Benchmarking machine learning for bowel sound pattern classification – from tabular features to pretrained models

This is the official code repository associated with the paper:

📄 Benchmarking machine learning for bowel sound pattern classification – from tabular features to pretrained models

✍️ Zahra Mansour, Verena Uslar, Dirk Weyhe, Danilo Hollosi and Nils Strodthoff. ArXiv. https://arxiv.org/abs/2502.15607

This repository contains a Bowel sounds classification pipeline that supports both Deep Learning and Machine Learning models for audio classification. The pipeline is built using PyTorch, Hugging Face Transformers, and Scikit-Learn.

Features

Supports multiple model architectures: - Deep Learning Models: VGG, ResNet, AlexNet, CNN-LSTM - Pre-trained models: Wav2Vec, HuBERT - Machine Learning Models: SVM, XGBoost, KNN, Decision Tree (DTC), CatBoost

Automatic feature extraction: - Extracts spectrogram, log-mel, MFCC, or raw waveform for deep learning. - Supports GeMAPS & ComParE feature extraction for machine learning.

Stratified Data Splitting: - Ensures class distribution remains balanced across train, validation, and test sets.

Configurable settings using config.yaml: - Easily change dataset paths, model type, and hyperparameters without modifying the code.

Automatic Model Training & Evaluation: - Computes accuracy, F1-score, confusion matrix, and AUC (Area Under Curve) for evaluation.

Usage

• Install dependencies sh pip install torch torchvision torchaudio transformers pandas scikit-learn pyyaml
• For Machine and deep learning models: after choosing the features and models by updating the config file, Simply run: sh python main.py
• For finetuning pretrained models: sh python pre_trained_Wav2Vec.py python pre_trained_HuBERT.py This will:
• Load the dataset from config.yaml

• Split the dataset into train/validation/test sets

• Extract features

• Train the specified model

• Evaluate model performance

Data Preparation

Your dataset should be in CSV format with the following columns:

| Column Name | Description | |-------------|------------| | path | File path to the audio file in .wav formate | | label | Class label for the audio sample (for bowel sound patterns: SB, MB, CRS, HS, and Silence period labelled NONE) | | patent_id | Identifier for subject grouping |

Data Availability

The dataset used in this study, consisting of recordings from four subjects, is publicly available at the following link:

Dataset on Figshare

Example CSV File

Below is an example of how your dataset should look:

Bowel Sound patterns Classification (BS_segments.csv)

| path | label | patentid | |-------------|--------|-----------| | /101SGPTsegment1.wav | SB| 101 | | /102SGPTsegment1.wav | MB | 102 | | /103SGPTsegment1.wav| NONE | 103 |