Researchers have developed a domain-adaptive deep learning framework for smartphones, used for detecting abnormal lung sounds.

Seoul, South Korea, October 16, 2026 / PRNewswire / -- Respiratory diseases in children are an important issue of global concern. Although stethoscopes can help clinicians identify abnormal lung sounds, listening to the lungs usually requires a face-to-face consultation, which may limit the availability of such examinations.

Smartphones can potentially be used for home respiratory screening, but due to differences in sound collection methods, there are still challenges in directly applying artificial intelligence models trained on recordings from electronic stethoscopes to smartphone audio.

An research team led by Associate Professors Kim Seong to Eun from the Department of Applied Artificial Intelligence at the National University of Science and Technology in Seoul, South Korea, has developed a domain-adaptive deep learning framework designed specifically for deployment on smartphones. The team members include Jeong Seunggyu, who is also the first author of the paper. Kim stated, "For neural networks, smartphones and electronic stethoscopes are essentially like speaking two different languages. Our approach is not to simply mix the data from these two devices together, but to perform reliable domain adaptation: the model learns from recordings made with electronic stethoscopes that come with reliable labels, while also adapting to the smartphone domain." This research was published on July 30, 2026, in IEEE Journal of Biomedical and Health Informatics.

Researchers collected paired stethoscope-smartphone recordings from 611 children aged 0 to 6 years, among which 147 children had paired smartphone recordings. Pulmonary sounds were collected at the same four anatomical sites. Such paired datasets are relatively rare. In addition, the team also collected an independent external validation dataset that contained only smartphone recordings from 48 children.

To achieve effective domain adaptation during the training process, researchers adopted two complementary strategies. Firstly, they used a 1 kHz low-pass filter to pre-process the smartphone audio and downsampled the sampling rate from 48 kHz to 4 kHz. Kim noted: "This makes the frequency characteristics of the smartphone audio more similar to those of a stethoscope." Secondly, the model incorporated a directional MixStyle module that injected statistical information from the smartphone domain into the stethoscope domain, making the stethoscope data sound more "like smartphone recordings." Subsequently, this module was integrated into the Audio Spectrogram Transformer ( AST ) backbone network, which is a deep learning architecture used for audio classification.

In the smartphone recording tests of the main dataset, the model achieved a F1 score of 0.802 and a AUROC score of 0.911 in detecting abnormal lung sounds. In contrast, the model that used only a smartphone had a F1 score of 0.732, while the approach of simply mixing stethoscope recordings with smartphone recordings had a F1 score of 0.708. The research results indicate that when there are differences in the acoustic characteristics of different devices, simply merging the data does not necessarily improve performance.

In an independent external smartphone queue, the model F1 trained solely using a stethoscope scored only 0.034, while AUROC scored 0.388, highlighting the difficulty of applying a model trained on one type of recording device to another.

The model still maintained its performance in the independent smartphone queue, achieving a F1 score of 0.744 and a AUROC score of 0.862, which is superior to other testing methods. Researchers stated that these results indicate that this method has potential in smartphone-based respiratory screening, but further real-world validation is still needed.

Kim indicates: "Our framework provides proof of concept for scalable and low-cost pediatric respiratory screening. It can help parents conduct preliminary screenings for their children at home between two medical visits, while also maintaining continuous involvement of clinicians. In addition, such technology may also improve remote outpatient consultations and assist frontline community healthcare workers in remote areas with triage."

Researchers say that alignment at the signal level may help to utilize smartphones for more accessible respiratory screening and monitoring, but further verification is still needed in larger-scale, more diverse populations, as well as in real-world environments.

Reference information

About the institution: Seoul National University of Science and Technology ( SEOULTECH )

Contact Person: