Prediction of Ventricular Arrhythmias in Non-Ischemic Cardiomyopathy Using a 12-Lead ECG-Based Machine Learning Model

Project Overview

Arrhythmic risk stratification in non-ischemic cardiomyopathy (NICM) remains a major clinical challenge. Traditional predictors—such as left ventricular ejection fraction (LVEF < 35%), late gadolinium enhancement (LGE), and genetic markers—have limited accuracy in identifying patients at risk of sustained ventricular arrhythmias (SVA).

This project develops and evaluates an ECG-based machine learning (ECG-ML) model capable of predicting SVA in NICM patients, outperforming established clinical predictors.

---

Background

• Cohort: 314 patients with NICM, no prior SVA.

• Data: Baseline 12-lead ECGs, cardiac MRI (for LGE), and genetic testing results.

• Outcome: Sustained ventricular arrhythmia (SVA), defined as sudden cardiac death, sustained VT, or appropriate ICD intervention.

• Approach:

  • Neural network trained on 80% of ECGs.

  • Validation on an independent 20% test set.

  • Two model variants:

    1. ECG-only model (ECG-ML)
    2. ECG + Age

---

Repository Structure

data_processing/ – Scripts for ECG data preprocessing and feature extraction
evaluation/ – Model evaluation scripts and performance metrics
explainability/ – Model interpretability (e.g., Grad-CAM, SHAP)
models_training/ – Neural network training pipelines
requirements.txt – Python dependencies

Methods Summary

• Model architecture: Neural network trained on raw and processed ECG signals.

• Training strategy: Multiple ECGs per patient were included to capture intra-individual variability, and both upsampling and downsampling techniques were applied to prevent overfitting.

• Comparison: ECG-ML model vs. a multivariate clinical model including:

  • LVEF
  • LGE (presence and high-risk type)
  • Pathogenic genetic variants

• Evaluation metric: AUC (area under the ROC curve) and sensitivity/specificity balance.

Results

Metric	ECG-ML (ECG only)	Clinical Model
Sensitivity	75%	—
Specificity	85%	—
AUC (95% CI)	0.86 (0.57–1.00)	0.53 (0.24–0.82)

The ECG-ML model demonstrated strong predictive performance, significantly exceeding traditional risk factor–based models. No combination of clinical covariates improved upon ECG-ML performance.

---

Installation

Prerequisites

• Python ≥ 3.8
• GPU (recommended for model training)

Install dependencies

pip install -r requirements.txt

---

Usage

1. Preprocess the multivariate clinical and get ECGs for each patient

data_processing/NICM_Data.ipynb

2. Train the Models

models_training/NICM_WorkSpaceTF_ECG_.ipynb models_training/NICM_WorkSpaceTF_ECG_With_vars.ipynb

3. Evaluate Performance

evaluation/NICM_evaluation.ipynb

4. Explain Model Predictions

explainability/NICM_Expl.ipynb

5. Use the pre-trained models

pre_trained_models/model_Resnet_50_ecg_With_Vars.h5 pre_trained_models/model_Resnet_50_ecg.h5

Key Findings

• ECG-ML achieved AUC 0.86 for predicting sustained ventricular arrhythmia in NICM patients.
• No clinical variable (LVEF, LGE, genotype) independently improved performance.
• The model highlights the potential of ECG-based deep learning for risk stratification in cardiomyopathy.

Limitations

• Retrospective, single-center cohort
• Modest sample size (n = 314)
• Requires external validation on larger, multicenter datasets before clinical use