Machine learning-based prediction of postoperative restenosis risk following direct repair of paediatric coarctation of the aorta
Ruge Duan, Zixu Wang, Kexin Yang, Nan Ding, Hanlu Yi, Yao-Bin Zhu, Song Bai, Xiaofeng Li, Feng Tong, Zhi-Qiang LiAbstract
Objectives:
Coarctation of the aorta is a congenital cardiovascular disease with focal aortic luminal narrowing, and paediatric patients face a high postoperative restenosis risk. This study aimed to develop and validate an interpretable machine learning model for early predicting restenosis after paediatric coarctation of the aorta direct repair using preoperative and intraoperative data.
Methods:
A total of 117 patients (2016–2024) were retrospectively enrolled, divided into restenosis (21 cases, 17.9%) and non-restenosis (96 cases, 82.1%) groups (restenosis was defined as a peak systolic pressure gradient >20 mmHg measured by echocardiography). Recursive feature elimination with cross-validation screened key variables; six machine learning models were built with 5-fold randomised search cross-validation tuning, using the area under the curve as the primary metric. SHapley Additive exPlanation analysed feature contributions.
Results:
The multilayer perceptron model performed best (mean area under the curve = 0.8333, 95% CI: 0.7111–0.9555, accuracy = 0.8376) with balanced precision-recall. SHapley Additive exPlanation identified low body surface area as the top risk factor. Resection and extended end-to-end anastomosis/end-to-side anastomosis were preferred surgically, while resection with end-to-end anastomosis should be avoided; end-to-side anastomosis reduced restenosis risk in patients with aortic arch hypoplasia.
Conclusion:
Machine learning models enable personalised, high-accuracy restenosis prediction. SHapley Additive exPlanation-facilitated risk factor identification optimises treatment strategies. Future prospective studies are needed to validate the models and develop clinical tools.