Evaluating prediction algorithm of malignant ventricular arrythmia for earlier prediction time on heterogenous databases

Abstract

Prediction of malignant ventricular arrhythmia (mVA) is utmost imperative to enable earlier medical intervention and prevent sudden cardiac death (SCD). However, patients with a history of coronary artery disease (CAD) and congestive heart failure (CHF) are at higher risk of SCD. This thesis aimed to develop a reliable mVA prediction algorithm with high performance and an earlier prediction time and evaluate in a more authentic situation mixed with other cardiac diseases which are CAD and CHF. This was done by testing the algorithm on multiple online databases which are Sudden Cardiac Death Holter Database (SDDB), MIT-BIH Normal Sinus Rhythm Database (NSRDB), Long Term ST Database (LTSTDB) and BIDMC Congestive Heart Failure Database (CHFDB). Heart rate variability (HRV) analysis with support vector machine (SVM) was employed in the prediction algorithm due to its reliability observed in previous works. To investigate the statistical relationship between all databases, 65 features were extracted from first, second, third, and fourth minute HRV signal before mVA onset and before two hours mark of control signals. Experimental results show a significant difference in HRV of mVA signals and other non-mVA signals, including six time-domain features and seven nonlinear features. Six feature combinations from time-segment-specific classification were found to perform best in predicting imminent mVA in situation mixed with CAD and CHF. High accuracy of 97.33% with 89.47% sensitivity and 100% specificity was achieved. For classification of the four distinct databases, four feature combinations of pNN50, MaxNN and CVI with CVNN, SD2, SD1a, or SDNNa achieved a high accuracy of 98.67% with 100% sensitivity and 98.21% specificity. For exploration of earlier prediction time, the six best-performing feature combinations in predicting imminent mVA with other non-mVA signals were selected for classifier training and testing in leave-one-out cross-validation classification on 120-minutes signal. A balanced performance with reasonably high accuracy of 73.33%, sensitivity of 73.68%, specificity of 73.21% and 91.14 minutes of earliest prediction time was achieved by combination of pNN50, SD1d, SDNNa with Gaussian radial basis function (RBF) SVM and moving average of 15 minutes.