Development of an intellectual system to support medical decision-making on forecasting chronic kidney disease in children.

Introduction. Chronic kidney disease is often diagnosed too late. Currently, the diagnostic accuracy is 44.1%, which highlights the urgent need to improve diagnostic methods. The purpose of the study: to develop a model – a system of support of medical decisions to predict chronic kidney disease in children. Materials and methods. А one-center retrospective cohort study (2011–2022) of children with chronic kidney disease of 1–4 stages in the age from 1 to 17 years. To construct a predictive model for diagnosing chronic kidney disease in children, an ensemble learning method was used, using which the models obtained by machine learning algorithms were combined: multi-factor logistic regression and decision tree. The models use five variables: asthenic physique in a child, loss of protein and red blood cells with urine, ESR and blood serum sodium. Results. The study involved 158 patients. The main group includes 128 children with chronic kidney disease of stage 1–4 aged 1–17. The comparison group was 30 children with no diagnosed kidney pathology aged 1 to 17. The children of the two groups did not differ statistically by sex and age. A model has been obtained to predict chronic kidney disease in children on a test sample with an accuracy of 93.5% [87.1; 100.0]%; a sensitivity of 92.0% [82.1; 100.0]; a specificity of 100.0% [100.0; 100.0.0]; ROC-AUC = 98.7%;100.0].0. Obtained model of excellent quality (>90%). The model describes 90.3% [83.8; 96.1]% variance. Conclusion. The proposed model demonstrates excellent predictive ability and may be of important clinical importance for predicting the chronic process in primary health care, where symptoms associated with the risk of chronic kidney disease, may be overlooked. Predicting and developing early nephroprotective strategies can lead to better treatment outcomes and prolong life.