Presenter of 1 Presentation
O069 - REAL-TIME COMPUTERIZED ANALYSIS OF AUSCULTATION FOR POINT-OF-CARE PEDIATRIC PNEUMONIA DIAGNOSIS IN FRONTLINE CLINICAL ENVIRONMENTS (ID 897)
Abstract
Background
Frontline workers use World Health Organization guidelines for child pneumonia care in resource-limited settings, which prioritize sensitivity over specificity and result in antibiotic overtreatment. Chest radiography (CXR) and other imaging tools are not available in many clinical contexts, and without standardized training, have high variability. Chest auscultation offers a non-invasive and low-cost tool for improving pneumonia diagnosis but is undermined by the need for trained listeners, inter-listener variability, subjectivity, and vulnerability to noise.
Methods
Feelix, a digital auscultation tool with onboard algorithms, is presented to improve the accuracy and speed of pneumonia diagnosis. Lung sounds are captured at various positions on a patient, using real-time noise suppression schemes to eliminate ambient sounds and sensor motion artifacts. High-quality signals are then mapped onto a rich spectrotemporal feature space before undergoing classification of the presence/lack of pneumonia using a novel lightweight deep learning model. The model is trained and benchmarked against a dataset sourced from a pneumococcal vaccine effectiveness study of children aged 3-35 months in Sylhet, Bangladesh, with 97 patients diagnosed with primary endpoint pneumonia (PEP) and 355 without PEP (normal or other infiltrates), based on a CXR adjudication panel.
Results
The classification of auscultation signals is shown to achieve an accuracy of 90.9% in differentiating PEP from non-PEP cases, with a sensitivity of 100% and specificity of 88.46%.
Conclusions
This process is designed to run in real-time, can be widely deployed in any clinical setting, can provide a classification in seconds, and would dramatically improve the diagnostic accuracy and speed of frontline health care workers.