Evaluation of Cardiovascular Disease Risk in the China Kadoorie Biobank Using Novelty Detection

We evaluate the risks of cardiovascular disease to the Chinese population by i) detecting ”abnormality” using 3 one-class classification methods (a discriminative one-class support vector machine (SVM), a generative kernel density estimate (KDE), and a discriminative KDE), and ii) predicting probabilities of ”normality”, arrhythmia, and ischemia using 3class classification method (a discriminative SVM). For one-class classification we constructed 5 possible criteria for ”normality” and used 14 automatically extracted ECG features by the Mortara device and blood pressure data. We evaluated classification performance in balanced test sets and the discriminative SVM model achieved area under curve (AUC) 0.79-0.83 and 71.6%-75.6% accuracy, depending on the different criteria for defining ”normality”. The generative KDE achieved 0.72-0.80 AUC, and 74.8% accuracy. For three-class classification we propose a set of inclusion criteria for ”normal”, ”arrhythmia” and ”ischemia”, and added 6 new features (amplitudes of P, Q, R, S, and T waves and STlevel) extracted from time-series ECG. With new features the accuracy improved from 64.6% to 74.5% for 3-class and from 74.5% to 88.7% for classification between normal and ischemia in balanced test sets.