A CNN-MAMBA-Based Framework for Salient Bowel Sound Detection and Gastrointestinal Health Assessment

Area of Science:

• Gastroenterology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Constipation is a growing concern in the aging global population.
• Bowel sounds offer non-invasive acoustic data for gastrointestinal assessment.
• Automatic bowel sound analysis is complex due to signal sparsity and non-stationarity.

Purpose of the Study:

• To propose a two-stage framework for analyzing bowel sounds to detect and classify constipation.
• To evaluate the performance of deep learning models for these tasks.

Main Methods:

• A Convolutional Neural Network-MAMBA (CNN-MAMBA) model for salient bowel sound detection.
• A Convolutional Neural Network-Conformer-Multiple Instance Learning (CNN-Conformer-MIL) architecture for patient-level constipation classification using spectral representations.
• Continuous abdominal sound recordings and physician-annotated Bristol Stool Form Scale (BSFS) labels.

Main Results:

• The CNN-MAMBA detection model achieved 0.87 accuracy, 0.78 F1-score, and 0.93 ROC-AUC.
• Patient-level constipation classification yielded a mean accuracy of 0.665 and F1-score of 0.755 via cross-validation.
• Results indicate preliminary feasibility for classification, with detection performance being robust.

Conclusions:

• Bowel sound analysis shows promise as an auxiliary tool for screening and monitoring constipation in the elderly.
• Further research is needed to improve patient-level classification accuracy for standalone diagnostic use.
• The proposed framework demonstrates the potential of AI in analyzing acoustic gastrointestinal signals.