Using iterative agentic retrieval-augmented generation to refine the risk of infection from social contexts during the COVID-19 pandemic: Development and Evaluation Study

In the Leuven student community in Belgium, an iterative Llama3.3 retrieval-augmented conversational framework predicted COVID-19 infection risk from social interaction text with ROC AUC up to 0.73.

• Why it matters: Contact tracing is essential but time-consuming, and existing infection-risk models do not adapt their estimates through a user-driven conversation to refine personalized social risk factors.
• What they did: An LSTM was trained on 19,434 textual descriptions of social interactions, then embedded in a Llama3.3-based conversational system that iteratively gathered additional social details and was evaluated using LLM-as-a-judge metrics and context comparisons.
• The result: A fine-tuned RobBERT reached ROC AUC 0.73 (LSTM and TF-IDF logistic regression: 0.72; QLoRA Mistral NeMo: 0.69), and the framework stratified synthetic users by infection risk while converging often and asking fewer redundant questions.

AI summary of the abstract · may contain errors · not medical advice · read the paper.