Prediction of haemodynamic parameters in pulmonary arterial hypertension using machine learning and non-invasive clinical data
T Kramer, H Weis, M Kramer, S Baldus, S Rosenkranz, S SpinlerAbstract
Background/Introduction
Non-invasive assessment of haemodynamic parameters in pulmonary arterial hypertension (PAH) remains challenging, while right heart catheterisation (RHC) is invasive and resource-intensive. Machine learning (ML) approaches may offer the potential to estimate key haemodynamic parameters using routinely collected clinical data.
Purpose
To develop and evaluate ML models for predicting key haemodynamic parameters in PAH using routinely available non-invasive clinical data collected within eight weeks prior to RHC, as an exploratory analysis.
Methods
Data from 181 patients with invasively confirmed PAH were analysed, incorporating 56 variables including demographics, echocardiography, blood gas analysis, six-minute walk distance, laboratory parameters and World Health Organization functional class. An 80/20 train–test split and five-fold cross-validation were applied across several ML models, including least absolute shrinkage and selection operator (lasso) regression, ridge regression, k-nearest neighbours, decision trees, random forest and gradient boosting machine.
Results
Lasso regression achieved the best performance for predicting mean pulmonary arterial pressure (mPAP), with a correlation coefficient of r = 0.80, R² = 0.64 and a root mean square error (RMSE) of 8.49. For pulmonary vascular resistance (PVR), ridge regression performed best (r = 0.71, R² = 0.51, RMSE = 3.60). Random forest and gradient boosting machine models showed modest but consistent performance for predicting cardiac index (CI), with correlation coefficients of r = 0.38 and r = 0.37, respectively. Prediction of pulmonary arterial wedge pressure (PAWP) remained limited across all models.
Conclusions
ML models can estimate mPAP and PVR from routinely collected, non-invasive clinical data obtained prior to RHC in patients with PAH. Prediction of CI is modest, while PAWP remains difficult to estimate. External validation is required to assess generalisability and the potential role of these models in supporting non-invasive haemodynamic assessment.