Identification of adventitious lung sounds in children with respiratory illnesses in Bangladesh using Artificial Intelligence and Digital Auscultation

Integrated management of childhood illness (IMCI) guidelines have high sensitivity but low specificity for pneumonia diagnosis. Artificial intelligence (AI)-enabled digital stethoscopes capable of analyzing lung sounds may improve IMCI diagnostic performance. We evaluated the performance of an AI algorithm trained to identify normal and adventitial lung sounds in children. Non-physician health workers recorded lung sounds from four chest positions using a digital stethoscope in under-five-year-old children with suspected pneumonia at community clinics in Bangladesh. A trained paediatrician listening panel classified chest position recordings as normal, abnormal (crackles and/or wheeze), or uninterpretable. The AI algorithm similarly classified chest position recordings except for the uninterpretable category. AI algorithm and listening panel comparisons were made at the child-level, and included chest positions considered interpretable, uninterpretable, and high- and low-confidence by the panel. Of 990 enrolled children, 867 (87%) had at least 3 interpretable chest position recordings by the panel and were analyzed. Compared to the panel, AI algorithm sensitivity and specificity for detecting abnormal sounds were 61.8% and 60.7% among all children, and 63.5% and 66.8% in IMCI pneumonia cases. Overall, the AI algorithm achieved moderate classification performance. Classification performance will benefit from further AI algorithm training to categorize recordings as uninterpretable.