing arbitration rules. In addition the design of an appropriate MuDT architecture for achieving the best classification performance for a specific problem, which could possibly include as ‘nodes’ not only decision trees but also other types of classifiers as well, is an open research question. Finally, the proposed decision trees structures and architectures can be applied to other heart sound diagnosis (or medical diagnosis in general) problems and should be further evaluated and improved.

424


A Multiple Decision Tree-based Method ACKNOWLEDGEMENTS The authors would like to thank Dr D. E. Skarpalezos for his clinical support, Dr G. Koundourakis, and Neurosoft S.A. for their support and provision of Envisioner, a data-mining tool that was used to execute algorithms related to the decision trees.

REFERENCES 1 Criley SR, Criley DG, Criley JM. Beyond heart sound: an interactive teaching and skills testing program for cardiac examination. Blaufuss Medical Multimedia, San Francisco, CA, USA, Computers in Cardiology 2000; 27; 591–94. 2 Stasis AC, Loukis EN, Pavlopoulos SA, Koutsouris D. Using decision tree algorithms as a basis for a heart sound diagnosis decision support system. Proc of the 4th Annual IEEE Conf on Information Technology Application in Biomedicine. UK, 2003, pp. 354–57. 3 Leung T, White P, Collis W, Brown E, Salmon A. Analysing paediatric heart murmurs with discriminant analysis. Proceedings of the 19th Annual conference of the IEEE Engineering in Medicine and Biology Society. Hong Kong, 1998, pp. 1628–31. 4 Durand L, Guo Z, Sabbah H, Stein P. Comparison of spectral techniques for computer-assisted classification of spectra of heart sounds in patients with porcine bioprosthetic valves. Med Biol Eng Comput 1993; 31: 229–36. 5 Wu CH. On the analysis and classification of heart sounds based on segmental Bayesian networks and time analysis. Journal of the Chinese Institute of Electrical Engineering, Transactions of the Chinese Institute of Engineers 1997; Series E, 4: 343–50. 6 DeGroff CG, Bhatikar S, Hertzberg J, Shandas R, Valdes-Cruz L, Mahajan RL. Artificial neural network-based method of screening heart murmurs in children Circulation 2001: 103; 2711–16. 7 Leung T, White P, Collis W, Brown E, Salmon A. Classification of heart sounds using timefrequency method and artificial neural networks. Engineering in Medicine and Biology Society. Proceedings of the 22nd Annual International Conference of the IEEE, 2000, Vol. 2, pp. 988–91. 8 Sharif Z, Daliman S, Sha’ameri AZ, Salleh SHS. An expert system approach for classification of heart sounds and murmurs. Signal Processing and its Application, 6th International Symposium. IEEE, 2001, vol. 2, pp. 739–40. 9 Stasis A. Decision support system for heart sound diagnosis, using digital signal processing algorithms and data mining techniques. Phd Thesis; National Technical University of Athens; 2003. 10 Epstein EJ. Cardiac Auscultation. Oxford: Butterworth–Heinemann, 1991. 11 Baracca E, Scorzoni D, Brunazzi MC, et al. Genesis and acoustic quality of the physiological fourth heart sound. Acta Cardiologica 1995; 50: 23–28. 12 Kamber M, Winstone L, Gong W, Cheng S, Han J. Generalisation and decision tree induction: efficient classification in data mining. Proc. of 1997 Int’l Workshop on Research Issues on Data Engineering (RIDE’97). Birmingham, England, 1997, pp. 111–20. 13 Koundourakis G. EnVisioner: a data mining framework based on decision trees, October 2001, Doctoral Thesis, University of Manchester Institute of Science and Technology. 14 Mitchell T. Machine Learning. Mc-Graw Hill Companies Inc., 1997. 15 Han J, Kamber M. Data Mining: Concepts and Techniques. Morgan Kaufman Publisher, 2001. The Journal on Information Technology in Healthcare 2004; 2(6): 413–426

425

Stasis, Pavlopoulos & Loukis APPENDIX: RECORDED HEART SOUND SOURCES A1 Karatzas N, Papadogiann D, Spanidou E, Klouva F, Sirigou A, Papavasiliou S. Twelve Recorded Lessons with heart Sound Simulator. Merk Sharp & Dohme Hellas, Medical Publications Litsas, Athens 1974. A2 Criley JM, Criley DG, Zalace C. The Physiological Origins of Heart Sounds and Murmurs. Harbor UCLA Medical Center, Blaufuss Medical Multimedia, 1995. A3 Littman. 20 Examples of Cardiac & Pulmonary Auscultation. A4 Cable C. The Auscultation Assistant, 1997. http://www.wilkes.med.ucla.edu/intro.html. A5 Glass L, Pennycook B. Virtual Stethoscope. McGill University, Molson Medical Informatics Project 1997. http://www.music.mcgill.ca/auscultation/auscultation.html. A6 VMRCVM Class of 2002. Virginia Maryland regional College of veterinary medicine. http://students.vetmed.vt.edu/2002/cardio/heartsounds.html. A7 Kocabasoglu YE, Henning RH. Human Heart Sounds. http://www.lf2.cuni.cz/Projekty/ interna/heart_sounds/h12/index.html. A8 Frontiers in Bioscience. Normal and Abnormal EKGs and Heart Sounds. http://www. bioscience.org/atlases/heart/sound/sound.htm. A9 Student Internal Medicine Society. Class 2000 Cardiology Heart Sounds Wave Files. American College of Physicians, Baylor College of Medicine. http://www.bcm.tmc.edu/class2000/ sims/HeartSounds.html.

426

HTI J


HTI J


Web-based Asthma Collaboration Management and Public Awareness Michael Glykas, Panagiotis Chytas* Departments of Financial and Management Engineering and *Department of Business Administration University of Aegean, Chios, Greece.

ABSTRACT Objective: To improve the management of asthma through a Web-based tool. Design: Developmental process and pilot trial. Setting: The system is used in primary care. Methods: AsthmaWeb has been designed to improve asthma management and increase independent patient living. It achieves this through user profiling and personalisation and collaborative work between health professionals, therapists, caregivers and patients through tele-care and tele-consultation. Patients enter their data regularly into AsthmaWeb and use the support function of the program to manage their asthma. Entries into the system can be made via a computer, mobile phone or personal digital assistant. The data can be accessed by the patient’s healthcare provider and the system can also alert the healthcare provider if pre-set conditions are met. The system was pilot tested by twenty patients and twenty healthcare professionals over a two week period. Results: Each patient made three entries a day giving a total of 840 entries over the two week test period. Ten patients made their entries using a computer and ten using a mobile phone. Healthcare professionals made 80 entries using a computer. No problems with using the system were reported and the system was rated highly by both patients and healthcare professionals on a number of parameters including ease of use and user satisfaction. Conclusion: AsthmaWeb is a versatile, easy to use tool that can achieve its aims of providing patient education, monitoring and management. It has many potential benefits for both patients and healthcare providers.

INTRODUCTION Asthma is a chronic lung condition that can develop at any age. It is most common in childhood, occurring in approximately 7–10% of the paediatric population and accounting for 25% of school absenteeism. Asthma affects twice as many boys as girls in childhood, but more girls than boys develop asthma as teenagers so that by adulthood the male to female ratio is equal. The severity of the disease is variable ranging from very mild, with attacks only occurring during vigorous Correspondence and reprint requests: Panagiotis Chytas, Department of Business Administration, University of Aegean, Chios, Greece. E-mail [email protected].

427

Glykas & Chytas exercise, to very severe where symptoms may occur every day causing lifestyle restriction. Worldwide it is estimated than there are 150 million asthmatics1 with almost half experiencing symptoms that disrupt their everyday lives2. It is of note that prevalence is increasing on a global basis, and in industrialised countries is associated with a trend of increased deaths and hospitalisations3. Every year approximately 180,000 patients die as a result of their asthma4. Many of these deaths can be prevented by appropriate education and treatment. Asthma is due to extra sensitive or hypersensitive airways, which can cause one or a combination of the following symptoms: • Wheezing • Coughing • Shortness of breath • Chest tightness To enable optimal management of their disease, asthmatics and their families need the following points to be addressed: • Adequate information and education • Up-to-date medication control • An environment in which their quality of living can be as good as nonasthmatics Effective management of chronic illnesses such as asthma, requires a close partnership between the patient and all healthcare providers5. Studies have demonstrated that long-term monitoring of asthma severity can reduce asthma exacerbations, optimise drug therapy and decrease the cost of asthma management6. The management of chronic patients is a collective and cooperative enterprise that may exploit information technologies (IT) to improve the overall quality of care7. In particular the development of the Internet and the World Wide Web has provided a new medium for interconnecting people throughout the world8–15. The Web has become a standardised infrastructure for giving access to sophisticated telemedicine applications from virtually any machine and operating system14,15. Such a standardised communication platform guarantees accessibility and usability advantages to both patients and physicians16. The aim of this paper is to present a Web-based asthma tool (AsthmaWeb) that significantly enhances public information and awareness to support illness prevention and independent living through user profiling, personalisation and collaborative work between health professionals, therapists, caregivers and patients through Tele-Care and Tele-Consultation. METHODS AsthmaWeb has been designed to fulfil three crucial processes related to patients with asthma or any other chronic disease: patient monitoring; management; and education.

428


Web-based Asthma Collaboration Management and Public Awareness • Monitoring is based on data gathering and analysis with respect to an individual patient’s condition. It is achieved through a portable spirometer. • Management involves an asthma action plan generated by the doctor that is customised for a particular patient. This is necessary because there is evidence that asthma education alone in the absence of an action plan and regular medical review is insufficient to improve patient outcomes17. The asthma action plan is developed by the patient’s doctor to help in the management of asthma episodes. It is a customised plan that tells the patient what to do based on changes in his/her symptoms and peak flow numbers. • Education is achieved through online information that was developed by a team of asthma experts. One way of improving patient education is to

The Caregiver, Doctor and Therapist can work together to monitor medication plans and goals via the Internet

Caregiver

Patient Medication Management

Medication Management

Doctor

Diagnosis and Care issues

Diagnosis and Care issues

User Interface: Scheduling System with Medication Management, Plan Management and Goal Management Capabilities

Mobile Level

Surveillance Centre Level

Therapist

New Plans and Goals via UMTS Technology

Execution Management Behaviour with time specifications and relations to plans and goals

I N T E R N E T S E R V E R

Execution Management Knowledge Base

For each Plan Executed for specific Goals the Server records the behaviour of the patient

Surveillance Server: Patient Positioning, Messaging to Caregivers and therapists

Caregivers

Nurses

Therapists

Doctors

Coworkers

Information and Knowledge items as well as Alerts in emergency cases are defined by caregivers, doctors and therapists. Access is provided to Execution Management Knowledge Base via the Internet Server. Messages are sent via UMTS. Electronic Patient Health Records

Figure 1. AsthmaWeb Infrastructure The Journal on Information Technology in Healthcare 2004; 2(6): 427–438

429

Glykas & Chytas provide computer-based access points where patients can browse through information screens explaining medical concepts. It has been demonstrated that such systems, if carefully designed, are considered helpful by patients18. The technological infrastructure that builds the aforementioned conceptual framework is based on different layers as shown in Figure 1. The system is composed of three main levels each one of which utilises a separate server or device: • The Mobile Level • The Surveillance Service Level • The Internet Service Level The AsthmaWeb system aims to achieve the following objectives: Public Awareness and Patient Education through Online Information The information content of the site includes information that increases public awareness and supports illness prevention in causes of asthma. More specifically it attempts to: • Increase awareness of asthma and its public health consequences • Promote identification of reasons for the increased prevalence of asthma • Promote awareness of the association between asthma and the environment • Reduce asthma morbidity and mortality • Improve asthma management • Improve availability and accessibility of effective asthma therapy • Provide patient education and support Surveillance Centre The system increases the level of personalised care for asthmatic patients. User profiles may be established directly or on-line through interactive Web forms. These information templates determine the pattern of information assembled for each asthmatic user. A complete user management function allows for central billing, the management of additions and deletions, administrative and statistical functions. The communications capability of the platform may be used to deliver information in the most appropriate format for voice, fax or e-mail, or by means of Universal Mobile Telecommunications System (UMTS) to mobile telephone networks and individual users. Data collection (patient measurements) are achieved through a spirometer. The device can easily transfer data to a desktop or laptop personal computer (PC), a mobile phone or a personal digital assistant (PDA), either through a serial connection or an infrared connection. When the device is connected to a PC, the data is stored on a local database, and with the help of the supporting software that accompanies the device, data is exported to a file. The patient or caregiver can upload the measurements to the Web-server through the Internet by selecting the appropriate file. If the devices are connected to a PDA,

430


Web-based Asthma Collaboration Management and Public Awareness the patient or caregiver can send the data to the Web-server through e-mail. With respect to use with a mobile phone, the data is passed from the device to the phone through an infrared connection. Then the user (patient or caregiver) sends the data to the Web-server via SMS (short message service). Internet Server for Collaboration through Tele-Consultation and Tele-Care An Internet server for collaboration through tele-consultation and tele-care has been created which all interested parties may access for diagnosis, treatment and group-working purposes. The Internet server can be directly linked to the surveillance server and to other health record databases that might contain critical information about the patient’s past medical history. At the heart of the Internet server is a computer equipped with facilities for interacting with external telecommunications and different computing systems. System Architecture The main architecture of the system is illustrated in Figure 2. The system is a Webbased tool and utilises the latest technology to meet the home healthcare needs of asthmatic patients. In addition it provides doctors with the necessary infrastructure to enable real-time collaboration for promoting better health and improving the quality of life for asthmatics. The AsthmaWeb system can be linked with hospital information systems and can be accessed on a 7 day/24 hour basis.

Figure 2. System architecture The Journal on Information Technology in Healthcare 2004; 2(6): 427–438

431

Glykas & Chytas

Figure 3. Unreviewed Exceptions SYSTEM DESCRIPTION Healthcare Professional Manager The healthcare professional manager makes the following services available to each healthcare professional. Unreviewed Exceptions These refer to conditions that the treating healthcare professional wants his attention called to, e.g. a specified number of readings in the red zone, or the frequency at which the patient is entering data into the system. The exceptions are easily

Figure 4. Patient Overview

432


Web-based Asthma Collaboration Management and Public Awareness configured by the treating physician. To open a list of patients with ‘Unreviewed Exceptions’ the physician clicks on the heading shown in the menu on the left of Figure 3. The patient manager opens a list of patients who have unreviewed exceptions. In Figure 3, the crosses indicate an exception for the patient. Patient Overview To review a patient’s online chart, the healthcare professional may click on an exception, or may click on the sub-menu labelled ‘Overview’ under the Patient menu (Figure 4). The graphs are powerful sources for identifying trends and changes in a patient’s condition. The healthcare professional may click to enlarge them and obtain options to change how they are plotted. The user can also edit the patient’s ‘Exceptions’, ‘Triggers’ and ‘Medications’. Measurement Diary The healthcare professional may review an individual patient’s data in a table formatted measurement diary to help him/her assess the patient’s health status and compliance. The diary includes PEF (peak expiratory flow rate) and FEV-1 (forced expiratory volume in 1 second) data and personal best percentages (Figure 5). Triggers The submenu ‘Triggers’ is a special feature used to notify the principal healthcare professional in case of an emergency (Figure 6). The healthcare professional may choose the means of notification (SMS or e-mail) and can set the limits at which a trigger will take place. It should be made clear that these triggers are a mechanism of the AsthmaWeb system and must not be confused with asthma triggers, i.e. factors that can precipitate an asthma attack.

Figure 5. Measurement Diary The Journal on Information Technology in Healthcare 2004; 2(6): 427–438

433

Glykas & Chytas

Figure 6. Triggers Asthma Action Plan The healthcare professional may enter and edit an asthma action plan for each patient. This plan details what actions a patient should take when his peak flow is in each of the PEF zones (Figure 7). This information is available to patients on their Web page but they are unable to edit it. Patient Manager Patients may collect their data through a peak flow meter, which they connect to a computer, a PDA or a mobile phone. They can upload the data from their computer to the Web server using an application that comes with the device tool, send it as SMS from their mobile phone or an e-mail from a PDA device.

Figure 7. Asthma Action Plan

434


Web-based Asthma Collaboration Management and Public Awareness

Figure 8. File Selection Upload Test Data To upload their test data the patient or caregiver first clicks on the menu labelled ‘Upload Test Data’ (Figure 8). This will take the user to the screen shown in Figure 8. By clicking the ‘Browse’ button on the right side, a pop-up window appears. With the help of the pop-up window, the user may select the appropriate file, which contains the data exported from the stand-alone application that comes with the spirometer device. Once the file has been selected the user simply has to click on the ‘Submit’ button and the data will be uploaded. The data will then be available to be viewed by both the patient or the doctor.

Figure 9. Health Status The Journal on Information Technology in Healthcare 2004; 2(6): 427–438

435

Glykas & Chytas Health Reports The patient status report summarises trends (Figure 9). The measurement diary report lists PEF and FEV-1 readings, date and time of readings and medication markers. For both reports, the patient may select the time range. SYSTEM TESTING The primary research question was whether AsthmaWeb would be a usable tool for healthcare professionals with respect to patient monitoring and management, and for patients, whether it would enhance their daily activities and lifestyle. The four specific usability attributes19–27 tested were: • Usefulness • Ease of use • Ease of learning • Satisfaction Usefulness refers to the ability of a system to be used to achieve a desired goal23, whereas satisfaction refers to the user’s perception of the pleasantness of using the system, i.e. whether they like it or not23. The pilot testing took place over a two week period and involved twenty Greek patients and twenty healthcare professionals (ten from Greece and ten from the UK). Each patient made three data entries per day (usually morning, noon and night) resulting in a total of 840 data entries over the two week period. Half the patients uploaded their data entries via a PC and half uploaded them via a mobile phone. The PDA module was not fully developed at this time and was consequently not evaluated in this trial. Healthcare professionals made 80 entries and all of these were made using a PC. No problems with using or accessing the system were reported by either patients or healthcare professionals. Specific features of AsthmaWeb, including portability28 were evaluated on a scale of 1–10 by both patients and healthcare professionals. The results of this evaluation are shown in Table 1.

Table 1. Evaluation of AsthmaWeb’s features by healthcare professionals and patients

Evaluating Group Healthcare Professionals Patients

436

Usefulness

Ease of Use

Ease of Learning

Satisfaction

Portability

6 7

7 8

7 8.5

7 8

8 8


Web-based Asthma Collaboration Management and Public Awareness DISCUSSION AsthmaWeb has been designed to improve patient monitoring, management, education and enhance collaboration between patients and healthcare professionals. Preliminary evaluation of the application has demonstrated the feasibility of using it in clinical practice and its acceptance by both patients and healthcare professionals. Both groups were satisfied with the user interface and the functionality of the system. Users found it helpful as it enabled them to easily access a large amount of useful information and dramatically reduced the time they spent retrieving data. The rapid expansion of Internet technologies and services has resulted in many households, schools and doctors having a multimedia PC connected to the Internet. The social benefits of the Internet and mobile technology are well-established and the potential healthcare benefits of these technologies are increasingly being recognised. AsthmaWeb enables healthcare professionals to use their time and skills more effectively and efficiently in caring for asthmatic patients. It also provides a practical solution to the problem of remote asthma monitoring and management. AthmaWeb permits doctors to monitor and manage patients whether they are at home, work, school or even on holiday. This, in turn, should lead to improved patient functionality and reduced mortality. However, the many potential benefits of AsthmaWeb for both healthcare professionals and patients remain to be demonstrated. REFERENCES 1 World Health Organisation (WHO). (2000a) Bronchial Asthma, Fact Sheet No. 206, http:// www.who.int/inf-fs/en/fact206.html. 2 Rabe KF. Clinical management of asthma in 1999: the Asthma Insights and Reality in Europe (AIRE) study. Eur Respir J 2000; 16: 802–7. 3 Lundback B. Epidemiology of rhinitis and asthma. Clin Exp Allergy 1998; 28(Suppl 2): 3–10. 4 World Health Organisation (WHO). World Health Report. 5 Lorig KR, Bodenheimer T, Holman H, Grumbach K. Patient self-management of chronic disease in primary care. JAMA 2002; 288: 2469–75. 6 National Heart, Lung and Blood Institute. Expert Panel Report 2: Guidelines for the diagnosis and management of asthma. NIH Publication, No. 97-4051, 1997. 7 Berg M. Patient care information systems and health care work: a socio-technical approach. International Journal of Medical Informatics 1999; 55: 87–101. 8 Detmer W, Shortliffe EH. Using the Internet to improve knowledge diffusion in medicine. Communications of the ACM 1997; 40: 101–8. 9 Raghupathi W, Tan J. Strategic uses of information technology in healthcare: a state-of-theart survey. Topics in Health Information Management, 1999; 20: 1–15. 10 Silver MS. Systems That Support Decision Makers: Description and Analysis. New York: Wiley, 1991. 11 Inmon WH, Hackathorn RD. Using the Data Warehouse. New York: Wiley, 1994. 12 Ba S, Lang KR, Whinston AB. Enterprise decision support using intranet technology. Decision Support Systems 1997; 20: 99–134. The Journal on Information Technology in Healthcare 2004; 2(6): 427–438

437

Glykas & Chytas 13 Brooks RW. Using an intranet for physician desktop data consolidation. Topics in Healthcare Information Management 1999; 20: 16–23. 14 Vanoirbeek C, Rekik YA, Karacapilidis N, Aboukhaled O, Ebel N, Vader JP. A Web-based information and decision support system for appropriateness in medicine. KnowledgeBased Systems 2000; 13: 11–19. 15 Hersch W, Brown K, Donohow L, Cambell E, Horacek A. CliniWeb: managing clinical information on the World Wide Web. Journal of the American Medical Informatics Association 1996; 3: 273–80. 16 Bellazzi R, Montani S, Riva A, Stefanelli M. Web-based telemedicine systems for home-care: technical issues and experiences. Computer Methods and Programs in Biomedicine 2001; 64: 175–187. 17 Gibson PG, Coughlan J, Wilson A, et al. The effects of limited (information only) asthma education on health outcomes of adults with asthma. Cochrane Reviews issue 3, Oxford: Update Software, 1998. 18 Jones RB, Navin LM, Murray KJ. Use of a community-based touch-screen public-access health information system. Health Bulletin 1993; 51: 34–42. 19 Rubin J. Handbook of Usability Testing: How to Plan, Design, and Conduct Effective Tests. New York: John Wiley and Sons, Inc, 1994. 20 Lund AM. USE Questionnaire Resource Page. http://www.mindspring.com/~alund/USE/ IntroductionToUse.html. 21 Nielsen J, Mack RL. Usability Inspection Methods. New York: Wiley, 1994. 22 March A. Usability: the new dimension of product design. Harvard Business Review 1994: 144. 23 Nielsen J. Usability Engineering. Boston: AP Professional, 1993. 24 Desurvire HW. Faster, cheaper! Are usability inspection methods as effective as empirical testing? In Nielsen J, Mack RL (eds.). Usability Inspection Methods. New York: Wiley, 1994, pp. 173–202. 25 Moyes J, Jordan PW. Icon design and its effect on guessability, learnability, and experienced user performance. Proceedings of the HCI’93 Conference on People and Computers. 1993; VIII:. 49–59. 26 Mayhew DJ. The Usability Engineering Lifecycle. San Francisco: Morgan Kaufman, 1999. 27 Pfleeger SL. Software Engineering: Theory and Practice. Upper Saddle River, NJ: Prentice Hall, 1998. 28 Mooney JG. Strategies for supporting application portability. IEEE Computer 1990; 23: 59–70.

438

HTI J


HTI J


Power to the Patient, using DI@L-log Lesley-Ann Black, Michael McTear, Norman Black Faculty of Engineering,University of Ulster, Northern Ireland.

ABSTRACT Chronic care patients today expect to be more actively involved in the management of their disease. Self-monitoring of blood glucose and blood pressure is strongly advocated for people with type 2 diabetes. We describe a system known as DI@L-log that enables patients to send their data (blood glucose, weight, blood pressure) to the point of care on a weekly basis using spoken dialogue technologies over the telephone. DI@L-log aims to act as a ‘teleconsultant’ in-between hospital visits by recording patient data more regularly. Data is analysed before being disseminated to authorised medical professionals who can examine the information, draw comparisons, and advise on therapeutic intervention, if required. The system also provides support to the patient at any time of day by proactively engaging them in the management of their disease. DI@L-log is currently being evaluated in diabetic, hypertensive patients in Northern Ireland.

INTRODUCTION The management of patients with diabetes is presenting a major challenge in many developed countries. This is due to an increase in the number of diabetic patients, and particularly those with non-insulin dependent or type 2 diabetes mellitus (T2DM). Better knowledge and understanding of the complications associated with poor diabetic control, has made the management of diabetes a high healthcare priority. Management of diabetes requires a strong partnership between health carers and their patients. Over the past two decades there have been substantial changes in diabetic care delivery. The traditional paternalistic ethos of patient care where the physician imparts medical knowledge is being gradually replaced by the ‘patient as a partner’ model, in which the patient is continually informed and educated about their disease status1. With an increase in patient numbers, and increased generation of information, rendering cost-effective management of diabetes is a complex issue worldwide. Effective organisation and communication of health information, and the acquisition and dissemination of patient data have become of paramount importance. Correspondence and reprint requests: Lesley-Ann Black, Faculty of Engineering, University of Ulster, Jordanstown, Belfast, Northern Ireland. E-mail: [email protected].

439

Black, McTear & Black Information and communication technology has the potential to improve the management and care of patients with chronic diseases. One of the more recent techniques employed to enable patients to relay home monitored information is by using interactive speech technologies over the telephone. These developments have enabled live operators to be replaced by intelligent agents which record, store and forward information to the hospital. In this paper we describe our proposed framework, DI@L-log, which enables diabetic patients to send their data to the point of care on a weekly basis using spoken dialogue technologies over the telephone. RATIONALE FOR DEVELOPING DI@L-log The rationale for developing DI@L-log was based on qualitative questionnaires and interviews conducted with forty patients and five healthcare professionals. These were designed to identify their perceptions about diabetes, compliance with treatment, efficacy of home monitoring, basic user needs and attitudes towards technology. Interviews with the medical professionals demonstrated high levels of anxiety towards the scale of the T2DM problem, with additional pressures from the problems associated with everyday management. A lack of patient motivation, multi-tablet therapy and obesity were frequently cited as major hurdles for doctors to contend with in managing the disease. Patient questionnaires revealed that the majority of patients had an incomplete awareness of the serious macrovascular complications associated with diabetes. In addition, the majority were not reaching their target range for weight, blood pressure or blood sugar. Many patients expressed a wish to see their health data more regularly and to have more support between clinic visits. When asked about technology, most patients claimed to be computer illiterate and expressed anxiety towards new technologies. The findings from this survey confirmed the need for both healthcare professionals and patients to have a better system for management of diabetes, including the provision of patient support and ready access to patient data by both patients and their carers. Although these requirements may be best met today through use of a personal computer (PC) and the Internet, the fact that many T2DM patients are computer illiterate and concerned about new technology presents a problem. The common telephone is an alternative method to meet these aims. It offers an inexpensive and practical solution that can be used by almost all patients. THE DI@L-log SYSTEM The system architecture is shown in Figure 1. To use the system, patients dial a designated number using their telephone. When connected, a voice recording requests them to provide their personal identification number (PIN). Once they have provided this they are asked to provide their weight, blood sugar and blood

440


Power to the Patient, using DI@L-log

Voxpilot Telecommunications Host “Welcome to DI@L-log Interactive. Please say your unique pin number”

Linked to LAN

Virtual VUI

PATIENT

Abnormal results colour-coded alerts generated

User modelling, inference Tangible GUI

Clinic Intranet EPR In clinic

Appropriate action taken PDA

Patient receives monthly report on progress

Data visualisation Trend analysis Decision support

Mobile Doctor/Nurse

Figure 1. The interactive DI@L-log prototype architecture

pressure in that order. Data may be provided either by speaking into the telephone or using the telephone keypad. The entered data is stored on a Local Area Network (LAN) clinical database. The system directed dialogues have been created using VoiceXML (VXML) version 2.0. A small set of grammars have been coded to verify the most likely user input. System prompts have been kept to a minimum length to reduce interaction times and the likelihood of user confusion. Prompts have also been written incrementally, to elaborate on information, for example, when the user requires help. Internal and external grammars based on WOZ (Wizard of Oz) scenarios have been developed to represent all possible user inputs. The system is modelled on a set of protocols, and clinical practice guidelines based on weight, blood sugar and blood pressure, stored in the knowledge base. Scripts have been mapped to represent all the possible dialogue paths through the network (for example, consulting the database, exiting the system or other eventualities that may arise during an interaction). Figure 2 shows a dialogue flow model for eliciting weight. The form interpretation algorithm will repeat or loop within the dialogue until a reasonable value is confirmed and the system can move on. The system, utilises the ODBC (Open Database Connectivity) and has been initially tested on the Apache Tomcat 4.0 Web server. JSP (Java Server pages) which consult the database (created in MS Access) dynamically convert VXML files and update the new readings obtained. These form the basis of the dialogue that the system will give in its feedback to the patient at the end of the call session. The The Journal on Information Technology in Healthcare 2004; 2(6): 439–445

441

Black, McTear & Black

Weight Prompt

Patient input (Speech /DTMF)

Input Recognised?

no

Help Exit

Event Handlers

yes Output is correct. Go to blood sugar dialogue

Figure 2. Dialogue flow for ‘weight’ task patient is unaware of the various functions that calculate and process the datasets ‘behind the scenes’. Structured sessions can become quicker as the patient becomes more familiar with the dialogue flow, and with entering data and answering questions. More expert users have the facility to use full-duplex (barge-in) facilities if they want to interrupt the system and speed up the interaction. Entered data is analysed before being disseminated to authorised medical professionals who can examine the information, draw comparisons and advise on therapeutic intervention if required. The system aims to provide support to the patient at any time of day by proactively engaging them in the management of their disease. DI@L-log Table 1. Task Specification for Weight, Blood Sugar and Blood Pressure Variable

Measurement Device

Effect of Control

Equipment Supplier

Task Familiarity

Clinic Target

Weight

Scales

Weight loss reduces likelihood of hypertension and increases insulin sensitivity

Patient

Familiar

Ideal Weight (BMI)

Blood Sugar

OneTouch Ultra Glucose Meter

Better glucose control reduces microvascular complications (e.g. blindness and nephropathy)

Lifescan Johnson & Johnson

Familiar

4–7mmol/l (NICE)

Blood Pressure

OMRON M5-I sphygmomanometermeter

Decreased blood pressure reduces cardiovascular complications (e.g myocardial infarction and stroke)

OMRON UK

Unfamiliartraining provided

140/80mm Hg (BHS)

BHS = British Hypertension Society, BMI = Body Mass Index, NICE = National Institute of Clinical Excellence.

442


Power to the Patient, using DI@L-log could extend its features to provide simple monthly progress charts to encourage and motivate patients. Three main variables have been chosen to be monitored by the system and these are weight, blood sugar and blood pressure. The variables, measurement devices, and the medically validated targets are shown in Table 1. TRIAL DI@L-log is being trialed in hypertensive T2DM patients from the Ulster Community Hospitals Trust (UCHT). This is a secondary healthcare provider to almost 3,000 diabetes patients in Northern Ireland. The specific group of hypertensive diabetics has been chosen because data indicates that co-morbid hypertension is not currently controlled in 65% of T2DM patients in the UCHT2. This has created a growing communication and management chasm between patients and their care providers with the potential for a large number of serious health complications. The trial is designed to assess the benefits of the system in improving control of blood glucose, blood pressure and weight. It will also allow system developers to assess its effectiveness as a computer-based patient management system, to determine the accuracy and quality of data received, and to evaluate if this method of data collection is preferred over a paper logbook or live contact centre. With respect to the latter point it will explore if patients prefer to use speech or the telephone keypad when interacting with the system. Patient empowerment will be evaluated by assessing patient satisfaction, acceptability and usability of the DI@L-log application. DISCUSSION The management of chronic diseases is presenting a major challenge world-wide and information and communication technology is increasingly being used in innovative ways to meet this challenge. Many new and developing systems based on use of PCs and the Internet have been demonstrated to confer significant benefits to both patients and healthcare providers3–5. However, a major drawback is that many patients and particularly the elderly are not computer literate and do not have access to a PC. Ideally any system should be able to be used by and benefit all patients6. With respect to this, the telephone is a tool that all patients are familiar with. It has been used in medical practice for over a century to deliver healthcare consultations and follow-up in a practical and inexpensive fashion. Its potential for replacing clinic visits and providing new telemedicine services has resulted from developments in technology and in particular the convergence of voice and data networks over the Web. The benefits of using the telephone to improve the management of patients with type 2 diabetes has been previously demonstrated7. In the UK a specialised conThe Journal on Information Technology in Healthcare 2004; 2(6): 439–445

443

Black, McTear & Black tact centre was created to empower patients through regular teleconsultations with trained ‘telecarers’. Patients provided their home-measurements of blood sugar levels and received advice on treatment regimens. Initial results indicated that almost 70% of the 600 patients felt more in control of their diabetes with the support of the centre, and there was a 0.4% decrease in HbA1c (glycosylated haemoglobin) for patients who had moderate or poor control. Other benefits demonstrated were a reduction in costs associated with type 2-associated complications, a reduction in staffing costs and an increased trust in the doctor/patient relationship. Many patients favoured not seeing their healthcare provider in person, instead preferring to use this form of interaction. Findings estimated savings of £300 million in hospital costs over ten years and further savings on social care costs for live triage consultation centres. These findings were supported by a report from the National Audit Office in 2002 which concluded that contact centres are a cost effective method to deliver public services and associated with high levels of customer satisfaction8. A drawback of live contact centres is the high staffing costs, particularly if they are to provide help and support 24 hours a day, 7 days a week. This cost can be eliminated through the use of an interactive voice response system coupled with an autonomous artificial system to collate patient data. An example of an interactive voice system is TLC (telephone linked care), which has been developed at the Boston Medical Centre. TLC has been implemented for a variety of chronic disease projects such as TLC Cancer Care9. It offers telephone counselling using automated pre-recorded digitised human speech to monitor and advise on medications. TLC ‘speaks’ to patients using computer controlled digitised human speech and patients use their telephone key pad to communicate responses. Results have demonstrated that the system is effective, inexpensive ($1 per call) and well-accepted by patients. Other projects such as Homey10, and Italy’s ITC-irst11 (Interactive Sensory System Division) are further developing telemedicine services through the realisation of intelligent, and adaptive medical dialogue systems capable of handling mixed initiative spoken dialogue interactions. Patients report vital signs information including weight, blood pressure, heart rate, and in addition their physical activity on a daily basis. These applications represent a new generation of telephone systems which can dynamically change and update as new user input is acquired. The system reported here has been developed based on our survey findings that many elderly patients are computer illiterate. It is designed to enable anybody to easily provide home measurements over the telephone using interactive voice technologies. Its use and potential benefits in clinical practice are currently being evaluated. CONCLUSIONS The benefits of using the telephone as a practical, efficacious way to relay homemonitored information over the traditional paper logbook and face-to-face consul-

444


Power to the Patient, using DI@L-log tations have been explored. Previous studies have shown high user satisfaction, and cost effectiveness in using both live and artificial teleconsultations. Furthermore, it seems an inexpensive and feasible method of helping contain the costs of managing type 2 diabetes mellitus. The benefits should help improve trust and communication between both patients and healthcare profesionals. Further work is, however, necessary to examine how such techniques can be successfully integrated into workflow practices to promote decision support without hampering the workflow processes of today’s ‘modern, dependable’ healthcare strategy. REFERENCES 1 Charles C, Whelan T, Gafni A. What do we mean by partnership in making decisions about treatment? BMJ 1999; 319: 780–82. 2 Jordan, A, Gingles, J. Ulster Community and Hospitals Trust Local Needs Assessment for Diabetic Services Diabetic Service Advisory Group Report. Northern Ireland, 2002. 3 Guendelman S, Meade K, Benson M, Chen YQ, Samuels S. Improving asthma outcomes and self-management behaviors of inner-city children: a randomized trial of the Health Buddy interactive device and an asthma diary. Arch Pediatr Adolesc Med 2002; 156: 114–20. 4 Krishna S, Francisco BD, Balas EA, Konig P, Graff GR, Madsen RW. Internet-enabled interactive multimedia asthma education program: a randomized trial. Pediatrics 2003; 111: 503–10. 5 Bomba D, Fulcher J, Dalley A. Construction of a diabetes database and pilot evaluation of iKey controlled GP–patient access. The Journal on Information Technology in Healthcare 2004; 2: 329–39. 6 Finkelstein J, Cabrera MR, Hripcsak G. Internet-based home asthma telemonitoring: can patients handle the technology? Chest 2000; 117: 149–55. 7 BT Health Newsletter Online. Positive results from Salford’s Contact Centre for Diabetes Patients. http://www.btwebtools.com/healthnewsletter/. 8 National Audit 2002. http://www.nao.org.uk/publications/nao_reports/02-03/0203134. pdf. 9 Mooney KH, Beck SL, Friedman RH, Farzanfar R. Telephone-linked care for cancer symptom and monitoring: a pilot study. Cancer Practice 2002; 10: 147–54. 10 “Project Homey”: Intelligent Dialogue Systems for Home Care of Chronic Patients. http://www. labmedinfo.org/research/homey/homey_flyer_english.pdf. 11 Azzini I, Falavigna D, Gretter R, Lanzola G, Orlandi M. First steps toward an adaptive spoken dialogue system in the medical domain. Proceedings of EuroSpeech 2001. Aalborg, Scandinavia: 1327–30.

HTI J


445

HTI J CALL FOR PAPERS

ONLINE COMMUNITY

The journal aims to educate healthcare professionals on the use of IT in healthcare, and to provide them with objective evidence of the benefits of IT in clinical practice. To achieve these aims the journal seeks to publish the following types of papers:

http://www.communityzero.com/JITH

(i) Articles that educate healthcare professionals on the principles and practice of the use of IT in healthcare. Articles should be written in a format that can be understood and appreciated by readers without in-depth knowledge of computers or medical informatics.

Access papers published in the journal Participate in discussions Vote in polls Post notices of meetings

(ii) Articles that provide practical advice for selecting and implementing IT solutions into clinical practice and/or optimising use of existing IT systems. (iii) Papers that provide objective evidence of the benefits of IT in clinical practice. All articles submitted to the journal undergo peer review. To ensure their international relevance, and that they can be understood and appreciated by an international readership, at least 2 reviewers for each article are not from the same country as the author(s). In addition, for research papers, at least 2 reviewers are clinicians. If you would like to submit an article or paper to the journal, please see the ‘Instructions for Authors’.

An online community has been created for readers of the journal to enable them to exchange ideas, knowledge and experience. Among its features members can:

Membership to the community is free. The website can be explored by clicking on the Preview button and following the instructions. If you have any questions or would like further information about the online community, please e-mail [email protected]. REVIEWERS The journal is dependent on reviewers to ensure the quality and originality of the papers it publishes. If you would like to act as a reviewer for papers submitted to the journal, please e-mail your details and areas of interest to [email protected].

HTI J

HJ I T

The Journal on Information Technology in Healthcare

INSTRUCTIONS FOR AUTHORS Aims and Scope: The Journal on Information Technology in Healthcare aims to improve the quality and safety of patient care, by encouraging and promoting the use of information technology (IT) in healthcare. The journal acts as a medium for the international exchange of knowledge and experience of the benefits of IT in healthcare. It principally publishes papers that objectively demonstrate clinical and cost benefits of IT in healthcare. It also publishes papers that offer practical advice for selecting and implementing IT systems. Submission of Papers: Papers in keeping with the aims of the journal should be submitted to the Editor. The journal will accept papers submitted by e-mail to editor@jith. net. Manuscripts submitted by post should be typed, with double spacing, on one side of A4. One copy of the typescript and illustrations, together with an exact matching copy on floppy disk or CD-ROM should be sent to: Clyde Saldanha, Editor, The Journal on Information Technology in Healthcare, 72 Churston Drive, Morden, Surrey, SM4 4JQ, UK. A copy of the manuscript should be retained as insurance against loss in the mail. Acknowledgements of Submissions: All submissions will be acknowledged on receipt by e-mail to the corresponding author. Further correspondence will be made as appropriate. Abstract: Research papers should have a structured abstract with the following headings: Objective, Design, Setting, Methods, Results, Conclusion. Style and Content of Manuscript: Papers should be set out under the headings: Introduction, Methods, Results and Discussion. They should as far as possible be written in a format that can be understood by readers who do not have in-depth knowledge of computers or medical informatics. Papers should ideally contain sufficient information for interested readers to be able to evaluate the feasibility and cost-effectiveness of implementing a described system into their practice. Studies should abide by high ethical standards, and if appropriate, should have received ethical approval from the local institutional human research committee. Use British rather than American spelling throughout the manuscript. References: References should be identified by Arabic superscripts in the text and numbered in the order in which they appear in the paper. They should be listed at the end of the paper in the order that they are first cited in the text. Give the full names and initials of all authors, unless there are more than six, when only the first three

should be given, followed by et al. Names should be followed by the title of the article; the title of the journal; the year of publication; the volume number and the first and last page numbers. Journal titles should be given in full or abbreviated according to the style of Index Medicus. Titles of books should be followed by the place of publication; the publisher and the year. Peer Review: All papers submitted to The Journal on Information Technology in Healthcare undergo a peer review process. On the basis of the reviewers’ responses, papers will be rejected, accepted subject to revision or accepted unconditionally. Rejected manuscripts will not be returned to authors unless expressly requested. Acceptance of Papers: On acceptance the Editor retains the right to make stylistic changes, decide on the date of publication and shorten material if necessary. Final versions of all accepted manuscripts must be submitted in electronic format. Proofs: Authors are sent one copy of the proofs. Corrections should be confined to typographical errors or matters of accuracy. Extensive amendments are not permissible. Authors should return proofs as soon as possible and not later than the date given in the covering letter. Reprints: These are available if ordered at the time of returning the proofs. Details of charges will be sent with the proofs. Copyright: All material received by The Journal on Information Technology in Healthcare are assumed to be submitted exclusively and not to have been previously published in the English language. The author(s) bear(s) the responsibility for checking whether material submitted is subject to copyright or ownership rights; for example in the use of diagrams or tables. If a submission is finally published the copyright becomes that of The Journal on Information Technology in Healthcare, and permission for any reproduction must be sought from the Editor. Dual Publication: A paper is accepted for publication on the understanding that it has not been submitted simultaneously to another journal in the English language. Accuracy and Liability: A contribution is accepted on the strict understanding that its author(s) is (are) responsible for the accuracy of all information contained in the contribution and that references to named people and/or organisations are both accurate and without libelous implications. Conflict of Interest: If authors have received payment/substantial gifts from companies whose products are featured in the paper, or have a commercial interest in the hardware and/or software used for the study, this must be acknowledged in a statement at the end of the text.

HTI J

HJ I T


The Journal would gratefully like to acknowledge the time and expertise of the following people who acted as reviewers for manuscripts submitted to the Journal in 2004 Michael Albisser (USA) Philippe Andrivet (France) Paul de Clercq (Netherlands) Isabelle Colombet (France) Carlos Costa (Portugal) Arnulf Ferlitsch (Austria) Omar Al Ghamdi (Saudi Arabia) Marcin Grabowski (Poland) Teodor Grantcharov (Denmark) Scott Hayek (USA) Ean-Wen Huang (Taiwan)

Chak-Wah Kam (Hong Kong) Rita Kobb (USA) Toby Lipman (UK) Jay Mabrey (USA) Greg Makoul (USA) James Mayrose (USA) Dinesh Mital (USA) James Ralston (USA) Anne-Marie Rassinoux (Switzerland) Dean Sittig (USA) Raimund Vogl (Austria)

SPECIAL ANNOUNCEMENT In 2005, the Journal will be available online only. The subscription rates have been reduced to reflect the fact that there will be no hard copies. There is also a special subscription rate for organisations which will allow members of large organisations to access the Journal for less than 1 pound or 1 dollar per year. To renew your subscription or start a new subscription, please fill in the form at the back of the journal. For all enquiries please contact [email protected].

HTI J

Subscription To

HJ


I T

❏ Please enter my 2005 subscription to The Journal on Information Technology in Healthcare (ISSN 1479-649X) for online access only to the 6 issues for 2005 and back issues of the Journal.

Name:

SUBSCRIPTION RATES

Postcode:

Institutional

Individual

£100 €180 US$180

£70 €120 US$120

Sterling Euros US Dollars Organisations Up to 100 members 101–500 members 501–1,000 members 1,000+ members

Country: Tel: Fax:

£200 £300 £400 £500

Developing Countries Institution/Organisation £50 or US$100

Address:

or US$350 or US$500 or US$700 or US$1,000 Individual £30 or US$60

❏ I enclose cheque/ banker’s draft made payable to The Journal on Information Technology in Healthcare. ❏ Please invoice me. ❏ I have arranged payment by Bankers Automated Credit Service (BACS) to Barclays Bank, 8 Alexandra House, Wimbledon, London SW19 7LA. Sort Code 20-96-89 Account Number 00728837 Name of account: The Journal on Information Technology in Healthcare

E-mail (required):

PLEASE POST TO: Subscriptions The Journal on Information Technology in Healthcare 72 Churston Drive Morden Surrey SM4 4JQ United Kingdom E-mail : [email protected] Fax : +44 (0)870 130 1572