Development of machine learning model to predict pulmonary function with low-dose CT-derived parameter response...

Area of Science:

• Pulmonary Medicine
• Radiology
• Artificial Intelligence

Background:

• Pulmonary function tests (PFTs) are crucial for diagnosing lung diseases like COPD.
• Low-dose computed tomography (LDCT) can provide detailed lung imaging.
• Parametric response mapping (PRM) quantifies lung tissue changes on CT scans.

Purpose of the Study:

• To develop and assess a machine learning (ML) model using LDCT-derived PRM for predicting PFT results.
• To evaluate the model's ability to classify individuals into normal, high-risk, and COPD categories.

Main Methods:

• Retrospective analysis of 615 subjects (40-74 years) with PFT and LDCT data.
• Calculation of 72 PRM parameters including emphysema and small airways disease.
• Construction and testing of random forest regression and MLP models for PFT prediction and classification.

Main Results:

• The random forest ML model demonstrated superior PFT prediction (R² of 0.749 for FEV1/FVC, 0.792 for FEV1%).
• High accuracy in differentiating normal from high-risk groups (88%) and non-COPD from COPD groups (99%).
• The model achieved high sensitivity and specificity in distinguishing between these groups.

Conclusions:

• An ML-based random forest model utilizing PRM effectively predicts PFT outcomes from LDCT data.
• The model reliably identifies individuals at high risk for COPD within a community screening setting.
• This AI-driven approach shows promise for early detection and risk stratification of lung disease.