Artificial intelligence in Optimizing Surgical Treatment of Tetralogy of Fallot: Development and Validation of Prognostic Model

INTRODUCTION: Tetralogy of Fallot (TOF) is the most common cyanotic congenital heart defect. With traditional examination methods precise prediction of complication risks is not always possible. AIM: To develop a machine learning model that recommends the optimal surgical tactics based on the analysis of clinical and instrumental data. MATERIALS AND METHODS: As part of a single-center cohort study, a retrospective registry was formed of clinical and instrumental data from 300 children with TOF who underwent primary surgical treatment in 2017–2022. The target variable was a four-class category corresponding to the type of intervention. The used surgical strategies were distributed as follows: radical TOF correction — 168 patients (57.1%), creation of a systemic-pulmonary anastomosis — 83 (28.2%), reconstruction of the outflow ways/tract from the right ventricle — 25 (8.5%), stenting or ballooning of the right ventricular outflow tract, including transluminal balloon plasty of the pulmonary valve — 24 (8.2%). RESULTS: To address this problem, a classification model with internal validation was constructed. The gradient boosting model demonstrated satisfactory predictive validity. Contributed most to the decision-making process: oxygenation parameters (blood oxygen saturation), the Nakata and McGoon vascular-anatomical indices, the pressure gradient between the right ventricle and pulmonary artery, and the left ventricular volume. The obtained results indicate that the proposed model can serve as a tool for supporting clinical choice of surgical strategy in TOF by aligning computational predictions with pathophysiologically justified predictors. CONCLUSION: In the context of this study on predicting surgical tactics in TOF, the use of 5-Fold cross-validation made it possible to verify that the LightGBM model demonstrates high accuracy not only on a specific subset of data, but also on all five partitioning options. This confirms its reliability and applicability in real clinical conditions.