Schema-Adaptive Tabular Representation Learning with LLMs for Generalizable Multimodal Clinical Reasoning

Researchers propose Schema-Adaptive Tabular Representation Learning, which uses LLMs to convert structured clinical data into semantic embeddings that transfer across different electronic health record schemas without retraining. When combined with imaging data for dementia diagnosis, the method achieves state-of-the-art results and outperforms board-certified neurologists on retrospective diagnostic tasks.

This research addresses a fundamental limitation in machine learning: the inability of tabular models to generalize across different data schemas. Traditional ML systems trained on one hospital's EHR structure fail when encountering another institution's data format, requiring costly retraining and feature engineering. The authors solve this by leveraging LLMs as semantic intermediaries, converting database columns and values into natural language representations that capture meaning beyond raw data types.

The approach reflects a broader trend of using foundation models to bridge domain-specific technical challenges. Rather than training specialized models for each data schema, the team bootstraps LLM understanding of clinical semantics to create universally interpretable embeddings. This aligns with how LLMs already excel at transferring knowledge across diverse tasks through language abstraction.

The clinical validation is particularly significant. Achieving diagnostic performance that exceeds practicing neurologists demonstrates real-world applicability in high-stakes domains. The multimodal integration—combining tabular EHR data with imaging—shows how structured and unstructured data can complement each other through LLM-driven reasoning. The zero-shot transfer capability eliminates a major barrier to healthcare AI deployment, where rigid schemas have historically prevented model portability between healthcare systems.

For healthcare technologists and AI infrastructure providers, this validates LLMs as practical tools for interoperability challenges that have plagued medical informatics. The work suggests expanding LLM applications beyond NLP into structured data processing, potentially accelerating AI adoption in regulated, heterogeneous environments where schema variance has been a persistent obstacle.

• →LLMs enable zero-shot transfer across incompatible EHR schemas by converting structured data into semantic language representations
• →The method outperforms board-certified neurologists on dementia diagnosis when combined with MRI imaging data
• →Schema-adaptive embeddings eliminate the need for manual feature engineering and retraining when deploying models across different healthcare systems
• →The approach demonstrates LLMs' utility for structured data problems beyond natural language processing
• →Successful validation on NACC and ADNI datasets suggests practical scalability in real clinical environments