A Machine Learning Model for Risk Stratification of Postdiagnosis Diabetic Ketoacidosis Hospitalization in Pediatric...

Area of Science:

• Pediatric Endocrinology
• Artificial Intelligence in Medicine
• Clinical Informatics

Background:

• Diabetic ketoacidosis (DKA) is a major cause of morbidity and mortality in pediatric type 1 diabetes (T1D).
• Existing prediction models for DKA hospitalization lack explainability and clinical readiness.
• There is a need for accurate, interpretable models to identify at-risk pediatric T1D patients.

Purpose of the Study:

• To develop an interpretable machine learning model for predicting DKA hospitalization risk in children with T1D.
• To utilize routinely collected electronic health record (EHR) time-series data for risk prediction.
• To provide clinic-ready insights for early intervention.

Main Methods:

• Retrospective case-control study using EHR data from 3794 pediatric T1D patients.
• Trained an explainable gradient-boosted ensemble (XGBoost) model with 44 EHR features.
• Evaluated model performance using AUC-weighted F1-score, precision, and recall; analyzed Shapley values for interpretability.

Main Results:

• The model achieved an AUC of 0.80, distinguishing DKA from non-DKA cohorts (P<.001).
• Key predictors identified include diabetes age and glycated hemoglobin levels at 12 months.
• The model stratified patients into risk groups (5%, 20%, 48%) for postdiagnosis DKA.

Conclusions:

• An explainable AI model was developed to predict and risk-stratify pediatric T1D patients for DKA hospitalization.
• The model identifies critical time points and risk factors for targeted clinical interventions.
• This tool has the potential to be integrated into clinical workflows to mitigate DKA risk and improve patient outcomes.