DRL-ECG-HF: Deep reinforcement learning for enhanced automated diagnosis of heart failure with imbalanced ECG data

Abstract

Heart failure (HF) is a prevalent cardiovascular condition requiring accurate and timely diagnosis for effective management. Electrocardiogram (ECG) data, as a non-invasive diagnostic resource, provides crucial temporal–spatial information essential for HF diagnosis. However, traditional automated systems struggle with the temporal–spatial complexity and class imbalance of ECG data. To address these challenges, we propose DRL-ECG-HF, a deep reinforcement learning (DRL)-based multi-instance model for enhanced HF diagnosis. By treating each ECG recording as a bag of instances and analyzing individual segments, the model captures fine-grained features related to HF. To mitigate data imbalance, we introduce a DRL strategy incorporating prioritized experience replay (PER), assigning different rewards to minority class instances. The SHapley Additive exPlanations (SHAP) technique is applied to enhance interpretability, providing clinicians insights into the model’s decision-making. The proposed method was validated on the MIMIC-IV-ECG dataset with 12-lead, 10-second ECG samples from 154,934 patients and compared against various methods, including techniques for handling imbalanced data and state-of-the-art time-series classification approaches. The DRL-ECG-HF model achieved an AUROC of 0.90, an F-measure of 0.58, and a G-mean of 0.80, significantly outperforming existing methods. Additionally, it demonstrated superior performance using 12-lead ECG data compared to single-lead, emphasizing the value of comprehensive temporal–spatial information. These results highlight the potential of DRL-ECG-HF as a reliable tool for improving HF diagnosis accuracy and interpretability, paving the way for clinical adoption.