experimental & clinical cardiology

EXPERIMENTAL & CLINICAL CARDIOLOGY

Volume 20, Issue 8, 2014

Title: "Improved Easi Ecg Method As a Future Tool in Diagnostics of Patients Suffering from Noncommunicable Diseases (ncds)"

Authors: Wojciech Oleksy, Ewaryst Tkacz and Zbigniew Budzianowski

How to reference: Improved Easi Ecg Method As a Future Tool in Diagnostics of Patients Suffering from Noncommunicable Diseases (ncds)/Wojciech Oleksy, Ewaryst Tkacz and Zbigniew Budzianowski/Exp Clin Cardiol Vol 20 Issue8 pages 3663-3667 / 2014

Improved Easi Ecg Method As a Future Tool in Diagnostics of Patients Suffering from Noncommunic...

Experimental and Clinical Cardiology

Improved EASI ECG method as a future tool in diagnostics of patients suffering from noncommunicable diseases (NCDs) Original Article

Wojciech Oleksy1, Ewaryst Tkacz2 and Zbigniew Budzianowski2

1 Institute of Informatics, Faculty of Automatic Control, Electronics and Computer Science at Silesian University of Technology, Gliwice, Poland 2 Department of Biosensors and Processing of Biomedical Signals, Faculty of Biomedical Engineering at Silesian University of Technology, Gliwice, Poland Correspondence: Wojciech Oleksy, Institute of Informatics, Faculty of Automatic Control, Electronics and Computer Science at Silesian University of Technology, Gliwice, Poland, e-mail [email protected] © 2013 et al.; licensee Cardiology Academic Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited

Abstract Electrocard iography, technique, w hich is the essential tool in the d iagnosis of heart d isease, as w ell

as other organs, is used by d octors for over 100 years. It is used to m easure electrical activity of the heart as a

Exp Clin Cardiol, Volume 20, Issue 8, 2014 - Page 3663


function of tim e and it presents it in d igital or analogue form . The m easurem ent is usually record ed from the bod y surface of the patient, w hich m akes the stand ard electrocard iogram com pletely d evoid of pain. Whilst the stand ard 12 lead ECG is the basic clinical m ethod of heart d iagnosis it has its d raw backs. Measuring all 12 lead s is often d ifficult and im practical, m ost of all it restricts patient m ovem ent. In 1988, Gord on Dow er d eveloped a system of quasi-orthogonal lead called EASI, w hich uses only 5 electrod es in ord er to register stand ard 12 lead ECG signals. The m ain goal of this w ork is to d evelop a m od el using m achine learning algorithm s w hich transform s electrocard iographic signals (ECG) perform ed by EASI into a stand ard 12-channel ECG. EASI w as proven to have high correlation w ith stand ard 12 lead ECG, it is easier an d faster to use because of sm aller num ber of electrod es. Most of all it increases m obility of patients.

the A and I electrod es are at the left and right m id auxiliary lines, respectively. The S electrod e is at the sternal m anubrium . The fifth electrod e is a groun d and is typically placed on one or the other clavicle, see Figure 1. EASI w as proven to have high correlation w ith stand ard 12 lead ECG, as w ell as w ith Mason Likar 12-Lead ECG. Apart from that it is less susceptible to artefacts, it increases m obility of patients, it is also easier and faster to use because of sm aller num ber of electrod es. What is m ore, sm aller num ber of electrod es red uces cost of a d evice [3].

Keyw ord s EASI, ECG, Regression, Machine Learning

Figure 1. Lead placem ent for EASI m ethod .

1.

Introd uction

In Eu rope noncom m unicable d iseases (N CDs) are responsible for the largest share of m ortality: about 80% of d eaths in 2009. Am ong all of them , d iseases of the circulatory system caused nearly 50% of all of d eaths. This num ber is higher am ong m en and it ranges in d ifferent countries from less than 30% to m ore than 65% of all d eaths. Those num bers are greater in less w ealthy countries, w here the social aw areness of the risks and access to qualified m ed ical personnel are lim ited . As a com parison, cancer is responsible for 20% of d eaths, ranging from around 5% to m ore than 30% in som e countries [1]. That is w hy it is so im portant to find w ays to increase the availability of d iagnostic m ethod s w hich w ill assist in the d etection of card iovascular d isease and to d evelop such a d evice, w hich w ith help of appropriate algorithm s, as w ell as telem ed icine, w ill allow m ore accurate or even autom atic and rem ote d iagnosis of the patient. All of these features has a d evice m easuring ECG using the EASI m ethod . In 1988 Dow er and his team introd uced EASI ECG system , w hich derives stand ard 12 lead ECG using only 5 electrod es [2]. The E electrod e is on the sternum w hile,

The electrod es are positioned over read ily id entified land m arks w hich can be located w ith m inim al variability, ind epend ent of the patient’s physique, assuring high repeatability. The electrod e placem ent m ake the chest largely unencum bered , allow ing physical or im aging exam ination of the hear t and lungs w ithout rem oving the electrod es. 2.

Theoretical d escription of the EASI m ethod

In the classical approach introd uced by Dow er, using the EASI lead configuration, 3 m od ified vectorcard iographic signals are record ed from the follow ing bipolar electrod e pairs:  A-I (prim arily X, or horizontal vector com ponent)  E-S (prim arily Y, or vertical vector com ponent)  A-S (containing X, Y, Z, the anteriorposterior com ponent) Each of the 12 ECG lead s is d erived as a w eighted linear sum of these 3 base signals using the follow ing form ula:



–

–

–

(1) (9)

w here L represents any surface ECG lead and a, b, and c represent em pirical coefficients. These coefficients, d eveloped by Dow er, are positive or negative values, accurate to 3 d ecim al places, w hich result in lead s very sim ilar to stand ard lead s. Our id ea to im prove EASI ECG perform ance w as to find new m od el used for 12 ECG lead s calculation. To d o that w e treated the system as a black box w ith 4 input variables: E, A, S, I and 12 output variables: I, II, III, aVR, aVL, aVF, V1, V2, V3, V4, V5, V6 and w e used various regression and m achine learning techniques to build a m od el.

(10) (11) (12) (13) N onlinear m od els tend to be m ore effective, but hard er to und erstand and build . 4.

3.

Results obtained

Method s used

Our w ork w as focused on im proving Dow er m od el by using som e regression and m achine learning techniques. Several d ifferent m ethod s w ere tested to find a best fitting m od el, nam ely LARS m ethod , Lasso m ethod , Forw ard Stagew ise m ethod [4], Linear Regression - Least Squares m ethod [5], Least Med ian of Squares m ethod [6], Regression Pace [7], Bagging Pred ictors m ethod [8], Grad ient Boosting m ethod [9], Artificial N eural N etw orks - Multilayer Perceptron m ethod [10-12] and Support Vector Machine m ethod [13]. Som e of them prod uced a linear m od el, other created a nonlinear one. Both m od els, linear and nonlinear, have d ifferent ad vantages and d isad vantages. Linear m od els, so in other w ord m od els d escribed by set of linear equations, usually are easier to und erstand and build , their total error could be easily estim ated . H ow ever, they are less effective in m ost cases. Exam ple of such a m od el is given below :

Every m ethod tested prod uced thousand s of d ifferent m od els. To d eterm ine perform ance of all m od els, for each of them correlation coefficient, root m ean squared error and m ean absolute error w as calculated . Each m od el calculation w as 10 fold cross valid ated . All results are based on d ata from PhysioN et [14] d atabase and also on synthetic d ata generated using ECG->EASI m od el d escribed by follow ing set of equations:

(14)

(15) – (16)

(2) (17) (3) (4) (5) (6) (7) (8)

Proced ure of find ing this m od el is presented in the paper Investigation Of A Transfer Function Betw een Stand ard 12-Lead ECG And EASI ECG [15]. Calculated m od els w ere com pared w ith results obtained using classical Dow er approach and also w ith Im proved EASI Coefficients d escribed in the paper “Im proved EASI Coefficients: Their Derivation, Values, and Perform ance” [16]. From all linear and nonlinear m od els obtained w e have chosen one best linear and one best nonlinear for further analysis. The best results obtained for each m od el are presented in



Table 1 (correlation coefficient), Table 2 (root m ean squared error) and Table 3 (m ean absolute error). Each of the obtained result w as proven to be statistically significant by Stud ent's t-test (w ith P < 0.01).

aVF aVL aVR I II III V1 V2 V3 V4 V5 V6

Linear Mod el 0.939 0.966 0.984 0.985 0.964 0.941 0.990 0.984 0.975 0.971 0.992 0.997

N onlinear Mod el 0.970 0.981 0.987 0.990 0.977 0.973 0.996 0.994 0.986 0.983 0.995 0.999

Dow er's Mod el 0.984 0.955 0.985 0.971 0.994 0.963 0.882 0.968 0.971 0.981 0.977 0.888

Field 's Mod el 0.776 0.922 0.966 0.973 0.894 0.786 0.849 0.872 0.751 0.851 0.970 0.985

Dow er's Mod el 28.41 35.45 31.86 40.75 32.57 37.19 99.24 177.75 120.25 144.60 119.93 93.17

Field 's Mod el 66.29 34.22 55.30 42.47 78.20 60.03 86.42 119.61 141.69 129.96 49.90 33.10

Table 1. Correlation coefficient.


Linear Mod el 27.31 21.81 15.95 17.83 26.03 31.19 20.83 40.18 46.24 55.23 24.55 10.63

N onlinear Mod el 18.32 16.63 14.70 14.99 21.13 21.55 13.78 25.24 34.58 43.04 19.49 7.41

Figure 2. Sam ple plot of an obtained linear and nonlinear m od els.

5.

H ard w are im plem entation

Currently a w ork is being d one on hardw are im plem entation of obtained m od els. Few hardw are configurations are going to be created . First there w ill be a sim ple d evice w ith a linear m od el im plem ented using operational am plifiers, transm itting ECG signals via Bluetooth, see Figure 3. Second d evice w ill be a m ore com plex solution, w hich includ es m icroprocessor, w ith a nonlinear m od el im plem ented , transm itting ECG signal via WiFi and Bluetooth.

Table 2. Root Mean Squared Error [m V].


Linear Mod el 18.10 15.20 11.64 12.72 17.53 21.09 10.94 27.19 28.18 34.89 14.17 6.86

N onlinear Mod el 14.99 12.45 11.20 11.37 15.91 16.71 10.70 18.64 23.42 29.35 13.07 5.56

Dow er's Mod el 20.33 16.38 20.93 22.32 24.74 20.63 45.90 82.90 52.76 53.24 43.92 35.14

Field 's Mod el 28.16 19.06 19.48 17.01 29.72 29.16 37.05 52.51 47.57 44.94 21.37 13.68

Figure 3. Prototype of the d evice using obtained linear m od el.

6.

Table 3. Mean Absolute Error [mV].

Sam ple plot obtained for som e m od els is show n in Figure 2.

Conclusions

Above results show that the best perform ance w as obtained for the nonlinear mod el. Second best m od el w as linear m od el. This w as the one obtained by Dow er. Surprisingly low perform ance w as observed for m od el that uses im proved EASI coefficients.



Because linear m od el w as almost as good as nonlinear one it should be consid ered as a first choice in case of hard w are im plem entation of the d evice. Further w ork in the topic of im proving EASI ECG coefficient using various regression and m achine learning techniques should be continued . 7.

[13] Alex J.Sm ola, Bernhard Schlkopf. A Tu torial on Support Vector Regression Statistics and Com puting. 2004;14:199–222. [14] PhysioN et d ata at http:/ / w w w .physionet.org/ challenge/ 2007/ d ata/

Acknow led gem ent

This w ork w as supported by the Eu ropean Union from the Eu ropean Social Fund (grant agreem ent num ber: UDA-POKL.04.01.01-00-106/ 09). 8.

[12] Frank Eibe, Witten Ian H .. Data Mining. Practical Machine Learning Tools and Techniques 2005.

References

[1] WH O Eu rope - The European health report 2012: charting the w ay to w ell-being, http:/ / w w w .euro.w ho.int/ en/ publications/ abstracts / european-health-report-2012

[15] Wojciech Oleksy and Ew aryst Tkacz. Investigation Of A Transfer Function Betw een Stand ard 12-Lead ECG And EASI ECG, BIOSIGN AL 2010 35 (2010) [16] Dirk Q. Feild Charles L. Feld m an, H orek B. Milan. Im proved EASI Coefficients: Their Derivation, Values, and Perform ance Journal of Electrocard iology. 2002;35:22–33.

[2] Patent at http:/ / w w w .google.com / patents/ US4850370/ [3] Bernice Red ley. EASI ECG Monitoring vs Trad itional 12-Lead ECG. A Review of the Literature 2005. [4] Brad ley Efron, Iain Johnstone, Robert Tibshirani. Least Angle Regression Annals of Statistics. 2004;32:407–499. [5] Trevor H astie, Robert Tibshirani, Fried m an Jerom e. The Elem ents of Statistical Learning. Data Mining,Inference,and Pred iction. 2013. [6] Peter J.Rousseeuw . Least Med ian of Squares Regression Journal of the Am erican Statistical Association. 1984. [7] Wang Yong, Witten Ian H .. Pace Regression Working Paper Series. 1999. [8] Leo Braim an. Bagging Predictors Machine Learning. 1996;24:123–140. [9] Jerom e H .Fried m an. Stochastic Grad ient Boosting Com putational Statistics and Data Analysis. [10] Sim on H aykin. N eural N etw orks: A Com prehensive Found ation Prentice H all. [11] Christopher M. Bishop. N eural N etw orks for Pattern Recognition Oxford University Press. 1995.
