Frustration Prediction from Heart Rate Data

This repository contains the code and analyses for my course project in Statistical Evaluation of Artificial Intelligence and Data (02445) at DTU, June 2025.

Author: Valdemar Stamm (s244742)
Study line: Artificial Intelligence and Data

---

📌 Project Overview

This project investigates how machine learning models can be applied to predict self-reported frustration levels from physiological heart rate (HR) features.

Two approaches were implemented and compared:

• Decision Tree (DT) → simple and interpretable, well-suited for small datasets.
• Artificial Neural Network (ANN) → capable of modeling more complex patterns, but more data-demanding.

The dataset is a subset of the EmoPairCompete dataset, consisting of repeated measures of heart rate signals from individuals across rounds and phases of a problem-solving task. The prediction target is the self-reported frustration level (0–10).

---

🧠 Key Methods

• Input features: HR Mean, HR Median, HR Std, HR Min, HR Max, HR AUC
• Task type: Classification (ordinal frustration levels)
• Cross-validation: GroupKFold (grouped by individual) to avoid data leakage from repeated measures
• Model optimization: Hyperparameter tuning via GridSearchCV

---

📊 Results in Brief

• Both models showed low overall performance, reflecting dataset challenges (small size, imbalance, high variability).
• Decision Tree: Higher stability in precision/recall/F1 across folds.
• ANN: Slightly higher mean accuracy, but less consistent.
• Statistical testing: Paired t-tests showed no significant difference in performance between the models.
• Conclusion: The limitations lie primarily in the dataset rather than the models. Larger, more balanced datasets are needed for reliable frustration prediction.

---

📂 Repository Structure

• project_code.ipynb → Interactive Jupyter Notebook with full analysis, visualizations, and outputs.
• project_code.py → Standalone Python script with identical code but without inline outputs.

---

🔬 Reflections and Future Work

• Current dataset size and imbalance strongly limit model performance.
• Overfitting risk was managed with careful cross-validation, but generalizability remains low.
• Future directions:
  • Collect larger and more diverse datasets.
  • Test models on external data for robustness.
  • Explore additional ML methods and feature engineering.

---

📖 References

• EmoPairCompete dataset: Das et al. (2024). Physiological signals dataset for emotion and frustration assessment under team and competitive behaviors. ICLR 2024 Workshop.
• scikit-learn documentation for GroupKFold and GridSearchCV.