Integrating quantitative aortic-isthmus imaging with clinical echocardiographic measurements for prenatal prediction of coarctation of the aorta: A development and validation study
Shuhao Song, Qunli Li, Pan Yang, Yushan Liu, Bo Liu, Zhe Xu, Abdoul Rachid Aboubacar Hamidou, Shi ZengObjectives
Prenatal assessment of suspected coarctation of the aorta (CoA) remains imprecise because routine echocardiographic measurements capture only part of the information available in the fetal aortic-isthmus image. We assessed whether a quantitative representation of this image adds predictive value to clinical echocardiographic measurements.
Methods
This retrospective study included 143 fetuses with suspected CoA; 67 had postnatally confirmed CoA and 76 did not. Thirteen clinical variables formed the baseline. We defined an aortic-isthmus imaging representation (AIIR) that combined handcrafted radiomic descriptors with features from five pretrained image encoders. Clinical-only and AIIR-augmented models used identical patient-level folds. Imputation, scaling, feature selection, model configuration and threshold selection were confined to 5-fold outer and 3-fold inner nested cross-validation, repeated five times. For temporal validation, the cohort was partitioned chronologically into a development cohort (n=80) and a temporal validation cohort (n=63), to which the development-fitted pipelines were applied unchanged.
Results
In nested cross-validation, AUC was 0.8132 (95% confidence interval [CI], 0.7363–0.8801) for the clinical model and 0.9098 (95% CI, 0.8549–0.9565) for the AIIR-augmented model. The paired difference was 0.0966 (95% CI, 0.0324–0.1649), with a two-sided patient-level prediction-swap P value of 0.003. The AUC difference was positive across all five repeats. In temporal validation, the AIIR-augmented model increased the AUC from 0.8247 to 0.9568 (ΔAUC, 0.1321; 95% CI, 0.0444–0.2346; P=0.0028).
Conclusions
AIIR added significant and repeatable discriminatory information beyond clinical echocardiographic measurements under patient-level nested cross-validation, and the increment persisted in temporal validation. These findings support quantitative aortic-isthmus imaging as a candidate adjunct for prenatal CoA assessment and provide a clear basis for locked external validation and recalibration.