Key result
Novel DA-BiLSTM-FNN model achieves ~93% accuracy in unsupervised cross-domain ECG arrhythmia classification.
Why the study?
Cross-domain ECG arrhythmia classification is limited by domain shift from heterogeneous devices, patient populations, and recording environments, which creates discrepancies between training and testing data and reduces the generalization ability of clinical decision support systems.
Does the unsupervised domain-adaptive DA-BiLSTM-FNN model with CSD joint loss improve cross-domain electrocardiogram arrhythmia classification between the PTB-XL and CPSC2018 datasets?
Does the unsupervised domain-adaptive DA-BiLSTM-FNN model with CSD joint loss improve cross-domain electrocardiogram arrhythmia classification between the PTB-XL and CPSC2018 datasets?
The proposed DA-BiLSTM-FNN deep learning model with CSD joint loss demonstrates high accuracy (93.35%) and generalizability in unsupervised cross-domain ECG arrhythmia classification across heterogeneous datasets.
May enhance ECG arrhythmia detection across devices; leaves open prospective validation before clinical use.
Cross-domain electrocardiogram (ECG) arrhythmia classification is challenging because of domain shift, which is related to the heterogeneity of medical data collected using various devices, collected from different patient populations, and obtained in differing recording environments. This heterogeneity leads to discrepancies between training and testing data that limit the generalization ability and practical utility of clinical decision support systems. To address the challenge of unsupervised domain adaptation, this study developed a robust deep learning framework called DA-BiLSTM-FNN that is constructed on the basis of adversarial learning. The model integrates bidirectional long short-term memory (LSTM) and an attention mechanism to extract temporal features and employs a fuzzy neural network (FNN) to manage data uncertainty and fuzzy reasoning to enhance decision stability. Additionally, a CSD joint loss function—comprising classification loss, domain loss, and a novel soft-margin disparity loss—is introduced to stabilize decision boundaries and mitigate drift in the unlabeled target domain. Under unsupervised cross-domain conditions and with the labeled PTB-XL and unlabeled CPSC2018 datasets employed as the source and target, respectively, the DA-BiLSTM-FNN model achieved 93.35% accuracy, 89.01% precision, 92.21% recall, and a 90.60% F1 score. Experimental and cross-validation results confirmed the model’s strong adaptability and generalization ability in heterogeneous ECG classification.
No takes yet. Share an insight, caveat, or question.
Lin et al. (2026) studied Cardiac arrhythmia. DA-BiLSTM-FNN model with CSD joint loss was evaluated on Classification accuracy in the unlabeled target domain (CPSC2018). The DA-BiLSTM-FNN model achieved 93.35% accuracy, 89.01% precision, 92.21% recall, and a 90.60% F1 score in unsupervised cross-domain ECG arrhythmia classification.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: