Feature normalization for enhancing early detection of cardiac disorders

Abstract

Early detection of cardiac disorders can help save many lives. Time and frequency domain statistical features derived from RR interval series of electrocardiogram (ECG) signals with a support vector machine (SVM) backend classifier can be used for distinguishing congestive heart failure (CHF) and sudden cardiac death (SCD) patients from the normal sinus rhythm (NSR) patients. We empirically found that ninety minutes of duration gave the optimal classification results after exploring with different heart rate variability (HRV) time durations. We obtained a classification accuracy of 92.85% for our baseline system using linear SVM kernel. In this work, the input statistical features consists of patient independent and patient specific variations. The patient specific variations were considered as noise in the input feature vector, while patient independent variations as informative. In this work, we experimented with two approaches. The first approach used was principal component analysis (PCA) to obtain dimensionality reduced features with maximum information stored. We obtained a performance improvement of 0.65% absolute over the baseline system. In the second approach, covariance normalization (CVN) was used to remove/minimize the effect of patient specific variations. The overall system performance was improved by 1.96% absolute over the baseline system.