Deep Intermediate Representation and In-Set Voting Scheme for Multiple-Beat Electrocardiogram Classification

Electrocardiogram (ECG) analysis based on computer-aided diagnostic systems plays an important role in reducing the high mortality rate among cardiac patients, but this issue undergoes the challenges from complicated variations of the data. This paper conducts a pioneering research on a new branch of ECG analysis to recognize cardiovascular diseases: multiple-beat ECG classification. In the research, this paper proposes a novel approach that comprises two components, Deep Intermediate Representation (DIR) and In-Set Voting Scheme (ISVS), for representation and classification, respectively. By combining feature learning and selection, DIR can overcome the subjectivity of feature design that fully relies on human empirical knowledge, so that the discrimination and the generalization abilities are well balanced. In the feature space, ISVS can avoid the biased judgment caused by the noisy and anomalous samples of each set by means of committee voting and majority decision, which overturns the current belief in minority-based strategies for set-based discriminative measure. The effectiveness, efficiency, and adaptability of the proposed method have been fully demonstrated via theoretical analysis and experimental validation.