Stress Detection – WESAD

Physiological stress detection using the WESAD dataset. Classifies three states — Baseline, Stress, and Amusement — from chest-worn sensor signals using classical machine learning with subject-independent evaluation.

---

Project Structure

stress-detection-wesad/
├── data/
│   ├── raw/              # Original WESAD .pkl files (not included)
│   ├── processed/        # Segmented signals as .npy files
│   └── features/         # Extracted feature CSVs per subject
├── notebooks/
│   ├── 01_exploratory_analysis.ipynb
│   ├── 02_feature_engineering.ipynb
│   └── 03_modeling.ipynb
├── results/              # Saved models, LOSO summaries, per-subject reports
├── streamlit-app.py
└── requirements.txt

---

Dataset

WESAD (Wearable Stress and Affect Detection) contains multimodal physiological recordings from 15 subjects (S2–S17, excluding S12) wearing a RespiBAN chest device and an Empatica E4 wrist device.

This project uses chest signals only:

Signal	Sensor	Sampling Rate	Filter Applied
ECG	RespiBAN	700 Hz	Bandpass 0.5–40 Hz
EDA	RespiBAN	700 Hz	Lowpass 1 Hz
Respiration	RespiBAN	700 Hz	Lowpass 1 Hz

Labels used: 1 = Baseline, 2 = Stress, 3 = Amusement (other protocol labels excluded)

---

Pipeline

1. Preprocessing (01_exploratory_analysis.ipynb)

• Bandpass filters ECG (0.5–40 Hz); lowpass filters EDA and respiration (1 Hz)
• Z-score normalizes each signal
• Segments into 60-second windows with 50% overlap (42,000 samples per window at 700 Hz)
• Assigns labels via majority vote within each window
• Saves segments and labels as .npy files per subject

2. Feature Engineering (02_feature_engineering.ipynb)

Extracts 12 hand-crafted physiological features per segment:

Domain	Features
HRV (ECG)	mean_rr, sdnn, rmssd, ecg_peak_count
Electrodermal (EDA)	eda_mean, eda_std, eda_peak_count, eda_slope
Respiration	resp_mean, resp_std, resp_zero_crossings

R-peaks are detected using a 75th-percentile height threshold and a minimum 0.5s inter-peak distance.

3. Modeling (03_modeling.ipynb)

• Evaluation: Leave-One-Subject-Out (LOSO) cross-validation — trains on 14 subjects, tests on the held-out subject, repeated for all 15
• Hyperparameter tuning: RandomizedSearchCV (20 iterations, 3-fold CV) run inside each LOSO fold on training subjects only
• Models: Random Forest and XGBoost

---

Results

Model	Mean Accuracy	Mean F1 (weighted)
Random Forest	82.9%	80.7%
XGBoost	81.4%	~79%

Top features by importance: eda_mean, eda_peak_count, resp_std, mean_rr

---

Dashboard

An interactive Streamlit dashboard for exploring results:

streamlit run streamlit-app.py

Features:

• Model selector (Random Forest / XGBoost)
• Per-subject classification report
• Confusion matrix heatmap
• SHAP feature importance

---

Setup

pip install -r requirements.txt

Run notebooks in order:

1. notebooks/01_exploratory_analysis.ipynb — preprocessing
2. notebooks/02_feature_engineering.ipynb — feature extraction
3. notebooks/03_modeling.ipynb — training and evaluation

Note: Raw WESAD data must be downloaded separately and placed under data/raw/S{id}/S{id}.pkl. The dataset is available from the UCI ML Repository.