A machine learning approach to leveraging electronic health records for enhanced omics analysis.

Summary

COMET is a multimodal transfer-learning framework that pretrains on large EHR datasets and fuses clinical and omics data to improve modeling and discovery in small omics cohorts. Across two independent datasets, COMET outperformed traditional omics-only methods in prediction and biological insight. It enables more granular patient stratification beyond binary case–control labels.

Key Findings

• Introduces COMET, a transfer-learning framework integrating EHR and omics via early and late fusion.
• Across two independent datasets, COMET improved predictive performance versus omics-only analyses.
• COMET enhanced biological discovery and enabled more precise, non-binary patient classifications.

Clinical Implications

While not a clinical trial, COMET may accelerate biomarker discovery and risk stratification in perioperative medicine (e.g., predicting delirium, pain phenotypes, or AKI) by enabling more powerful analyses with existing EHR-linked cohorts.

Limitations

• Exact cohort sizes and public code/data availability are not specified in the abstract.
• Clinical utility still requires prospective validation and workflow integration.

Future Directions

Prospective studies embedding COMET into perioperative registries to drive biomarker validation and clinical decision support; broader benchmarking across surgical/anesthesia indications.