NDPAI: Neuroscience-Derived Predictive Active Inference for Calibrated Cross-Subject Wearable Stress Detection

Continuous wearable stress monitoring requires models that generalise across individuals, quantify predictive uncertainty, and capture the temporal dynamics of affective states. While discriminative classifiers achieve competitive performance on physiological benchmarks, they provide limited mechanisms for uncertainty-aware inference and interpretable, action-oriented decision making. This study proposes the Neuroscience-Derived Predictive Active Inference (NDPAI) architecture, a generative framework grounded in the Free Energy Principle for continuous physiological stress monitoring. NDPAI implements a biologically motivated dual pathway design comprising a fast amygdala-analogous pathway for low-surprise observations and a deliberative cortex-analogous pathway for high-surprise observations, with routing governed by a per-window Variational Free Energy (VFE) score whose threshold is derived ex clusively from training data. At each observation window, the architecture performs affective state inference and Expected Free Energy (EFE) minimisation, producing both a stress-state prediction and a directional monitoring-policy output. The framework was evaluated on the WESAD multimodal benchmark under leave-one-subject-out cross-validation across 15 participants, using physiological features extracted from chest-worn electrocardiography and electrodermal activity signals for three-class affective state recognition. NDPAI achieved a mean balanced accuracy of 60.3%, with no statistically significant difference detected from XGBoost (64.4%) under paired fold-level testing, while reducing EFE computation by 77% through VFE-gated routing, with only 3.4% of windows requiring deliberative processing. Post-hoc temperature scaling reduced expected calibration error from 0.319 to 0.056, lower than uncalibrated XGBoost on the same concatenated prediction set (0.128). These results demonstrate that active inference can provide competitive cross-subject stress recognition while additionally offering interpretable action outputs, adaptive computational allocation, and well-calibrated probabilistic predictions that may support future clinical decision-support pipelines after validation on ambulatory and clinical cohorts