ChatHealthAI bridges EHR data and LLMs for interpretable clinical reasoning
New framework aligns structured patient records with language models for better predictions.
Large language models excel at natural-language reasoning for clinical decision support but struggle with structured longitudinal electronic health records (EHRs). Conversely, EHR foundation models learn predictive patient representations but lack interpretable language-based reasoning. To bridge this gap, researchers introduce ChatHealthAI — a multimodal reasoning framework that aligns structured EHR embeddings from a pretrained EHR foundation model with the semantic space of a frozen LLM using a task-aware resampler. This integration allows the LLM to perform clinically grounded reasoning while maintaining competitive predictive performance.
ChatHealthAI was evaluated on three clinical predictive tasks from the EHRSHOT benchmark. Results demonstrate that it improves both reasoning quality and interpretability without sacrificing accuracy. The framework effectively transforms longitudinal patient data into refined clinical event descriptions, enabling the LLM to generate explainable predictions. This work highlights the potential of combining EHR foundation models with pretrained LLMs for interpretable clinical decision support, paving the way for more transparent AI-assisted healthcare.
- ChatHealthAI aligns structured EHR representations from a pretrained foundation model with a frozen LLM using a task-aware resampler.
- Evaluated on three clinical predictive tasks from the EHRSHOT benchmark, maintaining competitive accuracy while boosting interpretability.
- Enables clinically grounded natural-language reasoning by integrating longitudinal patient data with refined event descriptions.
Why It Matters
Makes AI-driven clinical predictions interpretable, combining structured data and language reasoning for safer healthcare decisions.