P-164: Youth diabetes risk prediction using machine learning and personalized prevention plan using generative AI
S Wang, G GoldszteinAbstract
Introduction
Adolescent diabetes, particularly type 2 diabetes, has increased dramatically in the United States over recent decades. Due to the limited availability of clinically validated comprehensive youth diabetes datasets, relatively few studies have used machine learning to investigate the risk factors specifically associated with youth diabetes.
Objectives
This study aimed to develop high-accuracy machine learning models to predict youth diabetes risk and evaluate the efficacy of diverse Large Language Models (LLMs) in generating personalized prevention and management plans.
Methods
We utilized a comprehensive cohort of 15,149 individuals from the most recent comprehensive youth diabetes dataset, NHANES (1999–2018), to train and validate machine learning models representing diverse algorithmic approaches. Model performance was rigorously compared through standard classification metrics to identify optimal models and key risk factors. In the second phase, generative AI was employed to generate customized risk management plans, including detailed dietary diaries and physical activity advice. Large-language models ChatGPT, Grok, Gemini, and DeepSeek were systematically cross-evaluated using an “LLM-as-a-Judge” framework to assess clinical accuracy and relevance.
Results
Thirteen machine learning models were evaluated using 94 features to identify the top 20 risk factors. Following model comparison, three LLMs were used to generate and assess personalized prevention plans.
Conclusions
Logistic Regression emerged as the top-performing approach with the highest ROC-AUC score 0.683, demonstrating strong accuracy on prediction of youth diabetes. DeepSeek was the top-rated LLM for generating personalized prevention plans, providing detailed advice on diet and physical activity.