Interpretable Machine Learning Model for Pulmonary Hypertension Risk Prediction: Retrospective Cohort Study

Area of Science:

• Cardiology
• Medical Imaging
• Machine Learning

Background:

• Pulmonary hypertension (PH) is a progressive condition leading to right heart failure.
• Early PH detection is crucial for better patient outcomes.
• Current diagnostic methods like right heart catheterization are invasive.

Purpose of the Study:

• To identify key predictors of PH from echocardiographic, laboratory, and demographic data.
• To develop a machine learning-based predictive model for early, noninvasive PH diagnosis.
• To improve upon the limitations of invasive diagnostic procedures.

Main Methods:

• Utilized comprehensive datasets of echocardiography, lab tests, and demographics from PH patients and controls.
• Employed Recursive Feature Elimination for echocardiographic variable selection.
• Applied XGBoost and LASSO regression for feature selection and model construction.
• Validated the predictive model using ROC curves, calibration plots, and decision curve analysis.

Main Results:

• Identified 16 echocardiographic parameters and 2 lab biomarkers (prothrombin time activity, cystatin C) as optimal predictors.
• The developed PH prediction model achieved high accuracy (AUC 0.997 internal, 0.974 external validation).
• Model performance was confirmed through rigorous validation, demonstrating clinical applicability.

Conclusions:

• The noninvasive predictive model significantly enhances early PH diagnosis.
• High accuracy facilitates timely intervention and personalized treatment strategies.
• Potential applications extend to broader cardiovascular disease management.