Structured contrastive Self-Supervised representation learning for multimodal physiological stress detection

Stress is a pervasive physiological response to environmental and cognitive demands and is strongly associated with adverse physical and mental health outcomes. Wearable biosensors enable continuous and unobtrusive stress monitoring; however, learning robust and generalizable representations from multimodal physiological data remains challenging due to limited labeled samples, strong inter-feature correlations, and heterogeneous modality structure. While self-supervised learning (SSL) provides an effective paradigm for leveraging unlabeled data, existing tabular contrastive approaches typically rely on feature-independent augmentations that disregard physiological organization and modality-specific characteristics encountered in wearable sensing. This study introduces Structured Feature Corruption (SFC), a physiology-aware augmentation strategy for contrastive self-supervised representation learning in multimodal physiological stress detection. Instead of corrupting features independently, SFC applies modality-level corruption (e.g., heart rate and electrodermal activity), preserving intra-modality dependencies while encouraging informative invariances. The proposed framework follows a two-stage pipeline comprising contrastive pre-training on unlabeled data and supervised fine-tuning using limited labeled samples. The approach is evaluated on two multimodal physiological tabular datasets, StressData and SynthesizedStressData. Results show that SFC consistently outperforms established baselines, including SCARF, TabCLR, and VIME. SFC achieves an accuracy of 85.12% on StressData and 87.43% on SynthesizedStressData, demonstrating robust performance across datasets and corruption levels. Representation and explainability analyses further indicate improved class separability and physiologically meaningful feature contributions. These findings demonstrate that incorporating domain-specific multimodal structure into contrastive learning enhances stress detection without increasing model complexity.