AI-Driven Glycemic Instability Risk Modeling for Proactive Intervention and Chronic Disease Management in U.S. Healthcare Systems

Abstract

Hospitals with diabetic patients experience serious clinical problems because patients with diabetes develop glycemic instability which causes their blood glucose levels to fluctuate unpredictably. The process of identifying high-risk patients should start immediately because it requires accurate detection methods which help to protect patient safety. Healthcare experts face multiple difficulties when attempting to predict glycemic instability because clinical data exhibits extreme class imbalance and laboratory results plus medications and patient characteristics show complex interactions. The UCI 130-hospital diabetes dataset serves as the foundation for our complete machine learning and deep learning system which we developed to predict glycemic instability risk. The combination of Synthetic Minority Over-sampling Technique (SMOTE) with cost-sensitive learning provides us with a solution to tackle the challenges that arise from extreme class imbalance. The seven predictive models which include Logistic Regression, Random Forest, Gradient Boosting, XGBoost, Support Vector Machine, Multi-Layer Perceptron (MLP) and TabNet use a clinically informed decision threshold which helps them to detect medical conditions with high accuracy. The evaluation of model performance examines five metrics which include accuracy, precision, recall, F1-score and ROC-AUC. The experimental results show that machine learning through deep learning and ensemble methods achieves better results for detecting glycemic instability than traditional classifiers. The deep learning models TabNet and MLP exhibit high sensitivity for detecting unstable patients while Gradient Boosting and Random Forest demonstrate superior discriminative ability with ROC-AUC values close to 0.99. The analysis of feature importance shows that HbA1c levels, maximum glucose concentration, number of diagnoses, and insulin usage are the most influential predictors of instability. The research results demonstrate that models which use imbalance-aware machine learning together with explainable models can achieve accurate predictions.