•  
  •  
 

Document Type

Conference Proceedings

Abstract

Objective: To systematically evaluate machine learning (ML) and artificial intelligence (AI) models developed to predict diabetic ketoacidosis (DKA) or related outcomes in pediatric populations. Methods: PubMed and Google Scholar were searched using combinations of terms for DKA, ML/AI, and pediatric populations. Eligible studies developed or validated ML/AI models predicting DKA or DKA outcomes in individuals ≤18 years. Non-ML or adult-only studies were excluded. No language or date limits were applied. Fifty-five records wemimra003@fiu.edure identified; seven met inclusion criteria. Data were extracted on study design, model type, predictors, and performance. Results: The seven studies primarily used retrospective electronic health record data, with sample sizes ranging from under 100 to over 600,000 participants. Algorithms included logistic regression, random forests, gradient boosting, ensemble learners, and neural networks. Common predictors were glucose, bicarbonate, age, and prior DKA history. Reported discrimination varied (AUC 0.70–0.98). Only a minority performed external validation; most relied on internal cross-validation. Risk of bias was moderate to high, driven by small sample sizes and single-center data. Conclusion: ML/AI approaches show promise for predicting DKA or identifying pediatric patients at heightened DKA risk, with several models demonstrating strong internal discrimination (AUROC often reported in the high 0.8–0.9 range with top-reported values up to ~0.98). However, pediatric literature is limited in quantity and quality as many models lack prospective validation and have variable sample sizes. Future research should emphasize multicenter data, standardized outcomes, transparent reporting, and external validation to support clinical application.

Download

Share

COinS