Class-specific weighted broad learning system-based domain adaptation for patient-specific ECG classification

With the growing prevalence of wearable devices, fast and accurate classification of patient-specific electrocardiogram (ECG) is crucial for daily heart monitoring. Despite the surge in generic arrhythmia detection methods, automated systems with real-time capabilities and satisfactory performance for patient-specific ECG classification remain rare. Particularly, the morphological differences in ECG waveforms among individuals and the imbalances between beat classes pose major challenges to any model. In this paper, we propose two novel domain adaptation algorithms by improving the previously established class-specific weighted broad learning system (CSWBLS) with high resistance to imbalance data, named one-step CSWBLS-based domain adaptation (OCSWBLS-DA) and two-step CSWBLS-based domain adaptation (TCSWBLS-DA). OCSWBLS-DA achieves individual adaptation by simultaneously learning knowledge from a large number of common heartbeats without distinguishing patients and a small number of patient-specific heartbeats in one step. TCSWBLS-DA is first pre-trained on common heartbeats and then fine-tuned on patient-specific heartbeats to adapt to the corresponding individual. Both algorithms not only inherit high learning efficiency and the ability to address imbalance problems from CSWBLS but also have better generalization. Experimental results on the MIT-BIH arrhythmia database following the recommendations of Association for the Advancement of Medical Instrumentation (AAMI) EC57: 2012 standard show that both methods outperformed state-of-the-art techniques in detecting ventricular and supraventricular ectopic beats, with TCSWBLS-DA achieving the highest F1-scores of 96.1% and 80.6%, respectively. Moreover, TCSWBLS-DA takes only 0.07 seconds to adapt to a new individual and 0.034 milliseconds to identify a single beat due to its simple structure, demonstrating significant potential for practical application.