Daily Fatigue Prediction
Predicting daily fatigue scores from wearable data with high precision.
Context
During my internship at Arcascope, I tackled the challenge of quantifying and predicting daily fatigue using data from wearable devices. Fatigue is subjective and noisy, making standard regression approaches prone to overfitting.
Methodology
We approached this as an ordinal regression problem rather than simple regression or classification.
Key Innovations:
-
CORAL (Consistent Rank Logits) Layer: We replaced the standard output layer with a CORAL layer to enforce the ordinal nature of fatigue scores (e.g., Scale 1-10).
-
Loss Function Engineering: Combined
CornLoss(Conditional Ordinal Regression) with standard MSE to penalize large deviations while respecting rank order. -
Physics-Informed Features: Integrated Runge-Kutta methods to model circadian rhythm dynamics directly into the feature set.
Results vs Baseline
We iterated through ~20 architectures. The final model significantly outperformed the baseline:
| Metric | Baseline (Linear Reg) | Deep Ordinal Model | Improvement |
|---|---|---|---|
| MAE | 1.17 | 0.25 | 78% |
| Accuracy | 52% (within ±1) | 85% | +33% |
This reduction in Mean Absolute Error (MAE) allowed the system to provide actionable daily alerts to users.