Graph Neural Networks for Modelling Emotion Regulation Pathways from Multimodal Wearable Signals

Emotion regulation emerges from coordinated dynamics across autonomic, cardiovascular, electrodermal, thermoregulatory, and motor systems. Although wearable devices can continuously capture these signals, most predictive models either analyse each modality in isolation or fuse them via simple feature concatenation, which obscures the structured cross-system interactions that likely underpin successful regulation. We introduce a graph neural network (GNN) framework that encodes multimodal wearable physiology as a subject-specific graph: nodes represent physiological subsystems (ECG, EDA, PPG, accelerometery, skin temperature, and HRV), and edges model functional coupling estimated from modality-level feature similarity. To enable controlled benchmarking, we generate a synthetic but physiologically plausible dataset of 150 subjects spanning three affective states (negative, positive, regulated) and derive regulation-success labels using a physiology-informed rule based on HRV and sympathovagal balance (LF/HF). A hybrid GCN-GAT architecture achieves strong held-out performance (accuracy = 0.967, AUC = 0.982, F1 = 0.982) and exceeds reference baselines in a comparative evaluation. Beyond prediction, we include attention-based interaction analysis, although the current configuration does not reveal distinct regulation circuits analyses via attention-based node and edge scoring, facilitating inspection of candidate regulation circuits across systems. Finally, we outline key barriers to translation including synthetic-only validation, class-imbalance sensitivity, and potential attention degeneracy and propose methodological steps required for robust deployment on real-world wearable emotion regulation cohorts.