Deep learning for spirometry quality assurance with spirometric indices and curves

BACKGROUND: Spirometry quality assurance is a challenging task across levels of healthcare tiers, especially in primary care. Deep learning may serve as a support tool for enhancing spirometry quality. We aimed to develop a high accuracy and sensitive deep learning-based model aiming at assisting high-quality spirometry assurance. METHODS: Spirometry PDF files retrieved from one hospital between October 2017 and October 2020 were labeled according to ATS/ERS 2019 criteria and divided into training and internal test sets. Additional files from three hospitals were used for external testing. A deep learning-based model was constructed and assessed to determine acceptability, usability, and quality rating for FEV(1) and FVC. System warning messages and patient instructions were also generated for general practitioners (GPs). RESULTS: A total of 16,502 files were labeled. Of these, 4592 curves were assigned to the internal test set, the remaining constituted the training set. In the internal test set, the model generated 95.1%, 92.4%, and 94.3% accuracy for FEV(1) acceptability, usability, and rating. The accuracy for FVC acceptability, usability, and rating were 93.6%, 94.3%, and 92.2%. With the assistance of the model, the performance of GPs in terms of monthly percentages of good quality (A, B, or C grades) tests for FEV(1) and FVC was higher by ~ 21% and ~ 36%, respectively. CONCLUSION: The proposed model assisted GPs in spirometry quality assurance, resulting in enhancing the performance of GPs in quality control of spirometry. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12931-022-02014-9.