629 - Community use of digital auscultation to improve diagnosis of childhood pneumonia in Sylhet, Bangladesh
Sunday, April 30, 2023
3:30 PM – 6:00 PM ET
Poster Number: 629 Publication Number: 629.315
Salahuddin Ahmed, Projahnmo Research Foundation, Dhaka, Dhaka, Bangladesh; Harish Nair, University of Edinburgh, Edinburgh, Scotland, United Kingdom; Steve Cunningham, University of Edinburgh, Edinburgh, Scotland, United Kingdom; Eric D. McCollum, Johns Hopkins School of Medicine, Baltimore, MD, United States
Projahnmo Research Foundation Dhaka, Dhaka, Bangladesh
Background: The IMCI guidelines for pneumonia diagnosis have high sensitivity but low specificity. Although it is challenging for frontline healthcare workers to use a conventional acoustic stethoscope effectively, a digital stethoscope enhanced with an automated lung sound classification algorithm may be feasible. Objective:
Whether Community Health Workers (CHWs) can effectively record quality lung sounds
Agreement between digital stethoscope-recorded lung sound classifications generated from an automated machine learning algorithm with a consensus listening panel.
Design/Methods: Using a cross-sectional design, government CHWs recorded lung sounds from four chest positions using a novel digital stethoscope in under-5 children with cough or difficulty breathing at first-level community clinics in Bangladesh from November 2019 to December 2020. Each child’s recorded lung sounds were classified by two paediatricians and compared their classification, and if not matched, a third reader was the arbitrator. A quality recording was defined a priori as a child with three of four chest positions assessed as interpretable by the listening panel. An interpretable chest position was classified by the listening panel as any of the following: no wheeze and no crackle, wheeze only,crackle only, or wheeze and crackles. A machine learning algorithm classified recorded sounds into the same categories, which were compared with the panel classifications. Results: Of 2434 children screened, 990 were enrolled. The listening panel classified 867 children as having a quality recording (87.6%; 95% CI: 85.4%, 89.6%). CHWs recorded 75.7% (656/867) of quality recordings within three minutes. Compared to paediatricians, the sensitivity, specificity, and positive and negative predictive values of detecting abnormal sounds (wheeze/crackles) by the machine learning algorithm were 61.8 (95%CI: 55.7, 67.6), 60.7 (56.6, 64.6), 41.8 (36.9, 46.8), and 77.6 (73.6, 81.3) and 63.5 (54.5, 71.9), 66.2 (60.1, 73.1), 52.7 (44.4, 60.8), and 75.9 (69.2, 81.8) among children with IMCI defined pneumonia.
Conclusion(s): This study demonstrates government CHWs at rural, first-level clinics in Bangladesh can record quality lung sounds using a novel digital stethoscope without a substantial increase in their workload. This study also shows an automated algorithm had moderate sensitivity and specificity for classifying lung sounds as either abnormal or normal when using a paediatric listening panel as the reference. Agreement between the machine learning algorithm and paediatric listening panel modestly increased among children with IMCI-defined pneumonia.
