level: research
large language models can reason well in clinical settings but have trouble with structured electronic health records. ehr foundation models predict patient outcomes but lack clear language-based reasoning. a new framework called chathealthai aims to combine both strengths. it uses a pretrained ehr model and a frozen llm, connected by a task-aware resampler. this resampler aligns the ehr data with the llm's semantic space. the system also uses refined clinical event descriptions to add context.
the framework was tested on three clinical prediction tasks from the ehrshot benchmark. it showed better reasoning quality and interpretability compared to existing methods. by integrating longitudinal patient data with language, chathealthai can produce natural-language explanations for its predictions. this helps clinicians understand the model's decisions. the approach keeps the llm frozen, so it does not need retraining. the ehr model provides the structured data representation.
the work addresses a key challenge in medical ai: making models both accurate and explainable. ehr data is complex and time-based, which is hard for standard llms to process. by aligning the two model types, chathealthai offers a way to use llm reasoning on top of reliable patient representations. this could lead to better clinical decision support tools. the resampler design allows the system to focus on relevant parts of the patient history for each task.
why it matters: this approach could make clinical ai more trustworthy by providing clear reasons for predictions, helping doctors make better decisions with patient data.