4 Hz, 4 Pages: Just-in-Time Substance Use Relapse Risk Detection from Wearable Time Series Data

Substance use relapse is strongly associated with heightened stress and affective states, underscoring the importance of early detection and intervention. Wearable sensing provides a scalable pathway for delivering just-in-time adaptive interventions (JITAIs), but deployment in real-world, resource-constrained environments demands models that are lightweight, deterministic, and CPU-efficient. We introduce a streamlined pipeline that resamples multimodal wearable signals to 4 Hz, applies sliding-window segmentation, and supports two model families: MiniRocket with ridge regression for deterministic accuracy, and a compact statistical feature baseline with logistic regression for online adaptability. Using three publicly available stress/affect datasets (WESAD, PhysioNet Stress, and CAN-Stress) as proxies for relapse risk, we evaluate performance across standard and early-warning metrics, including AUPRC, AUROC, F1 at the optimal threshold, time-to-detection (TTD) at 80% recall, and per-window CPU latency. Results demonstrate competitive predictive performance with latencies consistently under 2 ms per 30 s window, highlighting the feasibility of real-time, streaming inference on commodity hardware. By emphasizing transparent, reproducible evaluation and proxy-to-relapse framing, our work provides a robust foundation for future clinical validation and has potential to enable equitable, low-resource, and globally scalable digital health interventions.