Large language models improve transferability of electronic health record-based predictions across countries and...

Area of Science:

• Artificial Intelligence in Healthcare
• Biomedical Informatics
• Computational Epidemiology

Background:

• Healthcare systems exhibit variations in medical practices and reporting standards.
• This heterogeneity limits the transferability of predictive models developed from electronic health record (EHR) data.
• Existing models often struggle to generalize across diverse clinical datasets.

Purpose of the Study:

• To introduce GRASP, a novel transformer-based architecture for enhancing the generalizability of EHR-based prediction models.
• To embed medical codes into a unified semantic space using a large language model (LLM) for improved model transferability.
• To enable accurate and transferable disease risk prediction across heterogeneous healthcare systems.

Main Methods:

• Developed GRASP, a transformer-based architecture leveraging LLMs for medical code embedding.
• Applied GRASP to predict 21 diseases and all-cause mortality in over one million individuals from UK Biobank, FinnGen, and Mount Sinai datasets.
• Harmonized data to the OMOP common data model for initial training and evaluation.

Main Results:

• GRASP achieved an average ΔC-index 83% higher than language-unaware models when evaluated in FinnGen and 35% higher in Mount Sinai.
• Demonstrated significantly higher correlations with polygenic risk scores for 62% of diseases.
• Maintained robust performance predicting disease risk from non-harmonized ICD-10-CM codes, without direct OMOP mappings.

Conclusions:

• GRASP significantly enhances the generalizability and transferability of EHR-based prediction models.
• The LLM-driven semantic embedding of medical codes overcomes limitations of data heterogeneity.
• GRASP offers a resource-efficient solution for accurate, cross-system disease prediction.