Conversational AI framework refines COVID-19 infection risk from social contact data, preprint says

[medRxiv] A preprint study developed an iterative retrieval-augmented generation conversational framework to predict COVID-19 infection risk from textual descriptions of social interactions. According to the study, the team trained a long short-term memory (LSTM) model on 19,434 textual descriptions of social interactions collected during the pandemic in the Leuven student community in Belgium, and integrated it into a Llama3.3-based conversational framework that iteratively gathers additional social information that may represent infection risk. Based solely on social interaction data, a fine-tuned RobBERT model predicted infection risk with a ROC AUC of 0.73, similar to the LSTM model and a logistic regression on TF-IDF features, both at 0.72, and the three slightly outperformed a QLoRA-fine-tuned Mistral NeMo model at 0.69, according to the study. The study said Llama3.3 performed well in the framework, with predicted infection risk often converging and fewer redundant questions than the smaller Mistral 7B and Gemma3 12B models; the framework also stratified synthetic users with different types of social interactions by infection risk. The authors concluded that combining simple language models with large language models can refine individual infection risk from contact tracing data in a high-circulation context. The findings are from a preprint and have not been peer-reviewed.