Inter-patient ECG classification with intra-class coherence based weighted kernel extreme learning machine

The variability of the ECG patterns among patients often exists in real-world application of ECG classification and limits the generalization ability of existing ECG recognition approach. Furthermore, the class imbalance problem among ECG classes also poses a massive challenge to ECG recognition task. The skewed data distribution exhibited by class imbalance may produce a learning bias toward the majority class during model training, resulting in the deterioration of the recognition performance for underrepresented classes, thereby incurring the failure of the model. To cope with the above issues, a novel algorithm termed intra-class coherence based weighted kernel extreme learning machine (ICC-WKELM) is proposed for imbalanced heartbeat multiclass classification. A compact and discriminative feature set is constructed beforehand by the combination of multi-perspective features and implementation of mutual-information-based feature selection for characterization of heartbeat general features among individuals. For heartbeat classification, kernel extreme learning machine (KELM), due to its excellent classification ability, is introduced as a heartbeat classifier. In the face of imbalanced phenomenon existing in the arrhythmia classes, differing from the traditional quantity-based imbalance criterion, spatial distribution of arrhythmia samples is taken into account, and the class imbalance for arrhythmias is measured by intra-class coherence (ICC). On this basis, a novel weight assignment strategy for imbalanced arrhythmia classes is designed and ICC-WKELM algorithm for imbalanced arrhythmia multiclass classification is further proposed. The study follows the recommendations of the Association for the Advancement of Medical Instrumentation (AAMI) EC57:1998 standard and adopts the inter-patient evaluation scheme. The proposed approach is verified on the MIT-BIH arrhythmia dataset, the F1 scores for normal beat, supraventricular ectopic beat, and ventricular ectopic beat are 98.05%, 68.80%, and 93.52%, respectively, and the overall accuracy of the proposed approach reaches 96.15%.