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

Area of Science:

• Epidemiology
• Computational Biology
• Public Health

Background:

• Accurate infectious disease forecasting is vital for public health.
• Traditional statistical and machine learning models face challenges with complex data.
• Large language models (LLMs) offer a potential advancement in forecasting capabilities.

Purpose of the Study:

• To evaluate the performance of LLMs in forecasting infectious disease cases and deaths.
• To compare LLM forecasting accuracy against established statistical and machine learning models.
• To identify specific disease categories and geographical regions where LLMs excel or underperform.

Main Methods:

• Collected monthly infectious disease data from China and the US NNDSS (2009-2025).
• Evaluated seven models: 4 statistical (ARIMA, TGARCH, EGARCH, ETS), 2 machine learning (XGBoost, LSTM), and 1 LLM (Qwen-2.5-3B fine-tuned with LoRA).
• Assessed performance using MAE, RMSE, and MAPE, with statistical comparisons via Friedman and Nemenyi tests.

Main Results:

• LLM performance was broadly comparable to baseline models.
• The LLM demonstrated significant advantages in case-forecasting, particularly for zoonotic and intestinal infections.
• LLM performance varied geographically, showing significant gains in China but not the US.
• No significant differences were detected between models for death forecasting.

Conclusions:

• LLMs fine-tuned with LoRA offer a competitive alternative for infectious disease forecasting.
• LLMs show particular promise for predicting infectious disease cases, outperforming other models in specific categories.
• Further research may explore LLM applications in more complex epidemiological forecasting scenarios.