Deep learning for spirometry quality assurance with spirometric indices and curves.

Area of Science:

• Pulmonary Function Testing
• Artificial Intelligence in Healthcare
• Medical Diagnostics

Background:

• Spirometry quality assurance is a significant challenge, particularly in primary care settings.
• Deep learning (DL) offers a potential solution to enhance spirometry quality.
• Developing accurate and sensitive DL models is crucial for supporting high-quality spirometry.

Purpose of the Study:

• To develop a high-accuracy, sensitive deep learning-based model for spirometry quality assurance.
• To assist general practitioners (GPs) in achieving high-quality spirometry measurements.
• To generate actionable feedback, including warning messages and patient instructions.

Main Methods:

• A DL model was developed using 16,502 spirometry PDF files from four hospitals (Oct 2017-Oct 2020).
• Files were labeled according to ATS/ERS 2019 criteria and split into training, internal, and external test sets.
• The model assessed FEV1 and FVC acceptability, usability, and quality rating, generating system warnings and patient instructions for GPs.

Main Results:

• The DL model achieved high accuracy: 95.1% for FEV1 acceptability, 93.6% for FVC acceptability.
• Accuracy for usability and quality ratings for FEV1 and FVC ranged from 92.2% to 94.3%.
• GP performance improved significantly, with a ~21% increase in good quality FEV1 tests and ~36% for FVC tests.

Conclusions:

• The developed DL model effectively assists GPs in spirometry quality assurance.
• The model enhances GP performance in the quality control of spirometry tests.
• This AI-driven approach shows promise for improving diagnostic accuracy in primary care.