S1105 Unsupervised Machine Learning Based Automatic Detection and Characterization of Prominent Bowel Sounds: Feasibility Study in Healthy Subjects

Introduction: Non-Invasive gut health monitoring poses a pressing clinical need, based on which we reported our previous work demonstrating a You Only Listen Once (YOLO) deep learning model for automatic detection of prominent bowel sounds (BS) i.e., phonoenterogram (PEG). The YOLO model was developed based on expert manual annotations of prominent BS that is cumbersome and subjective. The purpose of this work was to develop an unsupervised machine learning platform to automatically annotate and classify BS recorded from healthy participants towards adaptive YOLO system for real-time PEG monitoring. Methods: 110 2-Minute PEGs sampled at 44.1 kHz were recorded using the EKO Duo® stethoscope from 8 healthy volunteers at 2 locations left upper quadrant and RLQ after IRB approval. PEG signals were processed to detect prominent BS events using an unsupervised model-free approach based on short-time energy (STE) and zero-crossing rate (ZCR). Next, we extracted multi-dimensional acoustic features from each prominent event, including Mel-frequency cepstral coefficients (MFCCs), spectral centroid, bandwidth, zero-crossing rate, and duration. We then applied KMeans clustering algorithm to organize events into clinically relevant categories. Clustering outputs were evaluated using internal metrics such as Silhouette Score with KMeans (k = 5) producing optimal separation (Silhouette ≈ 0.6). Results: From the full dataset, 42,975 total audio segments were extracted. Among these, 6,314 were labeled as prominent bowel sound events. Clustering revealed 5 reproducible morphologies across subjects from visual inspection as below: 1. Single Burst – brief, isolated, high-energy impulse 2. Harmonic – sustained waveforms with regular oscillations 3. Multiple Burst – sequences of impulse bursts 4. Random Continuous – irregular high-frequency sequences 5. Multi-Modal – overlapping or complex compound patterns. Conclusion: This study presents an unsupervised and scalable framework for characterizing bowel sounds using real-world recordings demonstrating the feasibility for non-invasive, label-free acoustic profiling of gastrointestinal activity and can serve as a foundation for clinical gastrointestinal monitoring applications for variety of bowel diseases.