Leakage-Aware Self-Supervised Cross-Modal Representation Learning for Subject-Independent PCG-Based Physiological Stress Inference Under Strict LOSO Evaluation

Biosignal-based monitoring is increasingly used to study physiological states, yet validated stress labels and rigorous subject-independent evaluations remain limited. We present a leakage-aware self-supervised framework for inferring heart rate variability (HRV)-derived autonomic arousal states from phonocardiogram (PCG) signals under strict leave-one-subject-out (LOSO) evaluation. The task addresses physiological autonomic states rather than psychological stress diagnosis. The framework combines fold-isolated preprocessing, contrastive PCG pretraining, electrocardiogram (ECG)-guided latent regularization based on first- and second-order statistics, and a CNN–BiLSTM–Attention classifier. Positive pairs comprise two augmented views of the same PCG segment. ECG and PCG recordings originate from independent, unpaired datasets; no synchronization, sample-level cross-modal pairing, reconstruction, or translation is assumed. Preprocessing, pseudo-label construction, augmentation, hyperparameter selection, and representation learning remain confined to training partitions, with held-out subjects reserved for evaluation. The method achieved 88.7% accuracy, an F1-score of 0.861, and ROC-AUC of 0.912. Ablation analyses showed that PCG pretraining provided the largest gain. Removing ECG-guided regularization reduced performance, but absent random-teacher or shuffled-ECG controls, this supports regularization benefits rather than ECG-specific physiological transfer. The main contribution is the leakage-aware evaluation framework, not architectural novelty. Validation in larger independent cohorts with validated stress labels is needed before clinical or real-world deployment.