Accurately detecting abnormal breath sounds is vital in clinical pediatric medicine, as the nature and presence of pathological sounds guides diagnosis and initial treatment of common respiratory conditions. However, use of a standard binaural stethoscope by human practitioners to detect abnormal chest sounds introduces assessment subjectivity and research has shown that significant inter-listener variability exists [
1‐
3]. This calls into question the accuracy of diagnoses made on the basis of human auscultation. Treatment decisions informed by the diagnosis made may therefore be misguided, leading to unnecessary side effects and delay in provision of effective treatment.
In recent years, stethoscopes capable of digitally recording breath sounds have become more widely available, offering the ability to capture breath sounds with superior sound quality and fidelity [
4]. However, human interpretation of the digital recordings can still exhibit significant inter-listener variability [
5]. As the soundwave properties of pathologic breath sounds such as crackles, wheezes and rhonchi have been well-studied and previously defined, computer algorithms and programs to automatically detect them have been developed [
6,
7]. Increasingly, artificial intelligence (AI) algorithms have been applied in medicine, and because they have the capability of self-improvement as they learn from new data and cases, they can evolve to outperform traditional signal processing techniques [
8,
9]. AI programs based upon neural network programming have been used to identify melanomas from photographs of skin and suspicious soft tissue / calcified lesions on routine mammograms with an accuracy greater than most dermatologists and radiologists respectively, who were interpreting those images for performance comparison [
10,
11]. Similarly, an AI algorithm designed to detect abnormal pediatric breath sounds based upon several thousand patient recordings collected using a digital stethoscope (DS) was reported to outperform pediatricians, especially in coarse crackle detection [
12]. However, the study, which was primarily conducted by developers of the technology, utilized breath sounds collected using only the StethoMe DS. Therefore, we aimed to establish the performance of the AI algorithm in detecting pathological pediatric breath sounds collected using other DS devices, to evaluate the algorithm’s generalizability. We used blinding techniques and real-world recordings to maximize the validity and applicability of our study.