Prediction of hypertension and restenosis under guideline-directed management in aortic coarctation: development and validation of machine-learning models
Background
Aortic coarctation (CoA) is a major cause of arterial hypertension in young individuals, with recurrence occurring in up to one-third of patients throughout life despite guideline-directed management.
Methods
We conducted a development and validation study at a single centre in Berlin, Germany, utilising routinely collected electronic health records, cardiovascular magnetic resonance (CMR), and mid-/long-term follow-up data from 218 visits (160 individuals with CoA receiving standard of care, guideline-based management) collected between January 2014 and April 2022. Machine learning (ML) models (CatBoost, XGBoost, random forest, support vector classifiers, neural networks, logistic regression, and K-nearest neighbours) were developed to predict three endpoints: re-coarctation requiring intervention (CoA-I), aortic surgery (CoA-S) as a subset of CoA-I, and persistent arterial hypertension. The dataset was divided by random stratified split into a training set (n = 159; for model development with five-fold cross-validation), and a hold-out test set (n = 59; for out-of-sample validation). Stratification was based on sex, age, and CoA-I status. We included a final set of 38 clinically relevant features, encompassing baseline characteristics, medication intake, echocardiography, CMR, electrocardiography (ECG), and treatment decisions. ClinicalTrials.gov Identifier: NCT02591940.
Findings
Tree-based and support vector classifier models performed best after Bayesian hyperparameter optimisation, yielding high performance in a stratified validation cohort: area under the receiver operating characteristic curve (ROC AUC) 0.90 ± 0.01 for CoA-I, 0.90 ± 0.01 for CoA-S, and 0.84 ± 0.01 for hypertension. Shapley Additive exPlanations (SHAP) highlighted peak Doppler gradient, time since index visit, and ventricular size indices as key predictors for CoA-I. In inverse-probability-weighted analyses, antihypertensive medication was associated with a lower CoA-I probability (−17.3%; 95% confidence interval [CI], −28.2 to −6.4; p = 0.002), with concordant propensity-score-matched findings. An open-access research interface (https://icm.dhzc.charite.de/calc_coa) incorporates treatment thresholds and personalised risk estimates from an updatable ML framework.
