Abstract 15889: An Artificial Intelligence-Derived Pediatric Pulmonary Hypertension Risk Prediction Model From the Pediatric Pulmonary Hypertension Network (PPHNet) Registry

Abstract

Introduction: Risk prediction is fundamental to guiding pulmonary hypertension (PH) treatment and is a key part of PH guideline based care. Data-driven risk prediction in pediatric PH has been lacking, relying instead on expert opinion. Objective: Use the multi-center Pediatric Pulmonary Hypertension Network (PPHNet) registry to develop a pediatric PH risk model. Methods: Pediatric PH patients (WSPH 1/3) enrolled in the PPHNet were used in this study and randomly split into train (80%) and test cohorts (20%). 176 variables were suitable for analysis including demographics, imaging, hemodynamics, functional/exercise testing, laboratory values, symptoms, and comorbidities. Boruta feature selection with random forest classifier identified variables that predicted the composite outcome of time to death, transplant or reverse Potts shunt. 2 The prediction model was built using an extreme gradient boost classifier with parameter tuning using Bayesian optimization. Model performance was quantified using confusion matrices, calculated sensitivity, specificity, AUC-ROC, and calibration slope. Results: A total of 1232 patients (985 Train, 247 test) were appropriate for inclusion; 134 patients (10.9%) met the composite outcome) with follow-up to 4 years from diagnosis. 37 variables were isolated for the prediction model. The Model performed strongly in the test cohort with an AUROC of 87% (Figure 1A), sensitivity of 85%, and specificity of 77%. Predicted probabilities were stratified into 3 major risk groups based on 1-year risk of event (Figure 1B; low <5%, medium 10-15%, high >15%) with good agreement between predicted and observed outcomes. Conclusions: Using the multicenter PPHNet cohort, we developed a well calibrated 1-year risk prediction model for pediatric groups 1 and 3 PH. This model is uniquely trained on incident and prevalent pediatric patients and could enable accurate identification of high-risk patients in need of aggressive therapy.