A cross-disciplinary team from Sichuan University's NICUs employs a machine learning pipeline to classify neonatal intestinal diseases using bowel sound recordings captured by a digital stethoscope. They preprocess audio with filters, extract time–frequency features such as MFCCs, and train a transformer-based model combined with a Random Forest to detect conditions like NEC, FPIAP, and obstruction, aiming to supplement subjective clinical assessment with objective, automated diagnostics.

Key points

• Collected neonatal bowel sounds via 3M Littmann 3200 digital stethoscope with 2-minute recordings from six abdominal regions, filtered to exclude noise exceeding 30%.
• Extracted acoustic features—zero-crossing rate, spectral centroid, chroma, MFCCs—after pre-emphasis, framing, and Hamming windowing, forming a multidimensional feature vector.
• Trained a Random Forest for disease detection and a transformer-based network for multi-class classification (NEC, FPIAP, volvulus, obstruction), validated via tenfold cross-validation and external cohorts with high AUC.

Why it matters: An AI-based bowel sound diagnostic tool offers rapid, noninvasive neonatal intestinal disease screening, potentially reducing delays and improving outcomes compared with subjective auscultation.