Smartphone-based lung auscultation, combined with artificial intelligence (AI), may offer a promising alternative for pediatric respiratory telemonitoring. The aim of our study was to evaluate the performance of an AI model on wheeze detection from pediatric respiratory sounds recorded via smartphone. An observational cross-sectional study was conducted at the Pediatric Department of a Portuguese tertiary hospital, including children aged 0–17 years (pre-school, school-aged, and adolescent), with or without respiratory disease. Respiratory sounds were recorded at four locations using smartphone microphones and independently classified by at least 2 blinded annotators for quality and presence of wheezes. A hybrid convolutional–recurrent neural network (CNN + LSTM) trained on public electronic stethoscope databases was used to detect wheezes. AI model performance was assessed using positive predictive value, sensitivity, specificity, accuracy, and F1-score. A total of 217 children (59.9% male; median 10 [Q1-Q3 4.5–13] years) were included. From 2020 respiratory sound recordings, 1500 (74.3%) met quality criteria. Manual annotation identified 271 recordings with wheezes, whereas AI detected 217, most frequently in pre-school children (64.5% AI; 53.9% manual). The model achieved an overall accuracy of 87% (95% CI 86—89) and an F1-score of 61% (95% CI 56—66). The best accuracy was found in adolescents (92%, 95% CI 90—94) and the best global F1-score in pre-school children recordings (64%, 95% CI 58–71).Conclusion: This study supports the feasibility of AI-assisted analysis of smartphone-recorded pediatric respiratory sounds in real-world settings. Future multicenter studies with larger datasets and model fine-tuning using smartphone recordings are expected to improve AI model performance and generalizability.<Table Float="No" ID="Taba"> <tgroup cols="1"> <colspec align="left" colname="c1" colnum="1" /> <tbody> <row> <entry align="left" colname="c1"> What is Known: • Continuous monitoring of pediatric respiratory diseases is essential to improve outcomes and reduce healthcare and family burden. • Smartphone lung auscultation is feasible in children  but AI wheeze detection models have only been applied to recordings from electronic stethoscopes. </entry> </row> <row> <entry align="left" colname="c1"> What is New: • First study applying an AI model to detect wheezes in smartphone-recorded lung sounds. • Quality recordings were obtained across all pediatric age groups and the AI model performance showed promising results, comparable to studies using electronic stethoscopes. </entry> </row> </tbody> </tgroup> </Table>

Wheeze detection in real-world pediatric care: AI applied to smartphone lung auscultation

期刊论文
出版模式：

开放获取
发表日期：

2026年5月12日

查看全文

Inês Pais-Cunha,
Diogo Pessoa,
Maria Catarina Silva,
Henrique Ferreira-Cardoso,
Bruno Rocha,
Rui Pedro Paiva,
Paulo de Carvalho,
João A. Fonseca,
Inês Azevedo,
Cristina Jácome

摘要

Smartphone-based lung auscultation, combined with artificial intelligence (AI), may offer a promising alternative for pediatric respiratory telemonitoring. The aim of our study was to evaluate the performance of an AI model on wheeze detection from pediatric respiratory sounds recorded via smartphone. An observational cross-sectional study was conducted at the Pediatric Department of a Portuguese tertiary hospital, including children aged 0–17 years (pre-school, school-aged, and adolescent), with or without respiratory disease. Respiratory sounds were recorded at four locations using smartphone microphones and independently classified by at least 2 blinded annotators for quality and presence of wheezes. A hybrid convolutional–recurrent neural network (CNN + LSTM) trained on public electronic stethoscope databases was used to detect wheezes. AI model performance was assessed using positive predictive value, sensitivity, specificity, accuracy, and F1-score. A total of 217 children (59.9% male; median 10 [Q1-Q3 4.5–13] years) were included. From 2020 respiratory sound recordings, 1500 (74.3%) met quality criteria. Manual annotation identified 271 recordings with wheezes, whereas AI detected 217, most frequently in pre-school children (64.5% AI; 53.9% manual). The model achieved an overall accuracy of 87% (95% CI 86—89) and an F1-score of 61% (95% CI 56—66). The best accuracy was found in adolescents (92%, 95% CI 90—94) and the best global F1-score in pre-school children recordings (64%, 95% CI 58–71).

Conclusion: This study supports the feasibility of AI-assisted analysis of smartphone-recorded pediatric respiratory sounds in real-world settings. Future multicenter studies with larger datasets and model fine-tuning using smartphone recordings are expected to improve AI model performance and generalizability.

What is Known:

• Continuous monitoring of pediatric respiratory diseases is essential to improve outcomes and reduce healthcare and family burden.

• Smartphone lung auscultation is feasible in children but AI wheeze detection models have only been applied to recordings from electronic stethoscopes.

What is New:

• First study applying an AI model to detect wheezes in smartphone-recorded lung sounds.

• Quality recordings were obtained across all pediatric age groups and the AI model performance showed promising results, comparable to studies using electronic stethoscopes.