Enhancing stratification for survival analyses across standardized data sources.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Data Standardization

Background:

• Patient stratification is vital for personalized medicine but hindered by fragmented healthcare data.
• The Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM) standardizes diverse data sources.
• Transformer models like BERT show promise in extracting patient insights from electronic health records.

Purpose of the Study:

• To evaluate BERT-based patient representation learning using OMOP CDM data for effective patient stratification.
• To assess knowledge transfer and improved survival analysis through standardized data and advanced machine learning.

Main Methods:

• Harmonized MIMIC-IV-2.2 and German lung cancer registry data within the OMOP CDM framework.
• Pre-trained BERT on MIMIC-IV-2.2, generated patient embeddings for cancer registry data.
• Employed k-means clustering on embeddings for stratification, followed by survival analysis and clinical expert review.

Main Results:

• Successfully stratified patients and identified similarities across disparate datasets.
• Demonstrated up to 11% improvement in survival analysis over naïve k-means clustering.
• Achieved up to 7% increase in survival analysis accuracy for specific patient subgroups.

Conclusions:

• Standardized data and transformer models facilitate knowledge transfer between diverse datasets.
• Effective patient stratification using embeddings enhances survival analysis accuracy for identified subgroups.