Variance-aware penalized panel models for temporal risk detection from wearable sensor data

Wearable devices generate continuous, high-resolution physiological data that offer opportunities for real-time assessment of stress, arousal, and early physiological deterioration, but existing pipelines often treat variability as nuisance noise or rely on labeled classifiers. We present a computational approach for subject-adaptive temporal risk detection that explicitly separates conditional mean and variance dynamics in high-dimensional multisubject sensor data.

The proposed penalized panel ARX-GARCHX model integrates subject-specific baselines, shared autoregressive dynamics, sparse multimodal covariate effects, and covariate-dependent volatility. It produces an exceedance-based risk score that estimates the conditional probability of crossing an individualized physiological threshold. Simulation experiments across stable, seasonal, transient-regime, and sustained-regime settings showed that modeling covariate-driven variance improves recovery of threshold-exceedance risk when volatility is structured. In the Wearable Stress and Affect Detection (WESAD) demonstration, the score provided an interpretable, label-free temporal summary that separated stress-associated windows more clearly than raw heart-rate summaries and remained lightweight for streaming use.

Variance-aware penalized panel modeling provides a reproducible methodology for converting noisy wearable streams into subject-adaptive risk-state scores. It is intended for translational feature extraction, with prospective validation required before clinical decision support.

Not applicable. This study is a methodological article that uses publicly available secondary data and does not constitute a clinical trial.