Confidence-Regulated Heart Murmur Classification via Joint Representation Learning and Decision Optimization

Accurate identification of heart murmurs from auscultation recordings is essential for early cardiovascular screening and diagnosis. While deep learning offers strong potential for automated heart murmur classification, existing models often exhibit overconfident, incorrect predictions and limited generalization due to dataset bias and class imbalance. To address these challenges, this study proposes a two-stage confidence-regulated learning framework that jointly optimizes feature representation and decision reliability. Rather than focusing solely on improving classification performance, this work emphasizes enhancing prediction reliability through confidence-aware decision-making. The proposed framework integrates supervised contrastive learning (SCL) to strengthen the discriminative structure of feature embeddings and reward-based optimization (RBO) to regulate prediction confidence under uncertainty. In this framework, a convolutional neural network-based encoder first extracts acoustic representations, and a long short-term memory-based classifier refines the learned embeddings before final prediction. SCL improves intra-class compactness and inter-class separability, while the proposed confidence-regulated mechanism enables the model to adaptively accept or defer predictions based on a dynamically adjusted threshold. This approach allows the model to balance prediction accuracy and decision reliability by reducing overconfident errors in uncertain cases. The proposed method is evaluated on the PhysioNet 2022 heart murmur dataset. Experimental results show that the proposed framework improves the validation Score from 0.8064 to 0.8233, where the Score is defined as the mean of sensitivity and specificity. More importantly, these results demonstrate improved reliability and more balanced decision-making under uncertainty, beyond a marginal increase in aggregate performance. These findings demonstrate that jointly optimizing representation learning and confidence-regulated decision-making provides an effective and clinically relevant approach for robust heart murmur classification.