GradePrediction

Wearable Stress Modeling for Academic Performance Prediction

Marissa Crevecoeur · David Juarez · Sumaiyah Rahman

---

Overview

This project investigates whether physiological stress signals recorded from wearable devices during controlled lab sessions can be used to build a generalizable stress model, and whether that model's predictions correlate with academic performance when applied to real exam recordings.

The pipeline has two stages:

1. Train a stress classification model on a controlled stress-induction dataset using self-reported stress labels
2. Apply the trained model to physiological data recorded during real university exams and test whether predicted stress correlates with exam grades

---

Research Question

How can stress be modeled as a continuous physiological signal from wearable sensor data, and to what extent does this stress representation exhibit meaningful associations with academic performance?

---

Datasets

Source Dataset — Stress Training Data

Wearable Device Dataset from Induced Stress and Structured Exercise Sessions Hongn et al., PhysioNet 2025. DOI: 10.13026/he0v-tf17

• 35 participants (18 V1 protocol, 17 V2 protocol)
• Empatica E4 wristband recordings
• Controlled stress induction protocol: Baseline, Stroop, TMCT, Opinion tasks, Subtract
• Self-reported stress scores (1–10 scale) per protocol stage
• Signals: EDA, BVP, HR, TEMP, ACC, IBI

Target Dataset — Exam Stress Data

Wearable Exam Stress Dataset for Predicting Cognitive Performance Amin et al., PhysioNet 2022. DOI: 10.13026/kvkb-aj90

• 10 students across 3 exams (Midterm 1, Midterm 2, Final)
• Same Empatica E4 wristband
• Linked to actual exam grades
• Signals: EDA, BVP, HR, TEMP, ACC, IBI

---

Methodology

Source Dataset                    Target Dataset
(Lab stress sessions)             (Real exam recordings)
        │                                  │
        ▼                                  ▼
  feat_eng.py                       new_feat_eng.py
  (feature extraction)              (same features)
        │                                  │
        ▼                                  │
  train.py                                 │
  (model training +                        │
   cross-validation)                       │
        │                                  │
        ▼                                  ▼
  Saved Model  ──────────────▶  Inference + Grade
  (gb_class3_z)                  Correlation

Feature Engineering (feat_eng.py)

• 60-second sliding windows with 30-second overlap
• Features extracted per window across all signals:
  • EDA: mean, std, min, max, range, slope, SCL mean/slope, SCR peak count/amplitude
  • Temperature: mean, std, min, max, slope
  • Heart Rate: mean, std, min, max
  • HRV (IBI-derived): mean IBI, SDNN, RMSSD, pNN50
  • BVP: mean, std, LF power, HF power, LF/HF ratio
  • Accelerometer: magnitude mean, std, max
• Tonic/phasic EDA decomposition via 0.05 Hz Butterworth low-pass filter
• Artifact flagging: EDA flatline, EDA dropout, motion, HR missing
• Per-subject z-score normalization

Model Training (train.py)

• Label schemes: 3-class stress (Low: 1–3, Medium: 4–6, High: 7–10)
• Algorithms: Random Forest, Gradient Boosting
• Feature schemes: raw features, per-subject z-scored features
• Validation: nested subject-grouped cross-validation
  • Outer: StratifiedGroupKFold (5 splits)
  • Inner: StratifiedGroupKFold (5 splits) with RandomizedSearchCV (50 iterations)
• Class imbalance: class_weight="balanced" for RF; compute_sample_weight for GB
• Imputation: median imputation for missing signal windows

Data Quality Decisions

Decision	Rationale
Exclude S02	Duplicated raw signal values from E4 Connect download
Invalidate f07 BVP/HR/IBI/TEMP	Protection dock never removed — PPG sensor blocked
Split f14 into f14_a / f14_b	Bluetooth loss mid-session; baseline in f14_a
Flat-responder filter (std < 0.8)	Subjects with near-zero label variance add noise

---

Results

Stress Classification (Source Dataset)

Model	Accuracy	Balanced Accuracy	Macro F1
RF — raw features	0.562	0.382	0.377
RF — z-scored features	0.481	0.376	0.375
GB — raw features	0.533	0.363	0.364
GB — z-scored features	0.481	0.387	0.383

Best model: Gradient Boosting on z-scored features (gb_class3_z)

Chance level for balanced 3-class classification = 0.333 macro F1. All models are above chance, with the performance ceiling attributed to self-reported label noise rather than feature poverty.

Iteration History

Iteration	Best Model	Macro F1	Balanced Accuracy
Basic Features	RF	0.384	0.398
Trend Features	RF	0.437	0.444
HR Variability + Motion	RF	0.439	0.462
Baseline Normalization	RF	0.421	0.449
Clean EDA + Feature Selection (ANOVA-50)	RF	0.473	0.492

---

Limitations

Limitation	Mitigation
Small sample size (27 training subjects)	Effect sizes + confidence intervals reported
Domain shift: lab vs. exam stress	Within-subject z-score normalization; baseline-relative scoring
Self-reported labels are noisy	Flat-responder filtering; 3-class rather than 10-class binning
V1/V2 protocol differences	Protocol version tracked as metadata
Single wearable device	Empatica E4 is validated research-grade hardware

---

Repository Structure

GradePrediction/
├── feat_eng.py              # Feature extraction pipeline
├── train.py                 # Model training and evaluation
├── features_stress.csv      # Generated feature table (not tracked if large)
├── models/                  # Saved model pipelines (.joblib)
│   ├── stress_rf_class3_raw.joblib
│   ├── stress_rf_class3_z.joblib
│   ├── stress_gb_class3_raw.joblib
│   └── stress_gb_class3_z.joblib
├── reports/                 # Evaluation outputs
│   ├── summary.json         # Full results across all configs
│   ├── summary_partial.json # Incremental results (crash-safe)
│   ├── feature_importance.csv
│   └── *.txt / *_confusion.csv
├── CHANGES.md               # Branch-level change documentation
└── README.md                # This file

---

Setup and Usage

Requirements

pip install numpy pandas scipy scikit-learn joblib

1. Generate Features

python feat_eng.py

Reads from wearable-data/Wearable_Dataset/STRESS/ and writes features_stress.csv.

2. Train Models

python train.py

Reads features_stress.csv, runs nested cross-validation across 4 configurations, saves models to models/ and reports to reports/.

Expected runtime: 1–3 hours on a standard CPU instance.

On Google Cloud VM

source venv/bin/activate
nohup python3 train.py > training_log.txt 2>&1 &
tail -f training_log.txt

---

Branches

Branch	Description
main	Stable pipeline
feature/cv-fixes-flat-filter	CV leakage fix, flat-responder filter, class10 removed

---

References

1. Hongn, A., Bosch, F., Prado, L., & Bonomini, P. (2025). Wearable Device Dataset from Induced Stress and Structured Exercise Sessions (version 1.0.1). PhysioNet. https://doi.org/10.13026/he0v-tf17

2. Amin, M.R., Wickramasuriya, D., & Faghih, R.T. (2022). A Wearable Exam Stress Dataset for Predicting Cognitive Performance in Real-World Settings (version 1.0.0). PhysioNet. https://doi.org/10.13026/kvkb-aj90

3. Çöpürkaya Ç., et al. (2023). Investigating the effects of stress on achievement: BIOSTRESS dataset. Data in Brief, 49, 109297. https://doi.org/10.1016/j.dib.2023.109297

4. Hsu, A. (2023). Quantifying Exam Stress Progressions Using Electrodermal Activity and Machine Learning. IEEE BIBE 2023, pp. 434–438. https://doi.org/10.1109/BIBE60311.2023.00077