Improving polygenic risk prediction performance by integrating electronic health records through phenotype embedding

Area of Science:

• Genetics
• Biomedical Informatics
• Computational Biology

Background:

• Electronic health records (EHRs) offer rich clinical data for disease prediction.
• Traditional polygenic risk scores (PRS) have limitations in capturing complex EHR phenotypes.
• Integrating EHR data with genomic information is crucial for advancing predictive models.

Purpose of the Study:

• To introduce EHR-embedding-enhanced PRS (EEPRS) for improved disease prediction using EHR phenotype embeddings.
• To leverage advanced embedding techniques (Word2Vec, GPT) with GWAS summary statistics.
• To enhance the accuracy and interpretability of PRS by incorporating complex EHR structures.

Main Methods:

• Developed EEPRS by deriving phenotype embeddings from EHRs.
• Utilized Word2Vec and GPT for EHR embedding generation.
• Conducted EHR-embedding-based GWASs and hierarchical clustering to identify trait clusters.
• Compared EEPRS performance against single-trait PRS and MTAG_PRS in UK Biobank and All of Us cohorts.
• Developed EEPRS_optimal and MTAG_EEPRS for data-adaptive and multi-trait PRS optimization.

Main Results:

• EEPRS consistently outperformed single-trait PRS across 41 traits in the UK Biobank, especially for a cardiovascular cluster.
• EHR-embedding-based PRS showed robust associations with circulatory system diseases.
• EEPRS_optimal and MTAG_EEPRS demonstrated further improvements in prediction accuracy.
• Validation in the All of Us cohort confirmed EEPRS benefits for selected diseases.

Conclusions:

• EEPRS provides a robust and interpretable framework for enhancing PRS.
• Integrating EHR embeddings significantly improves the predictive power of both single-trait and multi-trait PRS.
• EEPRS represents a significant advancement in leveraging large-scale biobank data for personalized disease risk prediction.