A penalized, synergistic deep learning architecture leveraging DBN and GAN for classifying highly asymmetric medical data: A multi-domain validation

Detecting ECG-based arrhythmias early and accurately is challenging due to significant class imbalance, noise, and privacy barriers across medical institutions. This paper proposes a penalized, synergistic deep learning solution comprising: (i) Deep Belief Networks (DBN) for hierarchical representation learning, (ii) Generative Adversarial Networks (GAN) for augmenting the minority class with DBN feature conditioning, and (iii) a novel adaptive cost-sensitive penalty coefficient that dynamically adjusts based on training dynamics and class distribution to enhance classification robustness. Our key innovations include: (1) a feedback mechanism where DBN features condition GAN generation and GAN samples inform DBN training; (2) a dynamic penalty function with adaptive formulation; and (3) the first comprehensive multi-domain validation of such synergistic architecture across four diverse imbalanced datasets. Evaluation on MIT-BIH (patient-wise split, 9 classes), PTB-XL (5 diagnostic superclasses), CPSC-2018,(9 rhythm classes), and a real-world credit card fraud detection dataset demonstrates state-of-the-art performance with 99.47% ± 0.10% accuracy and 0.9905 ± 0.004 MCC on MIT-BIH, with strong cross-domain generalization. The same framework can potentially be extended to Federated Learning for privacy-preserving training, which we discuss as a future direction.