AdaptStress: Online Adaptive Learning for Interpretable and Personalized Stress Prediction Using Multivariate and Sparse Physiological Signals
New model uses 5 days of smartwatch data to forecast stress 1 day ahead with MSE of 0.053.
A team of researchers has published a novel AI model, AdaptStress, that delivers interpretable and personalized stress forecasts using data from consumer-grade smartwatches. The model employs online adaptive learning to continuously update its predictions based on multivariate physiological signals, including heart rate variability, activity patterns, and sleep metrics.
Technically, the model is a time series forecaster evaluated across 16 different temporal horizons, using history windows of 3 to 9 days to predict stress 1 to 7 days ahead. In its optimal configuration—using 5 days of input data to predict stress 1 day ahead—AdaptStress achieved a mean squared error (MSE) of 0.053, a mean absolute error (MAE) of 0.190, and a root mean squared error (RMSE) of 0.226. It consistently outperformed established baseline models like Informer, TimesNet, PatchTST, and various CNN-LSTM architectures, showing improvements of 21.5% to 36.9% over the best baseline.
The context for this development is the growing need for scalable, continuous mental health monitoring. Current models often lack personalization and explainability. AdaptStress addresses this through its interpretability analysis, which revealed sleep metrics as the most dominant and consistent predictor (importance: 1.1, consistency: 0.9-1.0), while activity features showed high variability between individuals. Crucially, the model captures individual-specific patterns where the same feature can have opposite effects on different users, validating its personalization capabilities.
The practical implication is a foundation for real-world, explainable digital health interventions. By proving that consumer wearables combined with adaptive deep learning can provide relevant, individualized stress assessment, this research paves the way for integrated mental health features in fitness trackers and wellness apps, moving beyond generic insights to truly personalized care.
- Outperforms SOTA models like TimesNet and PatchTST by up to 36.9% with an optimal MSE of 0.053.
- Uses 5 days of smartwatch data (heart rate, activity, sleep) to predict stress 1 day ahead, identifying sleep as the most consistent predictor.
- Captures individual-specific patterns where identical features have opposing effects across users, enabling true personalization.
Why It Matters
Enables personalized, explainable mental health monitoring via consumer wearables, moving beyond generic fitness tracking to actionable stress forecasts.