Arrhythmia Beat Classification

This project is an implementation of an arrhythmia beat classification system using the publicly available MIT-BIH Arrhythmia Database. The aim of this project is to classify different types of heartbeats from ECG data using machine learning techniques. We build upon the approach outlined in the article Combining Low-dimensional Wavelet Features and Support Vector Machine for Arrhythmia Beat Classification.

Table of Contents

• Introduction
• Dataset
• Method
• Preprocessing
• Feature Extraction
• Classification
• Results
• Installation and Usage
• Contributors
• Citations

Introduction

The detection of heart arrhythmia from ECG signals is a key step in the diagnosis of cardiac disorders. This project focuses on building an automated arrhythmia detection system using a machine learning pipeline to classify different heartbeat types from ECG recordings.

Dataset

We use the MIT-BIH Arrhythmia Database, which has been a standard for arrhythmia research since 1980. The dataset consists of 48 half-hour ECG recordings, with annotations by cardiologists. These recordings include both common and less common arrhythmias, making it suitable for training classification models.

Method

1. Beat Segmentation: The ECG signal is segmented into episodes based on QRS intervals using the R-peak locations provided in the dataset.
2. Wavelet Decomposition: Features are extracted using wavelet multi-resolution analysis (WMRA).
3. Principal Component Analysis (PCA): PCA is used to reduce the dimensionality of the wavelet features.
4. Classification: Three machine learning algorithms are used to classify the heartbeats: K-Nearest Neighbors (KNN), Support Vector Machines (SVM), and Neural Networks.

Preprocessing

• Beat Segmentation: ECG recordings are segmented into beats using QRS intervals based on R-peak locations provided by MITDB.

• Feature Extraction: Wavelet analysis is applied to extract features from each beat. We used bior wavelet family and specifically the bior 6.8.

• Data Normalization: StandardScaler is used to normalize the data.

  Example code:

  import qrsintervals
  # Define extraction parameters
  mitdbih_db_path = 'mit-bih-arrhythmia-database-1.0.0'
  annotations = ['A', 'V', 'N', 'L', 'R']
  qrs_interval_length = 320
  qrsintervals.extract_qrs_intervals_and_save(mitdbih_db_path, 'ecg_signals.mat', annotations)
  ...
  # Define the feature extraction parameters
  classes = 'A', 'V', 'N', 'L', 'R'
  signals_per_class = 2546
  wavelet = pywt.Wavelet('bior6.8')
  decomposition_level = 8
  # Extract features
  ext_feat_signals = preprocessing.extract_features(
      ECG_signals,         # The ECG signals from the ecg_signals.mat
      classes,             # The classification classes (A, V, N, L, R)
      signals_per_class,   # How many signals per class we have
      wavelet,         # The Wavelet
      decomposition_level  # The decomposition level for the Wavelet decomposition
  )
  ...
   

Feature Extraction

Features are extracted from each ECG signal using wavelet transformations. These features are then reduced to a lower-dimensional representation using Principal Component Analysis (PCA).

Classification

We apply three machine learning algorithms for heartbeat classification:

• K-Nearest Neighbors (KNN)
• Support Vector Machine (SVM)
• Neural Networks

Metrics used for comparison:

• Accuracy
• Precision
• Recall
• F1 Score

Results

The table below summarizes the performance of each classifier on the test data:

Algorithm	Accuracy	Precision	Recall	F1 Score
Nearest Neighbors	88.71%	88.77%	88.71%	88.63%
Linear SVM	92.41%	92.32%	92.41%	92.34%
RBF SVM	83.50%	87.50%	83.50%	84.17%
Neural Net	95.50%	95.48%	95.50%	95.46%

Installation and Usage

1. Clone the repository:

   git clone https://github.com/your-username/arrhythmia-beat-classification.git
   cd arrhythmia-beat-classification

2. Install the required dependencies:

   pip install -r requirements.txt

3. Run the notebook:

jupyter notebook Arrhythmia_Beat_Classification.ipynb

Contributors

• Houssem Khalfi
• Chaima Jeljli

Citations

Moody GB, Mark RG. The impact of the MIT-BIH Arrhythmia Database. IEEE Eng in Med and Biol 20(3):45-50 (May-June 2001). (PMID: 11446209)

Goldberger, A., et al. "PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation [Online]. 101 (23), pp. e215–e220." (2000).

Qin, Q., Li, J., Zhang, L. et al. Combining Low-dimensional Wavelet Features and Support Vector Machine for Arrhythmia Beat Classification. Sci Rep 7, 6067 (2017). https://doi.org/10.1038/s41598-017-06596-z