Applying Machine Learning to Fetal Echocardiograms: A Novel Method for Predicting Critical Coarctation of the Aorta

Current fetal echocardiographic metrics are inadequate for confident prediction of neonatal critical coarctation of the aorta due to low specificity. Random forest algorithms, a subset of machine learning, have not been applied to fetal echocardiography. We aimed to determine whether a random forest classifier improved the accuracy of critical coarctation prediction compared to previously published risk metrics.

Patients with prenatal concern for coarctation at a single center were included. Eight fetal echocardiogram measurements were used to train a random forest classifier with 80:20 splits and 5-fold cross validation to predict coarctation intervention within 30 days of life. Gestational age at the time of fetal echocardiogram was also included. A Shapley additive explanations (SHAP) analysis assessed the marginal contribution of each feature. A consolidated model was then validated with an external patient cohort from a different academic center.

Inclusion criteria were met by 132 patients in the initial cohort and 64 patients in the external validation cohort, of whom 44% (n = 58) and 25% (n = 16), respectively, had coarctation requiring intervention. A SHAP analysis for both cohorts demonstrated aortic arch angles as the most influential features. Using internal cross validation on the initial cohort, the area under the receiver operating characteristic curve was 0.95 ± 0.01 (sensitivity 1.0, specificity 0.96) with an F1 score of 0.97 ± 0.03. Validation of a consolidated model with the external cohort produced a sensitivity of 0.81, specificity of 0.98, and F1 of 0.87.

A random forest classifier using fetal echocardiogram features predicted neonatal critical coarctation with higher accuracy than previously published metrics. The model maintained excellent specificity when validated with an external patient cohort. Aortic arch angles most significantly impacted the model's accuracy. Future directions include prospective validation and converting the model to a distributable clinical calculator.