Detection of pulmonary hypertension in idiopathic pulmonary fibrosis using random forest models and automated measures of central computed tomography structures.

Objectives: Since pulmonary hypertension (PH) worsens the prognosis of idiopathic pulmonary fibrosis (IPF), early prediction of PH is crucial for timely intervention. This study aims to develop and validate a machine learning model to predict PH using automated computed tomography (CT)-based 3D measurements, particularly central cardiovascular structures, segmented by a publicly available tool.

Methods: We retrospectively studied 163 IPF patients who underwent both thin-section chest CT (contrast-enhanced and non-contrast) and right heart catheterisation within 2 years (78.5% within 6 months). Central CT structures were segmented using the TotalSegmentator Neural Network Version 1. We also manually measured pulmonary artery (PA) and ascending aorta (A) diameters. Random forest (RF) and logistic regression (LR) models were created and the model's reliability was assessed with 10-fold cross-validation. Shapley additive explanation (SHAP) analysis was performed to understand the contribution of each variable to the RF model.

Results: Of the 163 patients, 75 had PH (46.0%). Significant differences were found in race, body mass index, right atrial (RA) volume, and PA volume between PH and non-PH patients. The RF model outperformed the LR model, showing higher area under the curve (AUC) (0.87 versus 0.82). Replacing PA volume with the PA/A ratio in the RF model decreased performance (AUC: 0.87 versus 0.79). SHAP identified PA and RA volumes as key features. No significant differences were observed between mean pulmonary arterial pressure and RA or PA volume in non-contrast CT compared to contrast-enhanced CT.

Conclusion: The RF model with volumetric measures showed superior predictive performance for PH. Notably, both the RF model and segmentations of central CT structures are automated, facilitating seamless integration into clinical practice.