Personalizing a Generic ECG Heartbeat Classification for Arrhythmia Detection: A Deep Learning Approach

Abstract

We propose an end-to-end model for generic and personalized ECG arrhythmic heartbeat detection on ECG data from both wearable and non-wearable devices. We first develop a deep learning based model to address the challenging problem caused by inter-patient differences in ECG signal patterns. This model achieves the state-of-the-art performance for ECG heartbeat arrhythmia detection on the commonly used benchmark dataset from the MIT-BIH Arrhythmia Database. We then utilize our model in an active learning process to perform patient-adaptive heartbeat classification tasks on the non-wearable ECG dataset from the MIT-BIH Arrhythmia Database and the wearable ECG dataset from the DeepQ Arrhythmia Database. Results show that our personalization model requires a query of less than 5% of data from each new patient, significantly improves the precision of disease detection from the generic model on each new subject, and reaches nearly 100% accuracy in normal and VEB beat predictions on both databases.