From Confounded Signals to Causal Affect: Proximal Inference with Hybrid Deep Learning for Wearable Emotion Recognition in Precision Mental Health

Emotion recognition from wearable physiological sensors is fundamentally challenged by latent confounders. Physical activity, individual physiological variability, and environmental noise obscure genuine affective states and limit clinical interpretability. This study presents a causal multimodal deep learning framework that addresses these limitations through three integrated contributions. First, proximal causal inference establishes identifiability of individualised treatment effects under latent emotional confounding using observable physiological proxies. Second, a hybrid LSTM-CNN architecture with Hilbert-Schmidt Independence Criterion regularisation learns disentangled, time-varying representations of emotional states, separating affective signals from motion-related confounders. Third, an off-policy evaluation (OPE) protocol assesses the safety, robustness, calibration, and fairness of clinical decision policies derived from the learned causal representations. Empirical evaluation on the Wearable Stress and Affect Detection dataset demonstrates strong subject-independent performance, achieving 92.69% accuracy on an unseen evaluation subject. The proximal bridge variable exhibits meaningful variation across emotional conditions, enabling downstream treatment-effect estimation. OPE reveals that doubly robust estimation produces conservative policy values (0.454) compared to naïve inverse propensity scoring (1.051), with threshold optimisation yielding a 6.38% improvement over observational policies. Calibration analysis (ECE = 0.441) identifies the need for post-hoc recalibration before clinical deployment. The framework establishes that integrating causal inference with deep multimodal learning enhances interpretability, robustness, and clinical credibility of emotion-aware AI systems for precision mental health.

Keywords: Emotion recognition; causal inference; LSTM-CNN; proximal inference; off-policy evaluation; physiological signals; precision mental health.