zackyabd/clinical-ecg-classifier

Clinical ECG Classifier

A multi-branch deep learning model for automated ECG classification and clinical interpretation.

Model Description

This is a PyTorch-based ECG classification model that uses a multi-branch architecture combining:

• Rhythm Branch: 1D ResNet for temporal rhythm analysis
• Morphology Branch: 2D CNN for local morphology patterns
• Global Branch: 2D CNN for global ECG interpretation

The model is trained to classify 71 different ECG conditions including arrhythmias, conduction disorders, ischemia, myocardial infarction, and hypertrophy patterns.

Performance

• Macro AUC: 0.9424
• F-max Score: 0.7573
• Training Dataset: PTB-XL (21,799 ECGs)
• Test Performance: Evaluated on 2,198 test samples

Model Architecture

The model combines three specialized branches:

1. Rhythm Branch (1D ResNet): Analyzes temporal patterns for rhythm detection
2. Morphology Branch (2D CNN): Focuses on local morphological features
3. Global Branch (2D CNN): Captures global ECG patterns

All branches are fused through a final classifier with dropout regularization.

Input Format

• Input Shape: [batch_size, 12, 1000]
• Sampling Rate: 100 Hz
• Duration: 10 seconds
• Leads: Standard 12-lead ECG (I, II, III, aVR, aVL, aVF, V1, V2, V3, V4, V5, V6)

Supported Conditions

The model can classify 71 different ECG conditions including:

Rhythm Abnormalities

• Atrial Fibrillation (AFIB)
• Atrial Flutter (AFL)
• Sinus Tachycardia (STACH)
• Sinus Bradycardia (SBRAD)
• Premature Ventricular Contractions (PVC)
• Ventricular Tachycardia (VTAC)

Conduction Disorders

• Bundle Branch Blocks (LBBB, RBBB)
• AV Blocks (1st, 2nd, 3rd degree)
• Fascicular Blocks (LAFB, LPFB)

Ischemia & Infarction

• ST-T Changes (STTC)
• Myocardial Infarction patterns (AMI, IMI, LMI)
• Ischemia patterns (ISCAN, ISCIN, ISCLA)

Hypertrophy

• Left Ventricular Hypertrophy (LVH)
• Right Ventricular Hypertrophy (RVH)
• Atrial Enlargement (LAE, RAE)

Usage

import torch
import numpy as np

# Load the model
model = torch.load('ecg_model.pth', map_location='cpu')
model.eval()

# Prepare ECG data (12 leads, 1000 samples)
ecg_data = np.random.randn(1, 12, 1000)  # Replace with actual ECG data
ecg_tensor = torch.tensor(ecg_data, dtype=torch.float32)

# Make prediction
with torch.no_grad():
    logits = model(ecg_tensor)
    probabilities = torch.sigmoid(logits)
    
# Get top predictions
top_indices = torch.topk(probabilities, k=5).indices[0]

Clinical Applications

This model can assist healthcare professionals in:

• Automated ECG Screening: Rapid identification of abnormal ECGs
• Clinical Decision Support: Highlighting potential cardiac conditions
• Educational Tool: Teaching ECG interpretation patterns
• Research: Large-scale ECG analysis for epidemiological studies

Limitations

• Model trained on PTB-XL dataset - performance may vary on other populations
• Not a replacement for clinical judgment
• Requires standard 12-lead ECG format
• Should be used as a supportive tool alongside clinical assessment

Technical Details

• Framework: PyTorch
• Model Size: ~166MB
• Training: 25 epochs with AdamW optimizer
• Regularization: Dropout, batch normalization
• Loss Function: Binary cross-entropy with logits

Citation

If you use this model in your research, please cite:

@misc{clinical-ecg-classifier,
  title={Clinical ECG Classifier: Multi-branch Deep Learning for ECG Analysis},
  author={[Abdul Zacky, Irma Nia Alwijah, Muttaqin Muzakkir]},
  year={2024},
  url={https://huggingface.co/clinical-ecg-classifier}
}

License

This model is provided for research and educational purposes. Please ensure compliance with applicable medical device regulations before clinical use.

Contact

For questions or issues, please contact: zacxvan@gmail.com