Large language models (LLMs) have strong capabilities in medical text understanding but struggle with EHR-based prediction tasks due to the complexity of temporally structured and high-dimensional data. Traditional methods often treat LLMs as static retrievers and rely on downstream deep learning models for prediction, which limits the intrinsic reasoning capacity of LLMs and inherits the generalization weaknesses of these models. To overcome these limitations, the authors propose EAG-RL, a two-stage training framework that intrinsically improves LLM reasoning by leveraging expert attention guidance from task-specific deep learning models trained on EHR data. The first stage uses expert-guided Monte Carlo Tree Search to create high-quality, stepwise reasoning trajectories that initialize the LLM’s policy effectively. In the second stage, reinforcement learning is applied to align the LLM’s attention with clinically important features identified by expert EHR models, further optimizing the reasoning policy. Extensive experiments on two real-world EHR datasets show that EAG-RL enhances the intrinsic EHR reasoning ability of LLMs by an average of 14.62%. Moreover, the approach improves robustness to feature perturbations and generalizes better to unseen clinical domains, addressing key challenges in clinical prediction tasks. These improvements suggest that EAG-RL can enable more accurate and generalizable clinical predictions, making it a promising method for real-world healthcare applications. The authors have also made their code publicly available, facilitating further research and adoption in the field.
👉 Pročitaj original: arXiv AI Papers
