The self-evlaution tool, powered by a large language model, processes user queries through a comprehensive generation, decision, action, and safety framework to deliver optimal responses. The system's key features include retrieval-augmented in-context learning, which enhances the responses generated by sourcing information from reliable websites. It also incorporates a Guardrail module to mitigate potential harmful content in the responses by validating the content before delivery. Additionally, the system features a Self-checking memory module that maintains essential clinical characteristics across multi-turn dialogues, ensuring consistent and continuous interactions with users.

Following the baseline assessment, participants will be guided to use a self-evaluation tool independently to assess their risk of diabetic retinopathy (DR). This tool is a fusion of a conversational AI system based on LLM and an existing logistic diagnostic model.

The AI system is designed to collect clinical variables, including age, duration of diabetes, Body Mass Index (BMI), and insulin usage. Additionally, clinical test data such as mean arterial pressure, HbA1c, serum creatinine, and microalbuminuria will be extracted from a local dataset using the patient's name and ID. Once collected, these data will be transmitted to a server-based diagnostic model for further analysis to determine the presence of DR.

The performance of the self-evaluation tool is evaluated with accuracy with reference to the diagnostic labels by senior ophthalmologists based on fundus photos.