Interpretable machine learning-enabled risk stratification for pulmonary arterial hypertension associated with unrepaired congenital heart disease: results from a national prospective registry.

BACKGROUND: Standard pulmonary arterial hypertension (PAH) risk models may not fully apply to PAH associated with unrepaired congenital heart disease (PAH-CHD). We aimed to develop and internally validate an interpretable machine learning (ML)-based risk model for adults with unrepaired PAH-CHD, comparing it against the recent European Society of Cardiology (ESC) model. METHODS: Utilizing a nationwide prospective registry in China, we included adults with unrepaired PAH-CHD. Five survival models were trained and internally validated to predict all-cause mortality. The optimal model was interpreted via Shapley Additive exPlanations (SHAP) to identify the top 15 predictors and clinical cutoffs. Novel two-strata and continuous risk models were constructed; predictive accuracy was evaluated using the C-index and survival curves. RESULTS: Among 601 patients (mean age 34; 56.4% Eisenmenger syndrome [ES]), 108 died during a median 76-month follow-up. The random survival forest model achieved the highest predictive performance (bootstrapping C-index 0.773). SHAP identified key predictors, including hemoglobin, body mass index, systolic blood pressure, and diastolic pulmonary artery pressure. The ESC two-strata model classified only 5.0% of patients as "Poorer prognosis" and failed to stratify the non-ES subgroup. The novel ML-derived risk model effectively stratified survival in both ES and non-ES cohorts (log-rank P < 0.001), outperforming the ESC model across all subgroups. CONCLUSIONS: Harnessing the strength of interpretable ML on detecting non-linear correlations, we developed a novel risk stratification model for unrepaired PAH-CHD, outperforming the ESC model. This tool assesses mortality risk regardless of Eisenmenger physiology, facilitating individualized clinical management.