🫀 Single-Beat Ventricular Late Potential Detection

This repository contains the code and experiments for the paper:

“Single-Beat Feasibility of Ventricular Late Potential Detection on Standard ECG Leads via Interpretable Gradient Boosting”

---

📌 Overview

Ventricular late potentials (VLPs) are low-amplitude, high-frequency signals associated with arrhythmogenic risk. Traditionally, they are detected using signal-averaged ECG (SAECG), which requires long recordings and specialized processing.

This project investigates:

Can VLP-like signals be detected from a single ECG beat using interpretable machine learning?

---

🧪 Key Contributions

• First patient-wise nested cross-validation evaluation
• Synthetic VLP injection framework
• Multi-domain feature engineering (>200 features)
• Gradient boosting (XGBoost, CatBoost)
• SHAP interpretability

---

⚠️ Disclaimer

This is a feasibility study using synthetic VLPs.
Results do NOT represent clinical performance.

---

⚙️ Pipeline

1. Beat extraction (R-peak aligned)
2. Synthetic VLP injection
3. Feature engineering
4. Model training (10 models)
5. Nested cross-validation
6. SHAP analysis

---

💻 Environment

• Python 3.10
• NumPy, Pandas, SciPy
• Scikit-learn
• XGBoost, CatBoost
• PyWavelets, NeuroKit2

---

🚀 Usage

git clone <repo-url>
cd repo
pip install -r requirements.txt
python run_experiment.py

---

📜 Citation

@article{novak2026vlp,
  title={Single-Beat Feasibility of Ventricular Late Potential Detection},
  author={Novak, Andrej and Barić, Domjan},
  year={2026}
}