Peer-reviewed article
The Pediatric Pulmonary Hypertension International Risk Score: A Prediction Model for Outcomes Using Machine Learning.
Griffiths M, Chinni BK, Lokhorst C, Douwes JM, Sleeper LA et al. ·
Summary
Written by AI, not yet AICS's view · claude-haiku-4-5-20251001This peer-reviewed study describes the development and validation of a machine learning prediction model for children with pulmonary hypertension, a serious lung disease. The model identifies which children are at higher risk of poor outcomes within one year using 16 clinical variables. The model performed well in testing and validation across multiple registries. However, this is a technical medical tool development study focused on disease prognosis, not on child well-being or interventions to improve children's lives.
Key points
- Machine learning model achieved 90% accuracy in test cohort for predicting 1-year outcomes in pediatric pulmonary hypertension
- External validation across Dutch and Spanish registries showed 76-77% accuracy
- Model identified three risk groups with significantly different outcomes
- 16 clinical variables selected from 176 candidate predictors
- Negative predictive value of 93-99% across cohorts
Limitations
This document describes a diagnostic/prognostic tool development study, not an intervention to improve child health outcomes or well-being. It cannot tell us whether using this model in clinical practice actually improves children's health, survival, or quality of life.
Abstract
Background Risk prediction is fundamental to pulmonary hypertension (PH) guideline-based care, yet pediatric-specific risk prediction models remain limited, relying primarily on single predictors, expert opinion, or application of adult models to children. The authors developed and externally validated a data-driven 1-year risk prediction model for pediatric PH. Methods Pediatric patients with PH (n=345; World Symposium on Pulmonary Hypertension groups 1 and 3) enrolled in the Pediatric Pulmonary Hypertension Network Registry (2014-2020; 50.4% male; median age, 4.9 years [interquartile range, 1.9-10.3]) were split into training (80%) and test cohorts (20%). The Dutch National Registry for Pulmonary Hypertension in Childhood (n=155 [1993-2020]) and the Spanish Registry of Pediatric Pulmonary Hypertension (n=327 [2009-2023]) were used for external validation. From 176 variables, BorutaSHAP feature selection with random forest identified 16 predictors for a 1-year outcome of time to death, transplant, Potts shunt, or atrial septostomy, modeled using extreme gradient boosting. Performance was assessed with the area under the receiver operating characteristic curve, confusion matrices, calibration, and Kaplan-Meier event-free survival. Results The final model achieved an area under the receiver operating characteristic curve of 0.90 (0.79-0.97) and 99% (96%-99%) negative predictive value in testing, dividing participants into 3 groups with strong outcome discrimination. External validation showed an area under the receiver operating characteristic curve of 0.76 (Dutch National Registry for Pulmonary Hypertension in Childhood, 0.70-0.81) and 0.77 (Spanish Registry of Pediatric Pulmonary Hypertension, 0.73-0.82) with negative predictive values of 93% (93%-97%) and 96% (93%-97%), respectively. Kaplan-Meier analysis significantly differentiated outcomes by risk group. Conclusions This multicenter, validated model provides good 1-year risk prediction in pediatric PH across World Symposium on Pulmonary Hypertension groups 1 and 3, providing a robust tool for clinical risk stratification to guide therapy and addressing a gap in pediatric PH care.