Validation of artificial intelligence spirometry diagnostic support software in primary care: a blinded diagnostic...

Area of Science:

• Medical Informatics
• Pulmonology
• Artificial Intelligence

Background:

• Primary care spirometry is crucial for diagnosing respiratory diseases.
• Accurate interpretation of spirometry data can be challenging.
• Artificial intelligence (AI) offers potential for improving diagnostic accuracy.

Purpose of the Study:

• To evaluate the discriminative accuracy of AI software in identifying COPD and other chronic respiratory diseases using primary care spirometry data.
• To compare AI-based diagnoses with expert pulmonologist consensus.

Main Methods:

• A diagnostic study utilizing retrospective primary care spirometry data from 1113 patients.
• AI software (supervised random-forest machine learning) interpreted raw spirometry and demographics for index diagnosis.
• Reference diagnosis established by expert pulmonologists reviewing comprehensive medical records.

Main Results:

• AI achieved high accuracy for COPD detection (85.4%), with sensitivity of 84.0% and specificity of 86.8% (AUC=0.914).
• AI demonstrated strong performance for interstitial lung disease (AUC=0.900) and asthma (AUC=0.814).
• AI software performed less effectively for other chronic respiratory disease categories compared to COPD.

Conclusions:

• AI software demonstrates high sensitivity and specificity for COPD diagnosis in primary care settings.
• AI may serve as a valuable tool to support accurate COPD diagnosis.
• AI performance is less robust for diagnosing other chronic respiratory diseases.