ECG classification and prognostic approach towards personalized healthcare

A very important aspect of personalized healthcare is to continuously monitor an individual’s health using wearable biomedical devices and essentially to analyse and if possible to predict potential health hazards that may prove fatal if not treated in time. The prediction aspect embedded in the system helps in avoiding delays in providing timely medical treatment, even before an individual reaches a critical condition. Despite of the availability of modern wearable health monitoring devices, the real-time analyses and prediction component seems to be missing with these devices. The research work illustrated in this paper, at an outset, focussed on constantly monitoring an individual's ECG readings using a wearable 3-lead ECG kit and more importantly focussed on performing real-time analyses to detect arrhythmia to be able to identify and predict heart risk. Also, current research shows extensive use of heart rate variability (HRV) analysis and machine learning for arrhythmia classification, which however depends on the morphology of the ECG waveforms and the sensitivity of the ECG equipment. Since a wearable 3-lead ECG kit was used, the accuracy of classification had to be dealt with at the machine learning phase, so a unique feature extraction method was developed to increase the accuracy of classification. As a case study a very widely used Arrhythmia database (MIT-BIH, Physionet) was used to develop learning, classification and prediction models. Neuralnet fitting models on the extracted features showed mean-squared error of as low as 0.0085 and regression value as high as 0.99. Current experiments show 99.4% accuracy using k-NN Classification models and show values of Cross-Entropy Error of 7.6 and misclassification error value of 1.2 on test data using scaled conjugate gradient pattern matching algorithms. Software components were developed for wearable devices that took ECG readings from a 3-Lead ECG data acquisition kit in real time, de-noised, filtered and relayed the sample readings to the tele health analytical server. The analytical server performed the classification and prediction tasks based on the trained classification models and could raise appropriate alarms if ECG abnormalities of V (Premature Ventricular Contraction: PVC), A (Atrial Premature Beat: APB), L (Left bundle branch block beat), R (Right bundle branch block beat) type annotations in MITDB were detected. The instruments were networked using IoT (Internet of Things) devices and abnormal ECG events related to arrhythmia, from analytical server could be logged using an FHIR web service implementation, according to a SNOMED coding system and could be accessed in Electronic Health Record by the concerned medic to take appropriate and timely decisions. The system focused on ‘preventive care rather than remedial cure’ which has become a major focus of all the health care and cure institutions across the globe.