MIFlu: Large Language Model-Based Multimodal Influenza Forecasting Scheme.

Area of Science:

• Epidemiology
• Computational Biology
• Artificial Intelligence

Background:

• Accurate influenza forecasting is vital for public health interventions.
• Current deep-learning models for influenza prediction are limited by their inability to incorporate contextual information.
• Large language models (LLMs) show promise in enhancing time-series predictions by integrating text data.

Purpose of the Study:

• To propose MIFlu, a novel multimodal influenza forecasting scheme.
• To leverage LLMs for fusing contextual text information with influenza time-series data.
• To improve the accuracy and robustness of influenza early forecasting.

Main Methods:

• MIFlu utilizes two LLMs: one for extracting text embeddings from user prompts and another for forecasting.
• It fuses text embeddings with time-series embeddings of influenza occurrences.
• The fused embeddings are then used by the forecasting LLM to predict future influenza trends.

Main Results:

• MIFlu demonstrated superior performance compared to existing predictive models in extensive experiments.
• Prediction accuracy improved by up to 26.2% over state-of-the-art models.
• Analysis confirmed the impact of text embedders, hyperparameters, and data volume on forecasting accuracy.

Conclusions:

• The proposed MIFlu scheme effectively integrates contextual text information into influenza forecasting.
• Multimodal approaches using LLMs offer significant advantages over unimodal methods for epidemiological predictions.
• MIFlu represents a promising advancement in public health surveillance and preparedness for influenza outbreaks.