Large language models as versatile predictive engines for notifiable infectious diseases.

Accurate forecasting of infectious disease cases and deaths is crucial for public health decision-making. Traditional statistical and machine learning approaches may be challenged by complex temporal patterns and heterogeneous surveillance data. We evaluated large language models (LLMs) for infectious disease forecasting. We collected monthly reported cases and deaths data from China National Notifiable Diseases Surveillance System (NNDSS) from January 2009 to February 2025, and monthly reported cases from the United States NNDSS from 2016 to 2023, covering 79 notifiable infectious diseases and 111 forecasting tasks. We evaluated seven models: four statistical models (ARIMA, TGARCH, EGARCH, ETS), two machine learning models (XGBoost, LSTM), and an LLM-based regression model built on Qwen-2.5-3B and fine-tuned using Low-Rank Adaptation (LoRA). Performance was assessed using mean absolute error (MAE), root mean squared error (RMSE), and mean absolute percentage error (MAPE). We used Friedman tests and Nemenyi post-hoc comparisons versus the LLM, reporting mean-rank differences (ΔR; > 0 indicates better LLM ranking). In pooled analyses, the Friedman test indicated significant heterogeneity across models (χ² = 26.24, P < 0.001). Nemenyi comparisons showed positive ΔR for all baselines against the LLM, with significance only in the XGBoost-LLM comparison (ΔR = 0.72, P < 0.001). By category, significant positive ΔR for all baselines against the LLM was observed in zoonotic infections (XGBoost: ΔR = 1.42, P < 0.001) and intestinal infections (TGARCH: ΔR = 1.28, P = 0.012; EGARCH: ΔR = 1.98, P < 0.001), whereas blood-borne infections showed LLM inferiority (ETS: ΔR=-2.29, P = 0.011; XGBoost ΔR=-2.86, P < 0.001). Geographically, positive ΔR was significant in China (XGBoost: ΔR = 0.80, P = 0.007), while the U.S. stratum was not significant (χ² = 9.04, P = 0.171). By outcome, between-model differences were detectable for case-forecasting (χ² = 20.72, P = 0.002; XGBoost: ΔR = 0.73, P = 0.005), with no detectable differences for death (χ² = 9.30, P = 0.158). In conclusion, LoRA fine-tuned LLM showed performance broadly comparable to statistical and machine learning models, with significant positive ΔR mainly observed in case-forecasting and in zoonotic and intestinal infectious diseases.