Title: ChatHealthAI: Harmonizing Electronic Health Record Embeddings with Large Language Models for Clinically Grounded Reasoning

Abstract: While Large Language Models (LLMs) demonstrate robust capabilities in natural-language reasoning for clinical decision support, they often falter when attempting to effectively model structured, longitudinal Electronic Health Records (EHRs). Conversely, foundation models designed for EHRs excel at generating predictive patient representations but typically lack the capacity for interpretable, language-based reasoning. To address this discrepancy, we introduce ChatHealthAI, a multimodal reasoning framework. This system aligns structured EHR representations derived from a pretrained EHR foundation model with the semantic space of a frozen LLM, utilizing a task-aware resampler. By synthesizing longitudinal patient data with refined descriptions of clinical events, ChatHealthAI facilitates clinically grounded natural-language reasoning without compromising accurate patient prediction. We assessed the framework’s performance on three clinical predictive tasks within the EHRSHOT benchmark. The results indicate that ChatHealthAI enhances both the quality of reasoning and interpretability, all while maintaining competitive predictive accuracy. These outcomes underscore the significant potential of combining EHR foundation models with pretrained LLMs to achieve interpretable clinical predictions.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC