Multimodal stress detection system that analyzes wearable physiological signals — ECG, electrodermal activity, skin temperature, and accelerometer data — to classify emotional states. Built with classical ML (XGBoost, LightGBM, CatBoost) and deep learning (1D-CNN, BiLSTM, Transformer with cross-modal attention), featuring SHAP/LIME explainability and a FastAPI prediction service.
| Model | Approach | Accuracy |
|---|---|---|
| XGBoost / LightGBM / CatBoost | Classical ML with Optuna tuning | Competitive baselines |
| 1D-CNN with residual connections | Deep learning | Strong performance |
| Transformer with cross-modal attention | Deep learning | Best overall |
Full benchmark results in
notebooks/08_deployment_ablation.ipynb
| Notebook | What it does |
|---|---|
01 |
Data exploration and quality checks |
02 |
Signal preprocessing — filtering, artifact removal, ectopic beat correction |
03 |
Feature extraction — 60+ biomarkers from HRV, EDA, temperature, motion |
04 |
Statistical analysis and dimensionality reduction |
05 |
Classical ML — Logistic Regression, SVM, Random Forest, XGBoost, LightGBM, CatBoost |
06 |
Deep learning — 1D-CNN, EfficientNet, Bi-LSTM/GRU, Transformers |
07 |
Explainability — SHAP global/local, LIME, clinical reports |
08 |
Deployment and ablation studies |
git clone https://github.com/urme-b/CalmSense.git
cd CalmSense
python -m venv .venv && source .venv/bin/activate
pip install -e .
# Run the API
uvicorn api.main:app --reload
# → http://localhost:8000/docs
# Or with Docker
docker compose up| Method | Endpoint | Description |
|---|---|---|
POST |
/predict |
Single prediction from preprocessed features |
POST |
/predict/batch |
Batch predictions |
POST |
/predict/raw |
Prediction from raw signal upload |
GET |
/models |
List available trained models |
POST |
/explain/shap |
SHAP explanation for a prediction |
POST |
/explain/lime |
LIME explanation for a prediction |
POST |
/explain/clinical |
Clinical-style report |
WS |
/stream |
WebSocket for real-time streaming |
WESAD (Wearable Stress and Affect Detection) — 15 subjects wearing chest (RespiBAN) and wrist (Empatica E4) sensors through 4 conditions: baseline, stress (Trier Social Stress Test), amusement, and meditation.
Available from the UCI Machine Learning Repository.
Python · FastAPI · scikit-learn · XGBoost · LightGBM · CatBoost · PyTorch · SHAP · LIME · Optuna · Docker
- Add wrist-only model for smartwatch deployment
- Implement online learning for user-specific adaptation
- Add meditation and recovery state classification
- Support real-time streaming from Empatica E4 via BLE
- Build personalized stress thresholds per user
- Add Grad-CAM visualizations for deep learning models
- Integrate federated learning for privacy-preserving training
- Expand to additional datasets (AMIGOS, DREAMER, K-EmoCon)
- Add mobile app with on-device inference (ONNX/TFLite)
- Implement confidence calibration and uncertainty estimation
- Multimodal — Longitudinal neurophysiological analysis
- Sensor — Biometric sensor comparison and reference