Heart Sound Classification

This project implements a heart sound classification system using 1D Local Binary Patterns (1D-LBP), 1D Local Ternary Patterns (1D-LTP), and a 1D Convolutional Neural Network (CNN). It is based on the research paper "Heart sounds classification using CNN with 1D-LBP and 1D-LTP features" by Er, Mehmet Bilal.

Features

• Feature Extraction: Uses 1D-LBP and 1D-LTP for robust texture feature extraction from audio signals.
• Feature Selection: Implements ReliefF algorithm to select the most relevant features.
• Classification: Uses a 1D-CNN to classify heart sounds into categories.
• Multiple Datasets: Supports both PASCAL and Physionet2016 datasets.

Datasets

The project supports the following datasets:

1. PASCAL Classifying Heart Sounds Challenge 2011
   • Classes: Normal, Murmur, Artifact, Extrahls
2. PhysioNet/Computing in Cardiology Challenge 2016
   • Classes: Normal, Abnormal

Usage

Training the Model

You can train the model using either the PASCAL or Physionet dataset.

Option 1: Using main.py (Interactive or Command Line)

Run interactively to select the dataset:

python main.py

Or specify arguments directly:

# Train on PASCAL dataset
python main.py --dataset PASCAL --epochs 140

# Train on Physionet dataset
python main.py --dataset Physionet --epochs 140

Option 2: Using train.py (Advanced)

# Train on PASCAL with 10-fold CV
python train.py --mode kfold --dataset PASCAL --epochs 140

# Train on Physionet with 10-fold CV
python train.py --mode kfold --dataset Physionet --epochs 140

# Quick simple training (no CV)
python train.py --mode simple --dataset PASCAL

Making Predictions

To classify a heart sound recording (WAV file):

python predict.py path/to/your/audio_file.wav

Example:

python predict.py PASCAL/Atraining_normal/201101070538.wav

Project Structure

• data_preprocessing.py: Handles feature extraction (LBP, LTP) and ReliefF selection.
• model.py: Defines the 1D-CNN model architecture.
• train.py: Core training logic.
• main.py: Main entry point with dataset selection.
• predict.py: Script for making predictions on new audio files.
• config.py: Configuration settings.
• PASCAL/: Directory for PASCAL dataset.
• Physionet2016/: Directory for Physionet dataset.

Citation

If you use this work, please cite the original paper:

Er, Mehmet Bilal. "Heart sounds classification using CNN with 1D-LBP and 1D-LTP features." Applied Acoustics 180 (2021): 108152.