Autonomous LLM-Guided Disease Prediction System: Real-Time Matching and Surpassing CDC Expert-Ensembled Models

This article introduces an autonomous disease prediction system using large language model (LLM)-guided tree search. The system underwent prospective real-time evaluation during the 2025-2026 U.S. respiratory infectious disease season, autonomously discovering prediction models for influenza, COVID-19, and RSV. Its ensemble results match or surpass the gold standard manually curated by the CDC, breaking through the labor bottleneck of traditional infectious disease prediction.

Probabilistic prediction of infectious diseases is crucial for public health decisions (e.g., resource allocation, vaccine planning). However, traditional methods rely on labor-intensive expert model curation: manual model development requires collaboration across multiple teams, has long iterative optimization cycles, and poor scalability (difficult to quickly respond to new pathogens or fine-grained regional predictions).

The core architecture of the system consists of three parts: 1. LLM-guided code generation (exploring model architectures, generating executable code, translating epidemiological theories); 2. Tree search optimization (Monte Carlo Tree Search explores the code space and expands based on performance feedback); 3. Automatic judge mechanism (checking theoretical consistency, structural fidelity, and interpretability). The system can autonomously discover diverse models such as mechanistic models, statistical models, machine learning methods, and hybrid approaches.

Evaluation setup: During the 2025-2026 U.S. respiratory season, predicting the number of influenza, COVID-19, and RSV cases for the next 1-4 weeks, with real-time predictions generated weekly. Core results: The system's ensemble model consistently matches or surpasses the gold standard of CDC expert ensembles, and in cold-start scenarios with scarce RSV data, it maintains competitiveness through transfer learning and adjusting model complexity.

Ablation experiments validate key designs: 1. Log-scale distance metrics can prevent reward hacking (avoiding model fitting to extreme values); 2. The automatic judge mechanism ensures generated models comply with epidemiological principles, avoiding black-box models that perform well statistically but are scientifically unreasonable.

Research implications: 1. Integration of automation and specialization (LLM explores model space + expert knowledge encoded into the judge mechanism); 2. Breaking through scalability bottlenecks (rapid deployment to new regions/pathogens, supporting fine-grained predictions); 3. Transparency and interpretability (generated code is readable, has clear theoretical support, and provides uncertainty quantification).

Limitations: High computational cost, dependence on data quality, unproven generalization ability for extreme events, and ethical considerations. Future directions: Developing efficient search algorithms, integrating multi-modal data, establishing real-time feedback mechanisms, and optimizing human-machine collaboration interfaces. Conclusion: This system marks the transition of infectious disease prediction from manual development to automated intelligent systems, and AI will play an important role in the field of public health.