Skip to content

Latest commit

 

History

History
117 lines (81 loc) · 5.06 KB

README.md

File metadata and controls

117 lines (81 loc) · 5.06 KB
Raw

Global ECG Classification by Self-Operational Neural Networks with Feature Injection

Overview

This repository implements an approach for global ECG classification using Self-Operational Neural Networks (Self-ONNs) with feature injection, designed to classify arrhythmias using the MIT-BIH dataset. Our model uses temporal and morphological features from ECG signals, achieving high accuracy while maintaining a compact and computationally efficient architecture.

The method outlined in this repository is based on the paper:

Global ECG Classification by Self-Operational Neural Networks with Feature Injection
Muhammad Uzair Zahid, Serkan Kiranyaz, and Moncef Gabbouj
IEEE Transactions on Biomedical Engineering, Vol. 70, No. 1, January 2023.

Project Structure

  1. Data Preprocessing: ecg_data_processing.py — Processes raw ECG signals, aligns R-peaks, calculates wavelet coefficients, and generates datasets for model training and testing.
  2. Model Training: main.py — Implements the 1D Self-ONN model with feature injection and trains it on preprocessed data.

Table of Contents

Installation

  1. Clone the repository:

    git clone https://github.com/MUzairZahid/Global-ECG-Classification.git
cd yourrepo

  2. Download the MIT-BIH dataset: Download the MIT-BIH Arrhythmia Database from PhysioNet.

  3. Install dependencies:

    • Install Python packages.
    • Install the fastonn library for Self-ONNs: 
      git clone https://github.com/junaidmalik09/fastonn
cd fastonn
pip install .

Data Preparation

  1. Preprocess the ECG data: Use ecg_data_processing.py to preprocess the raw ECG data, align R-peaks, compute wavelet coefficients, and create training/testing datasets.

    python ecg_data_processing.py --raw_data_dir ./mit-bih-arrhythmia-database-1.0.0 --processed_data_dir ./MITBIH_data_processed

    This will save the processed data as mitbih_processed.pkl in the processed_data_dir directory.

Running the Code

  1. Train the Model: Run main.py after preprocessing the data. Specify paths and parameters as required.

    python main.py --wavelet_type "mexh" --sampling_rate 360 --q 3 --processed_data_path ./MITBIH_data_processed/mitbih_processed.pkl --save_model_dir ./saved_models

    Arguments for main.py:

    • wavelet_type: Type of wavelet to use (default: "mexh").
    • sampling_rate: Sampling rate of ECG signals (default: 360 Hz).
    • q: Degree of non-linearity for Self-ONN (default: 3).
    • processed_data_path: Path to the preprocessed data.
    • save_model_dir: Directory to save the trained model.

Proposed Approach

The proposed 1D Self-ONN model effectively captures the morphological and temporal features of ECG signals. Key components of the approach include:

  • Self-ONN Layers: These layers extract morphological features from individual ECG beats, enabling the model to differentiate arrhythmia types.
  • Feature Injection: Temporal features based on R-R intervals are directly injected into the Self-ONN model, enriching the feature space for classification and providing critical information on the sequence and timing of beats.

Figure 1: Proposed Approach. Figure 1: The proposed approach and model architecture for classification of ECG signals.

Results

The proposed approach demonstrates high accuracy in arrhythmia classification. The main results, as shown in Table 1, provide a comprehensive performance comparison across arrhythmia types. These results demonstrate that our Self-ONN approach can perform high-accuracy arrhythmia classification with minimal computational overhead, making it feasible for real-time, low-power ECG monitoring devices.

Table 1: Classification Performance. Table 1: Classification performance of the proposed 1D Self-ONN.

Contact and Collaboration

For any questions, issues, or potential collaboration inquiries, please contact:

Muhammad Uzair Zahid
Email: muhammaduzair.zahid@tuni.fi
LinkedIn: https://www.linkedin.com/in/uzair-zahid/

References

  1. Zahid, M. U., Kiranyaz, S., & Gabbouj, M. (2023). "Global ECG Classification by Self-Operational Neural Networks With Feature Injection." IEEE Transactions on Biomedical Engineering, 70(1), 205–214.
  2. MIT-BIH Arrhythmia Database. Available at: https://physionet.org/content/mitdb/1.0.0/
  3. FastONN: GPU-based library for Operational Neural Networks. Available at: https://github.com/junaidmalik09/fastonn