Respiratory disease in children is a major global concern. Clinicians traditionally rely on stethoscopes to detect abnormal lung sounds, but access to auscultation can be limited by the need for in-person visits.
Smartphones, with their built-in microphones and widespread availability, could support home-based respiratory screening. However, applying AI models trained on electronic stethoscope recordings to smartphone audio remains challenging because of a substantial domain mismatch caused by differences in how the devices capture sound.
To address this mismatch, a research team led by Associate Professor Seong-Eun Kim, including first author Seunggyu Jeong, a Ph.D. student, from the Department of Applied Artificial Intelligence, Seoul National University of Science and Technology in South Korea, has developed a domain-adaptive deep learning framework specifically designed for smartphone deployment. "To a neural network, the smartphone and electronic stethoscope effectively speak different languages," explains Dr. Kim. "Instead of simply pooling the data from the two devices, our approach offers reliable domain-adaption, where the model learns from reliable labelled-electronic stethoscope recordings while adapting to the smartphone domain." Their study was published in the IEEE Journal of Biomedical and Health Informatics on July 30, 2026.
The researchers first collected paired stethoscope-smartphone recordings. For the primary dataset, they acquired electronic stethoscope lung sound recordings for 611 children, aged 0–6 years. Of these, 147 also provided paired smartphone recordings. In both recordings, lung sounds were recorded at the same four anatomical sites. Notably, such paired datasets are rare. Additionally, the researchers collected an independent external dataset consisting of smartphone-only recordings from 48 children for validation.
To achieve effective domain adaptation during training, the researchers applied two complementary strategies. First, the smartphone audio was preprocessed using a 1 kHz low-pass filter and down sampled from 48 kHz to 4 kHz. "This brings the smartphone frequency profile closer to the stethoscope," notes Dr. Kim. Second, the model integrates a directional MixStyle module that injects feature statistics from the smartphone domain into the stethoscope domain, making stethoscope data appear more "smartphone-like". This module was then integrated into an Audio Spectrogram Transformer (AST) backbone, an audio classification deep learning architecture.
When tested on smartphone recordings in the primary dataset, the proposed model achieved an F1 score of 0.802 and an AUROC of 0.911 for detecting abnormal lung sounds. The smartphone-only model achieved an F1 score of 0.732, while naïve pooling of stethoscope and smartphone recordings achieved 0.708. These findings show that simply combining data from different devices does not necessarily improve performance when their acoustic characteristics differ.
On the independent external smartphone cohort, the stethoscope-only model achieved an F1 score of just 0.034 and an AUROC of 0.388, highlighting the challenge of applying models trained on one recording device to another.
Importantly, the proposed model maintained performance on the independent external smartphone cohort, achieving an F1 score of 0.744 and an AUROC of 0.862 and outperforming the other smartphone-domain approaches tested. These findings support the potential of the framework as an assistive tool for smartphone-based respiratory screening, although further real-world validation is needed before clinical deployment.
"Our framework provides a proof-of-concept for scalable and cost-effective pediatric respiratory screening," remarks Dr. Kim. "It could help parents home screen children between visits while keeping clinicians in the loop. Moreover, such technology could improve remote clinic consultations and aid community healthcare workers in frontline triage in remote areas."
The study demonstrates how signal-level alignment can reduce acoustic discrepancies between electronic stethoscopes and smartphones, providing a potential pathway toward more accessible respiratory screening and monitoring. Further validation with larger, multi-site cohorts, real-world home recordings, and multiple smartphone models will be needed.
