From Real to Synthetic: An Analytical Machine Learning and Generative AI Approach for Reducing Early Readmissions in U.S. Diabetic Patients

Abstract

US healthcare systems struggle with hospital readmissions, especially within 30 days of release. ML-based prediction of hospital readmission is a significant milestone for healthcare professionals to efficiently manage healthcare quality, premature readmissions, post-discharge follow-ups, incur large costs and identify high-risk patients. Many disadvantages still prohibit these models from being widely adopted and successful. This research proposes a promising solution to the challenge of early hospital readmissions among diabetic patients in the U.S. by combining ML and generative artificial intelligence (AI) techniques. Utilizing the UCI Diabetes 130-US Hospitals Dataset, the study develops predictive models for 30-day readmission risks using conventional Random Forest (RF), K-Nearest Neighbors (KNN), and Gradient Boosting Classifier (GBC)—alongside Conditional Tabular Generative Adversarial Networks (CTGAN) and Variational Autoencoders (VAE) for synthetic data augmentation. These models effectively address the issue of class imbalance, significantly improving model accuracy, fairness, and interpretability. Among all classifiers, RF achieved the highest accuracy (83%), indicating its robustness in handling complex clinical data. The results underscore the value of integrating structured EHRs with advanced AI to enable early interventions, equitable resource distribution, and improved patient outcomes. Future research will incorporate unstructured clinical data and validate models across diverse hospital environments.