Innovative AI Converts Hospital Data into Readable Text to Enhance Emergency Care

Researchers at the University of California, Los Angeles (UCLA) have introduced a groundbreaking artificial intelligence (AI) system designed to transform complex electronic health records (EHR) into comprehensible narrative texts. Traditionally, hospital data is stored in detailed tables filled with codes, numbers, and categories, making it difficult for AI models—primarily designed to process text—to leverage this information effectively. To address this, the team developed the Multimodal Embedding Model for EHR (MEME), which converts fragmented EHR data into 'pseudo-notes' resembling clinical documentation.

These pseudo-notes are created by breaking down patient data into specific categories such as medications, vital signs, diagnostics, and more. Each category is transformed into natural language text using simple templates and then encoded with language models, emulating medical reasoning. This approach enables AI systems to analyze patient histories more accurately by treating different streams of health data as related narratives.

The system was tested across extensive emergency department datasets, including over 1.3 million visits from the Medical Information Mart for Intensive Care (MIMIC) database. Results demonstrated that MEME consistently outperformed traditional machine learning models, specialized healthcare AI frameworks like CLMBR and Clinical Longformer, and prompt-based methods in various clinical decision support tasks.

Importantly, the approach exhibited strong portability across different hospital systems and coding standards, suggesting broad applicability. The research team plans to further explore MEME's use in other clinical settings beyond emergency departments and work on enhancing its performance across diverse healthcare institutions. They aim to adapt the system for evolving medical concepts and data standards, making advanced AI more accessible for healthcare providers.

This development holds significant promise for emergency medicine, where rapid, accurate decision-making is critical. By converting EHRs into a text format that advanced language models can interpret, medical professionals could potentially improve patient outcomes and streamline clinical workflows.

The full study is published in the journal npj Digital Medicine. For more information, visit source.