ML: Logistic Regression, Random Forest, XGBoost, LightGBM
DL: 1D-CNN, SHAP
- Detects stress vs baseline from wearable signals: ECG, EDA (skin conductance), temperature, respiration, motion.
- Scored Leave-One-Subject-Out (LOSO): train on 14 people, test on the 15th, rotate.
- Shows where the usual high numbers come from: subject leakage, motion, dataset shift, calibration.
- Runs in the browser (ONNX, no backend). make demo runs the full pipeline offline on synthetic signals.
Binary (baseline vs stress), 15 subjects, LOSO, mean over held-out subjects.
| Model | Accuracy | F1 (macro) |
|---|---|---|
| Random Forest | 0.913 | 0.898 |
| XGBoost | 0.903 | 0.873 |
| Logistic Regression | 0.902 | 0.883 |
| LightGBM | 0.894 | 0.860 |
| 1D-CNN (raw signal) | 0.718 | 0.648 |
- The 4 feature models are a statistical tie (Friedman p = 0.81). RF 95% CI: [0.860, 0.960].
Key findings, one per check:
| Check | Question | Result |
|---|---|---|
| Subject leakage | Does same-person testing inflate scores? | 3-class 0.66 to 0.79 (+13 pts) |
| Motion confound | Is it just movement? | Drop all motion: 0.913 to 0.901 |
| Wrist vs chest | Is a cheap sensor enough? | 0.893 vs 0.913 (2 pts lower) |
| Dataset shift | Does it transfer to another dataset? | Near chance (0.50 balanced) |
| Calibration | Are the probabilities trustworthy? | ECE 0.070; isotonic map to 0.025 |
| Personalization | Does a short enrollment help? | 20 windows: ECE 0.146 to 0.069 |
| Model | Type | Key settings |
|---|---|---|
| Logistic Regression | Linear | C=1.0, L2, class-balanced |
| Random Forest | Bagged trees | 200 trees, depth 10, class-balanced |
| XGBoost | Boosted trees | 200 trees, depth 7, lr 0.1 |
| LightGBM | Boosted trees | 200 trees, 50 leaves, lr 0.1 |
| 1D-CNN | Deep net on raw signal | Residual blocks, AdamW, early stopping |
- Every model runs inside an impute (median) to scale to classifier pipeline, fit per fold, seeded.
| Group | Count | Examples |
|---|---|---|
| HRV time domain | 12 | MeanNN, SDNN, RMSSD, pNN50 |
| HRV frequency | 8 | LF/HF power, LF/HF ratio |
| HRV nonlinear | 10 | SampEn, DFA, SD1/SD2, CSI |
| EDA (skin conductance) | 15 | SCL level, SCR count, SCR amplitude |
| Temperature + respiration | 8 | temp slope, respiration rate |
| Accelerometer (motion) | 5 | magnitude mean, std, energy |
 Model comparison (LOSO) |
 Optimism gap (leakage) |
 Feature ablation |
 Wrist vs chest |
 Cross-dataset transfer |
 Calibration reliability |
 Few-shot personalization |
 Top features (SHAP) |
 Confusion matrix |
| Area | Tools |
|---|---|
| Modelling | scikit-learn, XGBoost, LightGBM, PyTorch |
| Signal processing | NeuroKit2, SciPy |
| Explainability | SHAP |
| Dashboard | React, TypeScript, ONNX Runtime Web |
| Tooling | GitHub Actions, ruff, mypy, pytest |
- 15 subjects, lab-induced stress. Underpowered, wide CIs. No clinical claim.
- Ablation, calibration, and personalization are exploratory, not multiplicity-corrected.
- The 1D-CNN is a small baseline, not a fair test of deep learning.
- Cross-dataset uses one confounded pair. Illustrative, not conclusive.
- A third corpus (SWELL / AffectiveROAD) for leave-one-dataset-out generalization.
- Real-world, non-lab stress data beyond the 15-subject benchmark.
- Real-time streaming inference from a live wearable.
- Physiological signals are sensitive personal data.
- This is a research benchmark, not a product.
- Data minimization: collect and keep only what an analysis needs.
- No surveillance: do not monitor or penalize people without informed consent.
- Datasets keep their own licenses and are not redistributed here.