1823-LB: DiaLOG—A Personalized and Interactive Diabetes Education Platform Empowered with Risk Assessment AI Algorithms and ChatGPT

Introduction & Objective: The latest breakthrough of large language models offers a great opportunity to create personalized and interactive education tools for self-management of chronic conditions.

Methods: In this study, we developed an individualized and interactive diabetes education platform for Chinese populations with prediabetes, which consists of three key components: 1) a dashboard with individual records; 2) the AI algorithms for individualized risk prediction of type 2 diabetes mellitus (T2DM) incidence that were constructed and validated with 17-year territory-wide electronic health records (EHR) in Hong Kong; 2) a Chatbot empowered by chatGPT.

Results: We developed the clinically explainable prediction models using machine learning (ML) and deep learning (DL) techniques, for two-, five- and ten-year risk predictions in the sub-cohorts of 187,785, 99,853 and 17,301 patients, respectively. We found the Dense Layer Neural Networks (DNN) models achieved the best performance, with the area under receiver-operating curve (AUC) of 81.1%, 79.9%, 75.03% for two-, five- and ten-year risk prediction, respectively. The most significant predictors were fasting glucose and HbA1c, followed by creatinine, age at baseline, sex, potassium, triglyceride, and HDL-C. The DNN models achieved satisfactory performance in external validation. The developed models were then integrated into a web-based prediabetes self-management platform, named DiaLOG, as part of prompt engineering for interaction with ChatGPT.

Conclusion: Validated risk prediction models based on EHR data could be integrated into an AI-powered diabetes self-management platform, together with a ChatGPT-powered chatbot, to facilitate clinical decision and empower patients with self-risk assessments and individualized self-management advice.