Late Breaking Abstract - Accuracy of artificial intelligence in detecting pathological breath sounds in children using digital stethoscopes

Abstract

Background: Manual auscultation to detect abnormal breath sounds has poor inter-observer reliability. Digital stethoscopes (DS) with artificial intelligence (AI) could improve reliable detection of these sounds.

Objective: We aimed to independently test the abilities of AI developed for the purpose of detecting wheezes/rhonchi and crackles in children.

Methods: 192 auscultation recordings collected from children using two different DS (Clinicloud™ and Littman™) were each tagged as containing wheezes/rhonchi, crackles or neither by a paediatric respiratory physician, based on audio playback and careful spectrogram and waveform analysis. Untagged versions of the recordings were submitted for analysis by a blinded AI algorithm (StethoMe™ AI) trained to detect pathologic paediatric breath sounds, which generated a probability score of the likelihood of presence of crackles or wheeze/rhonchi. AI outcome was compared with tagged outcomes on a per-recording basis, with receiver operating characteristic curves used to identify optimal cutoffs representing best AI performance.

Results: With optimised AI thresholds, crackle detection positive percent agreement (PPA) was 0.95 and negative percent agreement (NPA) was 0.99 for Clinicloud recordings; for Littman-collected sounds PPA was 0.82 and NPA was 0.96. Wheeze detection PPA and NPA were 0.90 and 0.97 respectively (Clinicloud auscultation), with PPA 0.80 and NPA 0.95 for Littman recordings.

Conclusions: AI can detect crackles and wheeze from breath sounds obtained using different DS devices with a degree of accuracy that approaches (or exceeds) that of clinicians. Careful integration into clinical practice may improve standards of care.