Improving Inter-Patient Electrocardiogram Diagnosis Through Patient Adaptive Feature Density-Based Clustering

Current computer-aided ECG diagnosis tools tend to perform poorly under the inter-patient evaluation paradigm, where the training and test sets are separated by patient to reflect real-world scenarios. We observed that decision boundaries of different arrhythmia types do not align across different patients. Consequently, a general decision boundary learned from the training set becomes inaccurate when applied to unseen patient data with different decision boundaries. This study proposes a novel prediction refinement approach for inter-patient ECG diagnosis system that utilizes patient adaptive feature density-based clustering to improve the classification accuracy of a simple supervised classifier. This test-time prediction refinement approach effectively shifts the decision boundary of the model to snap along more reasonable paths separating different clusters, thus adapting to different patient characteristics. We also propose a two-sided variable points beat segmentation scheme which better captures the morphological features of each beat and yields improved classification result. The proposed framework was evaluated on the MIT-BIH Arrhythmia Database using a standardized inter-patient evaluation paradigm. An overall accuracy of 98.7% was achieved. The F1-scores obtained for normal (N), supraventricular ectopic beats (S), and ventricular ectopic beats (V) classes were 99.3%, 89.2%, and 95.8%, respectively, competing strongly with the current state-of-the-art.