Developing a prediction model for early detection and classification of kidney dysfunction in children with sickle cell disease

Introduction Children with sickle cell disease are at an elevated risk of developing chronic kidney disease, yet early detection remains a challenge in low-resource settings like Uganda. Traditional diagnostic tools often fail to capture subclinical kidney dysfunction, particularly in pediatric populations where glomerular hyperfiltration and atypical progression are common. Machine learning offers a promising complementary avenue for early risk identification, but model interpretability and contextual adaptation remain barriers to clinical integration. We developed a prediction model for the early detection and classification of kidney dysfunction in children aged 18 years and below with sickle cell disease at Mulago National Referral Hospital Uganda. Methods This retrospective study used data from pediatric SCD patients at Mulago National Referral Hospital and the Uganda Sickle Pan Africa Research Consortium (SPARCo) registry. Kidney dysfunction was staged according to KDIGO 2012 guidelines. A Random Forest classifier was trained on demographic, clinical, and laboratory features, with stratified 5-fold cross-validation. Results A total of 893 children were included. Mean age was 10 years (SD 4.5); 52% were female. The Random Forest model achieved 97% accuracy, F1-score 0.94, and AUROC 0.875. The classifier reliably stratified patients across KDIGO-defined stages, with minimal misclassification. Conclusion A Random Forest–based model demonstrated high performance in detecting and classifying kidney dysfunction in pediatric SCD patients in Uganda. Integrating explainable AI with clinical staging enhances interpretability, offering a promising tool for early risk stratification and intervention in resource-limited settings.