Bayesian Double Feature Allocation for Phenotyping with Electronic Health Records.

Area of Science:

• Computational biology
• Statistical genetics
• Health informatics

Background:

• Electronic health records (EHR) offer valuable data for understanding complex human phenotypes.
• Statistical modeling of EHR data can reveal underlying disease patterns and patient subgroups.

Purpose of the Study:

• To propose a novel categorical matrix factorization method for inferring latent diseases from EHR data.
• To enhance the identifiability and interpretability of latent diseases using Bayesian approaches and prior knowledge of known diseases.

Main Methods:

• A double feature allocation model was developed to simultaneously allocate features to rows and columns of a categorical matrix.
• Bayesian inference was employed, incorporating prior information on known diseases like hypertension and diabetes.
• The method was validated through simulation studies and compared with sparse latent factor models.

Main Results:

• Application to a Chinese EHR dataset identified 10 latent diseases.
• These latent diseases were associated with specific health traits including lipid disorders, thrombocytopenia, polycythemia, anemia, infections, allergy, and malnutrition.
• The identified latent diseases showed agreement with findings reported in medical literature.

Conclusions:

• The proposed method effectively infers latent diseases from EHR data, offering insights into complex health conditions.
• This approach can assist healthcare officials in monitoring patient health, identifying risk factors, and developing preventive strategies.
• An R package ('dfa') and a web application are available for implementing the method and exploring the findings.