Trustworthy Stress Detection from Wearable Sensor Signals for Preventive Anxiety Monitoring: Subject-Independent Accuracy and Calibrated Confidence with the Autonomic-Branch Quantum-Graph Network

Wearable monitoring of autonomic dysregulation preceding anxiety disorders must generalize to unseen wearers; reported accuracy reflects the evaluation protocol as much as the model. We measured it with one fixed instrument, the Autonomic-Branch Quantum-Graph Network (AB-QGN), mapping each autonomic branch onto one named qubit of a shallow entangling circuit. With architecture and features fixed, heart-rate variability (HRV) and electrodermal (EDA) features from WESAD and SWELL-KW went through leakage-free subject-grouped and leave-one-subject-out (LOSO) protocols and two leaky ones: single-modality configurations inflated by 27-61 macro-F1 points. Honest fused macro-F1 was 86.2 (WESAD) and 61.4 (SWELL-KW), 84.0 and 55.1 under LOSO; that distance from near-ceiling figures is the result, not a shortfall. Leakage corrupted confidence: conventional calibration left a gradient-boosting baseline at 25.6% expected calibration error, 5.4% after subject-disjoint refitting. Over structure- and parameter-matched classical twins, the quantum core added 3-12 session-level points on SWELL-KW at every seed, not significant at 51 sessions, and none on WESAD when parameter-matched. On a third corpus, AffectiveROAD, honest performance is at chance per window, while window-overlap leakage lifts tree ensembles to 90-98 macro-F1. We claim no advantage over fixed-configuration tree ensembles: the contribution is an auditable encoder and the protocol needed to trust a number.