Benchmarking machine learning for bowel sound pattern classification - From tabular features to pretrained models.

Area of Science:

• Biomedical Engineering
• Computational Health
• Gastroenterology

Background:

• Automated analysis of bowel sounds (BS) is now possible with electronic stethoscopes and wearable sensors.
• This allows for data-driven insights into BS patterns, their relationships, and links to diseases.

Purpose of the Study:

• To evaluate machine learning models for detecting and classifying BS patterns.
• To compare the performance of different model types, including those using tabular data, CNNs, and pre-trained audio models.

Main Methods:

• A dataset from 16 healthy subjects with annotated BS patterns was used.
• Models evaluated included those using tabular features, CNNs on spectrograms, and pre-trained models (HuBERT, Wav2Vec 2.0).

Main Results:

• Pre-trained models significantly outperformed others, especially for classes with limited samples.
• HuBERT achieved an AUC of 0.89 for BS vs. non-BS detection.
• Wav2Vec 2.0 achieved an AUC of 0.89 for differentiating BS patterns.

Conclusions:

• Pre-trained models demonstrate superior performance in BS analysis.
• These findings support the development of machine learning-driven diagnostic applications for gastrointestinal examinations.