Machine Learning Approach to Standard Clinical and Biochemical Parameters for Prediction of Pathological Transcranial Doppler in Pediatric Sickle Cell Disease

ABSTRACT Background Children with sickle cell disease (SCD) are at increased risk of cerebrovascular complications, with elevated Transcranial Doppler (TCD) velocities predicting stroke. Reliable laboratory correlates of pathological TCD remain uncertain. This study applied machine learning (ML) to identify hematologic and biochemical predictors of abnormal TCD velocities in pediatric SCD. Methods We conducted a retrospective analysis of 134 children with SCD (HbSS, HbSβ 0 , HbSβ + , HbSC) followed between 2015 and 2024. A Gradient Boosting Binary Classification model (H2O framework) was trained using demographic, hematologic, biochemical, and treatment variables. Model performance was evaluated using five‐fold stratified cross‐validation, assessed by the area under the receiver operating characteristic curve (AUC) and the precision–recall curve (AUCPR). Results Across 3190 patient‐years, 447 TCD examinations were analyzed, of which 45 (10.1%) showed conditional or abnormal velocities (≥170 cm/s). The optimized model achieved an AUC of 0.776 and an AUCPR of 0.360, indicating moderate discriminative ability. Reticulocyte and neutrophil counts, hemoglobin, fetal hemoglobin, platelet count, and age were strong predictors, while ongoing therapy showed a protective effect. Conclusions ML models offer an interpretable, exploratory framework to investigate associations with TCD findings in pediatric SCD, supporting hypothesis generation within guideline‐based care. The absence of systematic Internal Carotid Artery velocity assessment represents an important limitation of this analysis.