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

Area of Science:

• Computational physiology
• Data science
• Wearable technology

Background:

• Wearable devices collect continuous, high-resolution physiological data for real-time stress and deterioration assessment.
• Current methods often discard data variability or require labeled data, limiting real-time analysis.
• A novel computational approach is needed to effectively utilize this data.

Purpose of the Study:

• To develop a subject-adaptive temporal risk detection method for high-dimensional sensor data.
• To explicitly separate conditional mean and variance dynamics in physiological data.
• To create a label-free risk score for identifying physiological threshold exceedances.

Main Methods:

• A penalized panel Autoregressive-GARCHX (ARX-GARCHX) model was developed.
• The model integrates subject-specific baselines, shared autoregressive dynamics, and multimodal covariate effects.
• Covariate-dependent volatility was modeled to estimate conditional probability of threshold crossing.

Main Results:

• The model successfully estimated risk scores, showing improved threshold-exceedance detection when volatility was structured.
• In the Wearable Stress and Affect Detection (WESAD) dataset, the risk score provided a clear, label-free temporal summary of stress.
• The method demonstrated superior separation of stress-associated windows compared to raw heart-rate summaries and was computationally lightweight.

Conclusions:

• Variance-aware penalized panel modeling offers a reproducible method for extracting subject-adaptive risk scores from wearable sensor data.
• This approach facilitates translational feature extraction for potential applications.
• Prospective validation is necessary before clinical decision support implementation.