HMN 2025: How AI model converts hospital information into textual content for higher emergency care choices

UCLA researchers have developed an AI system that turns fragmented digital well being information (EHR) usually in tables into readable narratives, permitting synthetic intelligence to make sense of advanced affected person histories and use these narratives to carry out scientific choice assist with excessive accuracy. The Multimodal Embedding Model for EHR (MEME) transforms tabular well being knowledge into “pseudonotes” that mirror scientific documentation, permitting AI models designed for textual content to investigate affected person info extra successfully.

The work is published within the journal npj Digital Medicine.

Electronic well being information comprise huge quantities of affected person info that might assist docs make quicker, extra correct choices in emergency conditions. However, most cutting-edge AI models work with textual content, whereas hospital knowledge is saved in advanced tables with numbers, codes, and classes. This mismatch has prevented well being care methods from absolutely leveraging superior AI capabilities. Emergency departments, where fast choices will be crucial, significantly want instruments that may quickly course of complete affected person histories to foretell outcomes and information remedy choices.

Researchers created a novel strategy that converts tabular digital well being document knowledge into text-based “pseudonotes” utilizing medical documentation shortcuts generally utilized by well being care suppliers. In different phrases, as a substitute of treating the EHR as a set of codes, the pseudonotes model creates a narrative composed of a number of narratives.

The system breaks affected person knowledge into concept-specific blocks (drugs, triage vitals, diagnostics, and so on.), remodeling every into textual content utilizing easy templates, after which encodes each individually utilizing language models. It basically emulates a type of medical reasoning.

The crew then fed this textual content to superior language models, treating several types of well being info—like lab outcomes, diagnoses, and drugs—as separate however associated knowledge streams. The crew examined their system in opposition to conventional machine {learning} strategies, specialised well being care AI models, and prompting-based approaches utilizing actual emergency division prediction duties.

Across over 1.3 million emergency room visits from the Medical Information Mart for Intensive Care (MIMIC) database and UCLA datasets, MEME constantly outperformed current approaches throughout a number of emergency division choice assist duties. The multimodal textual content strategy, which processes totally different elements of well being information individually, achieved higher outcomes than making an attempt to mix all info right into a single illustration.

The system demonstrated superior efficiency to conventional machine {learning} methods, EHR-specific basis models like CLMBR and Clinical Longformer, and customary prompting strategies. The strategy additionally confirmed good portability throughout totally different hospital methods and coding requirements.

The analysis crew plans to check MEME’s effectiveness in different scientific settings past emergency departments to validate its broader applicability. They additionally purpose to deal with limitations noticed in cross-site model generalizability, working to make sure the system performs constantly throughout totally different well being care establishments. Future work will give attention to extending the strategy to accommodate new medical ideas and evolving well being care knowledge requirements, doubtlessly making superior AI extra accessible to well being care methods.

“This bridges a crucial hole between probably the most highly effective AI models obtainable at the moment and the advanced {reality} of well being care knowledge,” stated Simon Lee, Ph.D. scholar at UCLA Computational Medicine.

“By changing hospital information right into a format that superior language models can perceive, we’re unlocking capabilities that had been beforehand inaccessible to well being care suppliers. The indisputable fact that this strategy is extra transportable and adaptable than current well being care AI methods may make it significantly worthwhile for establishments working with totally different knowledge requirements.”