± HEALTH For the early identification, diagnosis, and treatment of mental health illnesses, the integration of deep le…± LIFESTYLE Loneliness is a pressing global health issue, yet traditional interventions often fall short due to scalabili…± POLICY Abstract As AI systems increasingly influence everyday life, integrating diverse community values is both eth…+ HEALTH Artificial intelligence has transformed digital pathology by enabling biomarker prediction from high-resoluti…+ POLICY As generative AI (GAI) becomes increasingly embedded in higher education, universities worldwide are developi…+ HEALTH ABSTRACT Efficient, accurate phenotyping for antidepressant treatment response in electronic health records (…+ HEALTH Electronic health records (EHRs) offer considerable potential for clinical prediction, but their complexity a…− HEALTH Abstract Computational pathology foundation models (CPathFMs) have emerged as a powerful approach for analyzi…
TruaceTracing the truth around AIFriday, July 17, 2026
Subscribe
Health·G Space·Evidence-backed gain·Published 2026-07-17

Large language models are powerful electronic health record encoders

Electronic health records (EHRs) offer considerable potential for clinical prediction, but their complexity and heterogeneity challenge traditional machine learning. Domain-specific electronic health record foundation models trained on unlabeled EHR data have shown improved predictive accuracy and generalization. However, their development is constrained by limited data access and site-specific vocabularies. We convert EHR data into plain text by replacing medical codes with natural-language descriptions, enabli…

TRV-2026-0249Peer-reviewedPermanent record — cite & verify
Large language models are powerful electronic health record encoders

Adulterations detected, or, Plain instructions for the discovery of frauds in food and medicine [electronic resource] by Hassall, Arthur Hill, 1817-1894. Public domain

The quick read

By July 2026, researchers reported converting electronic health records into plain text by replacing medical codes with natural-language descriptions, then using general-purpose large language models to produce embeddings for downstream clinical prediction without access to private medical training data. They tested this approach on 15 tasks from the EHRSHOT benchmark and in an external validation using UK Biobank.

The finding matters because it suggests health systems could use portable, data-independent LLM encoders instead of building site-specific EHR foundation models that require broad data access. The reported parity with CLMBR-T-Base and improvements on some UK Biobank tasks point to better vocabulary coverage and generalization, but the trade-off with computational efficiency and performance across diverse populations and settings remains to be fully characterized.

Main points
  • Authors replaced medical codes with natural-language descriptions to create plain-text EHR inputs for general-purpose LLMs.
  • LLM embeddings were evaluated on 15 clinical tasks from the EHRSHOT benchmark against CLMBR-T-Base.
  • External validation was performed in UK Biobank, with improvements attributed to higher vocabulary coverage and slightly better generalization.
  • Study frames a trade-off between computational efficiency of specialized EHR models and portability and data independence of LLM embeddings.
Gain

General-purpose LLMs can encode EHRs as plain-text descriptions to produce embeddings for clinical prediction without private medical training data, matching a specialized EHR foundation model across 15 tasks and improving on some tasks in UK Biobank external validation.

The rundown

The method converts structured EHR codes into natural-language descriptions, allowing off-the-shelf LLMs to generate high-dimensional embeddings without training on private medical data, addressing constraints of limited data access and site-specific vocabularies.

Performance was benchmarked against CLMBR-T-Base on 15 EHRSHOT tasks and externally validated in UK Biobank, where higher vocabulary coverage was cited for statistically significant gains on some tasks, while authors note specialized models retain computational efficiency advantages.

Sources

Reader signal

How should this claim be treated?

The debate