AthleteGuard — Non-Invasive Cortisol Prediction
A machine learning system that predicts cortisol levels from wearable
physiological signals (HRV, ECG, EDA, EMG, Respiration) without blood tests.
Model Details
- Model: Gradient Boosting Regressor (tuned)
- Dataset: WESAD (15 subjects, chest-worn sensors)
- Target: RMSSD (cortisol proxy)
- Validation: Leave-One-Subject-Out Cross-Validation (LOSO-CV)
Performance
| Metric | Value |
|---|
| Test MAE | 6.82 ms |
| Test R² | 0.859 |
| LOSO Mean MAE | 9.93 ± 7.31 ms |
| LOSO Mean R² | 0.756 |
Outputs
- Recovery Score (0-100) — Daily training readiness
- Overtraining Risk Index — Low / Medium / High via ACWR
- Pre-Competition Zone — Underactivated / Optimal / Overactivated
Additional Models
- LSTM + Attention — temporal cortisol trend tracking
- Autoencoder — physiological anomaly detection (76.4% stress detection rate)
Authors
Shubhankur — SRM Institute of Science and Technology
