Beat-by-beat Classification of ECG Signals with Machine Learning Algorithm for Cardiac Episodes

Heart failure (HF) is a common clinical syndrome of cardiac episode leading to a variety of cardiac diseases. Detecting these cardiac episodes from electrocardiogram (ECG or EKG) data and classifying these large data automatically with high accuracy in real-time is critical for useful application of wearables targeting cardiac disease monitoring. With this motivation, in this study, we used the BIDMC Congestive Heart Failure (CHF) datasets (from PhysioNet database). A total of 15 patient records was analyzed, which have NYHA Class level III and IV patients from the database. Simultaneous measurements of the 2 leads of ECG were stored in the record. The captured data was sampled at 250 Hz. The extracted features were for three categories: temporal, spectral, and statistical. In total, we extracted 28 features out of which 7 were of amplitude types, 6 were based on frequency, and the remaining 15 were statistical features. Machine learning models explored include SVM, KNN, ensemble tree, neural network, decision tree, naive bayes, and logistic regression. We evaluated different model performance in each patient data and combined patient data. In our analysis, neural network was the best performer in terms of accuracy for cardiac patients. We further studied neural network to test sensitivity, specificity, accuracy, precision, f1-score to evaluate the best performer statistics. Neural network has 99.5% overall accuracy for interpatient data classification, and was also among the best performers. In interpatient classification, the performance was: sensitivity 99.80%, specificity 99.0%. accuracy 99.42%, precision 99.80%, and F1 score 99.64%. Accurate detection of ECG beat classes using this approach can allow real-time cardiac disease monitoring.