CardioNet-XL 🫀

CardioNet-XL is a specialized 1D Convolutional Neural Network (CNN) designed for multi-label classification of cardiac abnormalities from ECG signal data. Trained on the PTB-XL dataset, the model is optimized to detect five specific classes, providing a high-precision screening tool for clinical decision support.

✨ Features

Multi-label Classification: Simultaneously detects 5 cardiac abnormalities
Optimized Thresholds: Class-specific decision boundaries for clinical use
High Specificity: Excellent performance on ST/T Change (AUC: 0.918) and Normal rhythm (AUC: 0.915)
Lightweight Architecture: 4.1M parameters, suitable for edge deployment
Clinical Ready: Designed with medical decision support in mind
Trained on Large-Scale Data: Leverages the comprehensive PTB-XL dataset

📊 Dataset

PTB-XL: Large Publicly Available Electrocardiography Dataset

CardioNet-XL is trained on the PTB-XL database, one of the largest publicly available ECG datasets for machine learning applications.

Dataset Characteristics

Size: 21,837 clinical 12-lead ECG records
Duration: 10 seconds per record
Sampling Rates: 100 Hz and 500 Hz available
Patients: 18,885 unique patients
Annotations: Expert-validated diagnostic labels
Source: Physikalisch-Technische Bundesanstalt (PTB), Germany
Time Period: October 1989 to June 1996

Clinical Labels

The dataset includes comprehensive annotations for:

Diagnostic Classes: Normal ECG, Myocardial Infarction, ST/T Changes, Conduction Disturbances, Hypertrophy, and more
Form Annotations: Detailed morphological descriptions
Rhythm Annotations: Heart rhythm classifications
Demographics: Age, sex, and clinical metadata

Access the Dataset

# Download from PhysioNet
wget -r -N -c -np https://physionet.org/files/ptb-xl/1.0.3/

Citation for PTB-XL Dataset:

@article{wagner2020ptbxl,
  title={PTB-XL, a large publicly available electrocardiography dataset},
  author={Wagner, Patrick and Strodthoff, Nils and Bousseljot, Ralf-Dieter and Kreiseler, Dieter and Lunze, Fatima I and Samek, Wojciech and Schaeffter, Tobias},
  journal={Scientific Data},
  volume={7},
  number={1},
  pages={154},
  year={2020},
  publisher={Nature Publishing Group}
}

Dataset Link: PTB-XL on PhysioNet

🏗️ Model Architecture

CardioNet-XL utilizes a deep 1D-CNN architecture with 4.1M trainable parameters. The model processes raw ECG signals through three convolutional blocks followed by a dense classification head.

Architecture Overview

Layer Type	Output Shape	Parameters
Conv1D Block 1 (32 filters)	(None, 1000, 32)	2,720
BatchNormalization	(None, 1000, 32)	128
MaxPooling1D	(None, 500, 32)	0
Conv1D Block 2 (64 filters)	(None, 500, 64)	10,304
BatchNormalization	(None, 500, 64)	256
MaxPooling1D	(None, 250, 64)	0
Conv1D Block 3 (128 filters)	(None, 250, 128)	24,704
BatchNormalization	(None, 250, 128)	512
MaxPooling1D	(None, 125, 128)	0
Flatten	(None, 16000)	0
Dense (256 units)	(None, 256)	4,096,256
Dropout (0.5)	(None, 256)	0
Output Dense (5 units)	(None, 5)	1,285

Total Trainable Parameters: 4,135,717 (15.78 MB)

📈 Performance Metrics

Classification Report

The model demonstrates strong discriminative performance across major cardiac conditions:

Class	Precision	Recall	F1-Score	Support
NORM (Normal)	0.86	0.70	0.77	963
MI (Myocardial Infarction)	0.86	0.26	0.39	550
STTC (ST/T Change)	0.87	0.38	0.53	506
CD (Conduction Disturbance)	0.93	0.35	0.51	496
HYP (Hypertrophy)	1.00	0.01	0.02	262
Micro Average	0.87	0.43	0.57	2777
Macro Average	0.90	0.34	0.44	2777

Optimal Decision Thresholds

Class-specific thresholds optimized for clinical sensitivity-specificity balance:

NORM: 0.17
MI: 0.34
STTC: 0.43
CD: 0.41
HYP: 0.33

ROC-AUC Scores

STTC: 0.918
NORM: 0.915
MI: Strong performance
CD: Strong performance

🔧 Installation

Prerequisites

pip install tensorflow>=2.0.0 numpy pandas scikit-learn matplotlib wfdb

Clone Repository

git clone https://github.com/yourusername/CardioNet-XL.git
cd CardioNet-XL

🚀 Quick Start

Load Model and Predict

import tensorflow as tf
import numpy as np

# Load the pre-trained model
model = tf.keras.models.load_model('cardionet_xl.h5')

# Define optimal thresholds for each class
thresholds = np.array([0.17, 0.34, 0.43, 0.41, 0.33])
class_labels = ['NORM', 'MI', 'STTC', 'CD', 'HYP']

# Load your ECG data (shape: [batch_size, 1000, channels])
ecg_data = np.load('your_ecg_data.npy')

# Get model predictions
predictions = model.predict(ecg_data)

# Apply optimal thresholds
binary_predictions = (predictions > thresholds).astype(int)

# Display results
for i, sample in enumerate(binary_predictions):
    detected_conditions = [class_labels[j] for j, pred in enumerate(sample) if pred == 1]
    print(f"Sample {i}: {', '.join(detected_conditions) if detected_conditions else 'No abnormalities detected'}")

Input Data Format

Shape: (batch_size, 1000, channels)
Type: Normalized ECG signal (recommend z-score normalization)
Sampling Rate: 500 Hz (downsampled from PTB-XL)
Duration: 2 seconds per sample

Loading PTB-XL Data

import wfdb
import pandas as pd

# Load PTB-XL metadata
metadata = pd.read_csv('ptb-xl/ptbxl_database.csv')

# Load a single record
record = wfdb.rdsamp('ptb-xl/records500/00000/00001_hr')
ecg_signal = record[0]  # ECG data

📚 Training Details

Training Configuration

Dataset: PTB-XL (500 Hz sampling rate)
Train/Test Split: Standard PTB-XL split (stratified by patient)
Optimizer: Adam
Loss Function: Binary Crossentropy
Epochs: Early stopping at Epoch 4 (optimal generalization)
Batch Size: 32
Validation Split: 20%
Data Augmentation: Optional noise injection and time shifting

Training Observations

Consistent decrease in training loss
Validation loss begins to rise after Epoch 4, indicating optimal stopping point
Model achieves peak performance early in training
Class imbalance addressed through weighted loss function

🏥 Clinical Applications

CardioNet-XL is designed for:

Automated ECG Screening: First-line triage in clinical settings
Remote Monitoring: Wearable device integration
Emergency Departments: Rapid preliminary assessment
Telemedicine: Remote cardiac evaluation support
Research: Large-scale ECG analysis
Education: Teaching tool for ECG interpretation

⚠️ Important: This model is intended for research and clinical decision support only. It should not replace professional medical diagnosis.

📖 Citation

If you use CardioNet-XL in your research, please cite:

@software{cardionet_xl_2024,
  title={CardioNet-XL: Multi-label ECG Classification using 1D CNNs},
  author={Your Name},
  year={2024},
  url={https://huggingface.co/sid512206/CardioNet-XL}
}

@article{wagner2020ptbxl,
  title={PTB-XL, a large publicly available electrocardiography dataset},
  author={Wagner, Patrick and Strodthoff, Nils and Bousseljot, Ralf-Dieter and Kreiseler, Dieter and Lunze, Fatima I and Samek, Wojciech and Schaeffter, Tobias},
  journal={Scientific Data},
  volume={7},
  number={1},
  pages={154},
  year={2020},
  publisher={Nature Publishing Group}
}

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

Note: The PTB-XL dataset is licensed under the Creative Commons Attribution 4.0 International License.

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Areas for Contribution

Model architecture improvements
Additional cardiac condition classifications
Explainability and visualization tools
Deployment scripts for clinical environments
Documentation improvements

📧 Contact

For questions or collaborations, please reach out via:

GitHub Issues
Email: [your.email@example.com]

🙏 Acknowledgments

Dataset: PTB-XL ECG Database - Wagner et al., 2020
PhysioNet: For hosting and maintaining open medical datasets
Framework: TensorFlow/Keras
Inspiration: Clinical need for automated ECG interpretation
Community: All contributors and researchers advancing cardiac AI

📚 Related Resources

Dataset: PTB-XL on PhysioNet

Downloads last month: 6

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support