JMIR Publications

ABSTRACT

Background:

Traditional influenza surveillance systems rely on sentinel hospitals, often suffering from a reporting lag of 1–2 weeks. Social media platforms offer real-time data streams, but extracting high-quality epidemiological signals remains challenging due to linguistic noise.

Objective:

This study aims to develop a cost-effective, real-time influenza surveillance framework by leveraging Large Language Models (LLMs) to distill knowledge into a lightweight classifier (BERT), using Weibo data from Guangdong, China, during the 2025 influenza season.

Methods:

This retrospective study analyzed Sina Weibo posts from Guangdong Province, China, during the 2025 influenza season. Using a domain-specific keyword taxonomy, we collected 389,959 posts and applied multi-stage preprocessing to identify first-person influenza-like illness (ILI) self-reports, yielding a final analytic dataset of 99,160 posts. We then implemented a teacher-student knowledge distillation framework in which DeepSeek-Chat annotated 2,500 posts using zero-shot prompting, and annotation quality was checked against 200 human-labeled posts, showing high agreement(κ=0.846). These labels were used to fine-tune Chinese-RoBERTa-WWM-ext with training, validation, and test splits of 70%, 15%, and 15%, respectively. The fine-tuned model was applied to the full filtered corpus to construct a normalized weekly Weibo Flu Index, which was compared with official weekly ILI% surveillance data from the Guangdong CDC using time-lag cross-correlation and Granger causality analyses, with a keyword-based index used as the baseline.

Results:

The fine-tuned BERT model, distilled from the DeepSeek teacher, achieved an accuracy of 88.2% and an F1-score of 0.881 on the test set, successfully retaining the linguistic reasoning capabilities of the large model. Time-lag cross-correlation analysis indicated that the BERT-derived Weibo Flu Index was highly synchronous with official ILI% data, peaking at a one-week lead time (Lag 1: r=0.676, P<.001). In contrast, the keyword baseline yielded a significantly lower correlation (r=0.5679), with the BERT approach demonstrating a 17.7% performance improvement at the optimal lag. For the association between Weibo semantic signals and official surveillance data, Granger causality tests were performed, with results showing F=4.32 and P=.044.

Conclusions:

The framework offers a critical 7-day early warning window for rapid response. By leveraging a distilled BERT model, we demonstrate that high-performance surveillance is accessible to local health departments with limited budgets. This work underscores the potential of multi-source digital intelligence as a cornerstone of effective global disease defense.