Stress Detection — WESAD End-to-End Pipeline

A complete machine-learning pipeline for physiological stress detection using the WESAD (Wearable Stress and Affect Detection) dataset.

---

Overview

This project classifies three physiological states — Baseline, Stress, and Amusement — from multi-modal wearable sensor signals:

Signal	Device	Sampling Rate
ECG (Electrocardiogram)	RespiBAN (chest)	700 Hz
EDA (Electrodermal Activity)	Empatica E4 (wrist)	4 Hz
BVP (Blood Volume Pulse)	Empatica E4 (wrist)	64 Hz

Reference: Schmidt, P. et al. (2018). WESAD: A Multimodal Dataset for Wearable Stress and Affect Detection. ACM ICMI.

---

Pipeline

Raw Signals → Preprocessing → Feature Extraction → LOSO Cross-Validation → ONNX Export

1. Dependency setup — auto-installs XGBoost, skl2onnx, onnxruntime
2. Dataset loading — synthetic WESAD-faithful simulator (no files needed) or real .pkl loader
3. Signal preprocessing — zero-phase Butterworth filters (bandpass ECG/BVP, lowpass EDA)
4. Feature extraction — 60 s windows, 50% overlap; statistical + HRV + EDA features
5. EDA — signal plots, feature distributions, correlation heatmap
6. Model training — XGBoost, MLP, Random Forest with Leave-One-Subject-Out (LOSO) cross-validation
7. Evaluation — accuracy, F1-macro, ROC-AUC, confusion matrix
8. Feature importance — Gini importances (tree models) / permutation importance (MLP)
9. Real-time inference — confidence-gated prediction (threshold = 0.60; below → "inconclusive")
10. ONNX export — cross-platform deployment with numerical consistency check

---

Quick Start

Option 1 — Google Colab (recommended)

Click the Open in Colab badge above. All dependencies are auto-installed in the first cell.

Option 2 — Local

git clone https://github.com/MithraKL/stress-detection-wesad.git
cd stress-detection-wesad
pip install -r requirements.txt
jupyter notebook stress_detection_WESAD_colab.ipynb

---

Using the Real WESAD Dataset

The notebook ships with a synthetic simulator that runs without any files. To use the real dataset:

1. Request access at WESAD @ UCI
2. Download and extract to a folder named WESAD/
3. In the notebook, uncomment the load_subject() block and set WESAD_ROOT

---

Results (Simulated Data)

Model	Accuracy	F1-Macro	ROC-AUC
XGBoost	~0.92	~0.91	~0.98
Random Forest	~0.90	~0.89	~0.97
MLP	~0.88	~0.87	~0.96

Results will vary with the real WESAD dataset and actual LOSO splits.

---

Outputs

All figures and model files are saved to /content/ in Colab:

File	Description
signal_overview.png	ECG / EDA / BVP with condition overlay
filter_effect.png	Raw vs filtered signal comparison
feature_distributions.png	Per-class feature histograms
correlation_heatmap.png	Feature correlation matrix
boxplots.png	Feature distributions by condition
model_comparison.png	Accuracy / F1 / AUC bar chart
confusion_auc.png	Confusion matrix + ROC curves
feature_importance.png	Top features (Gini / permutation)
stress_model.onnx	Exported best model for deployment

---

Next Steps

• Replace simulate_subject() with real WESAD .pkl loader
• Add PyTorch LSTM for raw-signal end-to-end learning
• Wrap best_model in a FastAPI /predict endpoint
• Add MLflow experiment tracking
• Add SHAP for per-prediction feature attribution
• Personalise model per subject with transfer learning

---

Requirements

See requirements.txt. Core dependencies:

• numpy, pandas, scipy
• scikit-learn
• xgboost
• matplotlib, seaborn
• skl2onnx, onnxruntime, onnxmltools

---

License

MIT — see LICENSE.