A comprehensive study was conducted to differentiate cardiovascular disease (CVD) subjects from non-CVD subjects using short recording electrocardiogram (ECG) of 244 Malaysian adults in The Malaysian Cohort project. An automated peak detection algorithm to detect nine fiducial points of electrocardiogram (ECG) was developed. Forty-eight features were extracted in both time and frequency domains, including statistical features obtained from heart rate variability and Poincare plot analysis. These include five new features derived from spectrum counts of five different frequency ranges. Feature selection was then made based on p-value and correlation matrix. Selected features were used as input for five classifiers of artificial neural network (ANN), k-nearest neighbors (kNN), support vector machine (SVM), discriminant analysis (DA), and decision tree (DT). Results showed that six features related to T wave were statistically significant in distinguishing CVD and non-CVD groups. ANN had performed the best with 94.44% specificity and 86.3% accuracy, followed by kNN with 80.56% specificity, 86.49% sensitivity and 83.56% accuracy. The novelties of this study were in providing alternative solutions to detect P-onset, P-offset, T-offset as well as QRS-onset points using discrete wavelet transform method. Additionally, two out of the five newly proposed spectral features were significant in differentiating both groups, at frequency ranges of 1–10 Hz and 5–10 Hz. The prediction outcomes were also comparable to previous related studies and significantly important in using ECG to predict cardiac-related events among CVD and non-CVD subjects in the Malaysian population.

Cardiovascular disease; ECG fiducial point detection; electrocardiogram; feature extraction; machine learning