Predicting diabetic ketoacidosis in pediatric patients using machine learning

Background Machine learning is a powerful tool to define relationships between large data variables through computing algorithms. In medicine, machine can find the association given disease and disease-related complications such as relationship Diabetes development of diabetic ketoacidosis (DKA). The aim this study develop evaluate predicting model for among pediatric cases leading factors that predict ketoacidosis. Methods We evaluated medical records 3737 patients ages 0 18 years who attended clinics were diagnosed with diabetes. After initial preprocessing, we used Orange, an open source software, visualization, analysis. six prediction models: Decision Tree, Random Forest, kNN, Gradient Boosting, CN2 rule inducer AdaBoost. Data imbalance was managed using oversampling technique. Variables analyzed included age, sex, hemoglobin A1C level, visits education clinic, number appointments clinic. Models based on Area under Curve (AUC), accuracy, precision, recall F1-score stratified 5-fold cross validation Results results show Forest highest performance classifier (AUC=0.98; F1 score=0.92; recall=0.93). Furthermore, HbA1c most contributing factor model. Conclusion This shows importance effectiveness modeling diabetes DKA. Flagging those are at higher risk developing DKA provides better point care these patients.