Real-Time Influenza Surveillance Using DeepSeek-Augmented BERT Classification on Chinese Social Media: Infodemiology Study
Journal of Medical Internet Research ·
Background: Traditional influenza surveillance systems rely on sentinel hospitals and often experience a 1‐ to 2-week reporting lag, limiting timely public health response. Social media platforms provide real-time data streams, but extracting reliable epidemiological signals remains challenging due to substantial linguistic noise. Although large language models (LLMs) improve semantic extraction, their computational cost limits routine deployment in resource-constrained public health settings. Objective: This study proposes and evaluates a hybrid framework that distills knowledge from an LLM (DeepSeek) into a lightweight bidirectional encoder representations from transformers (BERT)–based classifier for real-time influenza surveillance using Weibo data from Guangdong Province, China, during the 2025 influenza season. The study further assesses the temporal alignment between the derived signal and official influenza-like illness (ILI) surveillance data. Methods: A total of 389,959 Weibo posts from epidemiological weeks 7 to 50 (February 10 to December 14, 2025) were collected using a domain-specific keyword strategy. After multistage filtering, 99,158 posts were retained for analysis. A DeepSeek-Chat model was used to generate pseudo-labels for 2500 posts via zero-shot prompting, which were evaluated against expert annotations (Cohen κ=0.846) and used to fine-tune a Chinese-RoBERTa-WWM-ext classifier (a robustly optimized BERT pretraining approach). A chronological train-validation-test split was applied to reduce temporal leakage. The trained model was used to construct a weekly Weibo Flu Index. This index was compared with keyword-based and Term Frequency–Inverse Document Frequency (TF-IDF) + logistic regression baselines using time-lag Pearson correlation (0‐5 wk) against official ILI% data. Savitzky-Golay smoothing and second-derivative analysis were applied to identify epidemic inflection points and peak timing. Results: The classifier achieved an accuracy of 85.1% and an -score of 0.800 on a chronologically held-out test set, identifying 29,779 posts (30.0%) as ILI self-reports among 99,158 posts. The BERT-derived index showed the strongest correlation with official ILI% at lag 0 (=0.523, 95% CI 0.268‐0.710; ²=0.274) and lag 1 (=0.446, 95% CI 0.169‐0.658), compared with the keyword baseline (lag 0: =0.481, ²=0.231) and TF-IDF baseline (lag 0: =0.185, =0.034). Inflection analysis indicated that the Weibo-derived signal preceded both the epidemic rise (week 44 vs week 45) and peak (week 48 vs week 49) in official ILI data, corresponding to an approximate 1-week lead at both transition points. Conclusions: The results suggest that a DeepSeek-augmented lightweight BERT classifier can extract meaningful influenza-related signals from noisy Chinese social media data and produce a surveillance index that shows an approximate 1-week lead relative to official ILI reporting, with improved performance over keyword-based and TF-IDF approaches. This framework provides a proof-of-concept for cost-effective real-time influenza nowcasting using social media data.