Development and External Validation of a Machine Learning Model to Predict Restriction from Spirometry.

Area of Science:

• Pulmonary Medicine
• Medical Informatics
• Machine Learning

Background:

• Current European Respiratory Society and American Thoracic Society (ERS/ATS) guidelines use forced vital capacity (FVC) lower limit of normal (LLN) to exclude restriction.
• Recent data indicate a lower negative predictive value (NPV) for FVC LLN, particularly in non-Hispanic Black patients.
• This study addresses the need for improved accuracy and equity in pulmonary function test (PFT) interpretation.

Purpose of the Study:

• To develop and externally validate a machine learning (ML) model to predict restriction from spirometry.
• To determine if ML models can improve the accuracy and equity of PFT interpretation.

Main Methods:

• Utilized PFTs with static and dynamic lung volume measurements from two health systems.
• Trained logistic regression, random forest, and boosted tree models using demographic, anthropometric, and spirometric data.
• Externally validated models using data from a separate health system, assessing NPV and racial equity.

Main Results:

• All three ML models outperformed the FVC LLN in predicting restriction.
• The random forest model demonstrated a significantly higher overall NPV (88.3%) compared to FVC LLN (72.7%).
• The random forest model showed improved NPV for both non-Hispanic Black (74.6%) and non-Hispanic White (90.9%) patients compared to FVC LLN.

Conclusions:

• ML models enhance the accuracy and equity of PFT interpretation for excluding restriction.
• While ML models improve upon current guidelines, they do not entirely eliminate observed racial differences in prediction.
• Further research may be needed to fully address racial disparities in PFT interpretation.