MixEHR-Guided: A guided multi-modal topic modeling approach for large-scale automatic phenotyping using the...

Area of Science:

• Clinical informatics
• Biomedical data science
• Machine learning for healthcare

Background:

• Electronic Health Records (EHRs) offer vast clinical data but lack reliable disease labels for research.
• Manual phenotyping using billing codes is labor-intensive, prone to bias, and often incomplete.
• Existing unsupervised methods struggle with topic identifiability and accurate phenotype representation.

Purpose of the Study:

• To introduce MixEHR-Guided (MixEHR-G), a novel multimodal hierarchical Bayesian topic model for automated EHR phenotyping.
• To leverage surrogate features for aligning latent topics with known clinical phenotypes.
• To enhance the interpretability and efficiency of extracting disease information from EHR data.

Main Methods:

• Developed MixEHR-Guided (MixEHR-G), a Bayesian topic model integrating EHR data.
• Utilized prior information from surrogate features to guide phenotype identification.
• Applied the model to large-scale EHR (MIMIC-III) and claims datasets (PopHR).

Main Results:

• MixEHR-G successfully identified interpretable phenotypes and revealed insights into phenotype similarities and comorbidities.
• The model demonstrated superior performance over existing unsupervised methods in phenotype label annotation.
• Accurate estimation of relative phenotype prevalence functions was achieved without gold-standard labels.

Conclusions:

• MixEHR-G offers an interpretable and automated approach to phenotyping using EHR data.
• This method facilitates improved disease risk prediction and clinical decision support.
• Automated phenotyping with MixEHR-G represents a significant advancement in leveraging EHR data for research and clinical applications.