Machine learning for prediction of key haemodynamic parameters in pulmonary arterial hypertension.

Abstract

Aims: Machine learning (ML) is increasingly recognized for its ability to identify and structure variables for predictive tasks. Pulmonary arterial hypertension (PAH) is a progressive disease characterized by elevated mean pulmonary arterial pressure (mPAP) and pulmonary vascular resistance (PVR) with normal pulmonary arterial wedge pressure (PAWP), as assessed by right heart catheterization (RHC). Despite increased awareness, delays between onset of non-specific symptoms and diagnosis continue to hinder early initiation of targeted therapies, leading to poorer outcomes. To develop and evaluate ML models for predicting key haemodynamic parameters in PAH, based on routinely available non-invasive data collected within 8 weeks prior to RHC, as a proof of concept.

Methods and results: We analysed data from 181 patients with invasively confirmed PAH, incorporating 56 variables, including demographics, echocardiography, blood gas analyses, 6-min walk distances, laboratory tests, and WHO functional class. An 80/20 train-test split and fivefold cross-validation were applied across multiple ML models, including least absolute shrinkage and selection operator (lasso) regression, ridge regression, k-nearest neighbours, decision trees, random forest, and gradient boosting machine. Lasso achieved best performance for predicting mPAP (r = 0.80, R² = 0.64, RMSE = 8.49). For PVR, ridge performed best (r = 0.71, R² = 0.51, RMSE = 3.60). Random forest and gradient boosting machines achieved modest but consistent performance for cardiac index (r = 0.38 and 0.37), while PAWP prediction remained limited across all models.

Conclusion: Machine learning models can estimate mPAP and PVR from routine clinical data obtained prior to RHC in patients with confirmed PAH. External validation is required to confirm generalizability and clinical applicability.