Potential contributions of health data science to Learning Health Systems Tom Kelsey Reader in Health Data Science
[email protected]
Overview 1. 2. 3. 4.
Learning Health Systems Hypothesis-led research Data-driven research Potential contribution of the Health Data Scientist
Learning Health Systems
Learning Health Systems Interpret Results
Analyze Data
Aggregate Data
Deliver Message
Improve Practice
Adapted from Friedman 2014
Learning Health Systems • Finding out what constitutes best care is important • How do we deliver this care? • Jason report: November, 2014: The learning health system needs to be “closed loop” to ensure a continuous and transparent cycle of research, analysis, development, and adoption of improvements relevant to health and wellness and to the delivery of health care.
The LHS as a Research Challenge Interpret Results Deliver Message to Academics
Analyze Data
Aggregate Data After Academic Study Design
Improve Practice?
Adapted from Friedman 2014
Hypothesis-led Research
STEP 1 • • • •
Study Design Sample Sizes Consent Recruit
STEP 2
STEP 3
Collect & Analyse Study Data
Carefully Interpret Results
STEP 4 Publish Results
Hypothesis-driven Science - Example 1
Hypothesis-driven Science - Example 2
Meta Analysis Published result
Published result
Meta-analysis
Published result
1
Inclusion & Exclusion Criteria
2
Assess for Bias & Study Quality
3
Summarise Effects
4
Publish Summary Results
Meta Analysis - Example 1
Meta Analysis - Example 2
Review Published findings
Published findings
Review
Published findings
1
Identify Empirical Evidence
2
Expert Appraisal
3
Synthesis
4
Publish Results
Review - Example 1
Review - Example 2
Hypothesis-led Research • The classical approach to medical research • Fully concordant with Popper’s theories of reproducibility & falsifiability
• Data science tools and techniques are well established • For the most part
• In principle, analysis methods are defined a priori • Data are collected after we’ve set out how they will be analysed
Data-driven Science
STEP 1 Assemble Data of Interest
STEP 2 Analyse Using Modern Techniques
STEP 3 Carefully Interpret Results
STEP 4 Publish Results
Data-driven Science - Example 1
Data-driven Science - Example 2
Data-driven Science • The modern approach to medical research • But well-established as a research paradigm • Mendeleev’s Periodic Tables of 1869 • Florence Nightingale’s analysis of complex mortality data
• Data science tools and techniques are modern • Machine learning & AI • Ensemble techniques in Statistical Learning
• Analysis methods are defined a posteriori • Data are collected before we’ve set out how they will be analysed
Learning Health Systems Interpret Results
Data-driven Research
Analyze Data
Aggregate Data
Deliver Message
Improve Practice
Adapted from Friedman 2014
Potential Contributions of the Health Data Scientist 1. Ensure the highest quality of data-driven research outputs 2. Disseminate to non-academic audiences 3. Engage directly with healthcare providers 4. Use case studies from this direct engagement to promote new research 5. Engage with the health informatics landscape
Potential Contributions of the HDS Interpret Results Deliver Message to Academics
1.
Analyze Data
2.
Aggregate Data After Academic Study Design
Improve Practice? 3. 4 & 5.
Adapted from Friedman 2014
1. High-quality of research outputs • The data-driven component of the LHS loop remains important • Rapidly improving techniques & technologies • For example: a. Population studies b. Unstructured data c. Normative modelling
1a: Population studies • Linking, for example, cancer registry to maternity data can yield important new insights
1b: Unstructured data • Potentially useful healthcare data is in the form of notes, comments & audio/video files • Can the techniques we use for financial analysis be adapted to these data?
1c: Normative Modelling • Age-related models in physiology & endocrinology • Normal scores for new patients allow personalisation of care • Uterine volume – Kelsey et al., 2016 • Inhibin B – Kelsey et al., 2016 • Ovarian follicle density – McLaughlin, Kelsey et al., 2015 • Testosterone – Kelsey et al., 2014 • Ovarian volume – Kelsey et al., 2014 • Anti-Müllerian hormone – Kelsey et al., 2011 • Human ovarian reserve – Wallace & Kelsey, 2010
2. Dissemination and Outreach • The standard approach is to publish in traditional journals • With access via library/personal subscription • Or $30 to download the paper
• I prefer the open-access paradigm, making outputs available to anyone with a browser • Data and code are also made instantly available • Facilitating both reproducibility and wider dissemination
2. Dissemination and Outreach • Specialist Meeting talks • Keynote, ISFP 2015, Shanghai • Invited, CFAS 2015, Nova Scotia
• Expert panel membership • Launch event on the future of artificial intelligence and machine learning 2018, London
• Research Council Impact Acceleration • EPSRC Impact Festival 2018, Edinburgh
3. Direct Engagement with Healthcare • Make a positive contribution to better practice involves a detailed understanding of the stresses, issues and culture within healthcare systems A. At the strategic level § Setting standards for best practice
B. At the operational level § Involvement with the deployment of new care structures
3A. MSN Children & Young People with Cancer • NHS Scotland Managed Services Network • Directly funded • Cabinet Office oversight
• Tasked with the provision of high-quality & standardised care throughout Scotland • I am on the main board of the MSN • Quarterly board meetings • Funding, appointments, strategies, risks, eHealth, KPIs, …
3B. MSN CYPC Specialist Advisor • eHealth, data quality, risk assessment, IG, compliance, etc. 1. Operational Delivery Group § Detailed oversight of neuro-oncology, palliative care, psychosocial care, pharmacy, MDT organisation, etc.
2. MyStoryNow smartphone app § Detailed digital record of treatment for survivors
3. Safety Checklists Project § Migration of surgical checklist methods to oncology & haematology
4. Case studies to motivate new research • MSN: Scheduling of MDTs • EPSRC £929,076 – PI Ian Miguel
• MSN: MyStoryNow data privacy
Funded
• EC 826278 £800,000 of €4.5m – PI Kevin Hammond
• Imperial College: Individual luteal-phase support • MRC Experimental Challenges • Wellcome Trust Career Development Fellowship
• MSN: Teenager & young adult transition • NCRI TYA subgroup award
Under Review
5. The Health Informatics Landscape • Governance and oversight are vital • Subject privacy must be a leading principle
• Scotland has a complex & effective landscape • Safe Havens for data • eDRIS assist with study design & compliance • Public Benefit & Privacy Panel • SHARE opt-in • SPIRE resource for GP data
• In addition to standard ethical approval
5. The Health Informatics Landscape • I am an advocate for the use of synthetic data 1. 2. 3. 4. 5.
The potentially-identifiable data is in a safe haven A safe haven colleague produces data with the same headings and descriptive stats I produce code for analysis and test it on the synthetic data I send the code to the safe haven The code is deployed on the actual data
• Compliance is built-in to the framework
Summary Interpret Results 1. • New techniques • New technologies • Focus on personalised medicine
Analyze Data
Deliver Message 2. • Open access • Non-academic talks
Aggregate Data
4 & 5. • Case studies -> New research projects • HI landscape navigation
Improve Practice
3. • Direct engagement • Strategic oversight • Operational delivery
Any Questions?
[email protected] 