BAYESIAN ANALYSIS FOR IMBALANCED POSITIVE-UNLABELLED DIAGNOSIS CODES IN ELECTRONIC HEALTH RECORDS

Area of Science:

• Health Informatics
• Machine Learning
• Biostatistics

Background:

• Electronic health records (EHR) offer abundant data but suffer from noise, missing values, and imbalanced labels.
• A significant challenge is classifying patients with undiagnosed conditions due to incomplete diagnostic information, creating a positive-unlabelled learning scenario.

Purpose of the Study:

• To develop a robust model-based approach for classifying unlabelled patients within EHR data.
• To address the challenges of positive-unlabelled learning and imbalanced datasets in predictive health analytics.

Main Methods:

• Utilized a Bayesian finite mixture model for patient classification.
• Implemented a consensus Monte Carlo approach to manage label switching and enhance computational efficiency in imbalanced data.

Main Results:

• The proposed model-based approach demonstrated superior performance compared to existing positive-unlabelled learning algorithms in simulation studies.
• Application to Cerner EHR data for detecting diabetic retinopathy (DR) revealed an estimated DR prevalence of 25%, despite only 3% confirmatory diagnoses.

Conclusions:

• The developed Bayesian approach effectively classifies patients in positive-unlabelled and imbalanced EHR datasets.
• This method provides a reliable tool for estimating disease prevalence and identifying at-risk patients, as demonstrated by the diabetic retinopathy detection example.