Back

Fine-tuned large language models enhance influenza forecasting

Li, C.; Zhang, Q.; Zhang, Y.; Zhao, B.; Yang, J.; Qi, L.; Ding, J.; Tian, D.

2025-03-31 infectious diseases
10.1101/2025.03.27.25324747 medRxiv
Show abstract

Influenza-like illness (ILI) continues to present significant challenges to global health, highlighting the need for accurate forecasting to guide timely public health responses. Traditional statistical and deep learning models, though widely applied, often face difficulties in capturing complex nonlinear dynamics and addressing data scarcity. This study examines the potential of fine-tuned large language models (LLMs), including Llama2 and GPT2, for multi-step influenza forecasting. A specialized fine-tuning framework is introduced, incorporating custom embeddings and a prediction block, and evaluated on seven real-world ILI surveillance datasets. Extensive benchmarking against SARIMA, LSTM, and PatchTST demonstrates that fine-tuned LLMs consistently deliver superior accuracy and stability, with particular advantages in long-term forecasts. Pre-trained LLMs, while able to capture broad temporal patterns in zero-shot scenarios comparable with SARIMA, gain substantial improvements in precision through fine-tuning. These results establish fine-tuned LLMs as practical and robust solutions for influenza forecasting, offering new opportunities to enhance epidemic modeling in data-limited public health environments.

Published in Cell Reports Methods · training set

Matching journals

The top 9 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.