Continuous-time probabilistic models for longitudinal electronic health records

Area of Science:

• Computational biology
• Biostatistics
• Machine learning in healthcare

Background:

• Longitudinal Electronic Health Record (EHR) data analysis is crucial for precision medicine.
• Heterogeneity and irregular sampling of EHR data pose challenges for traditional Machine Learning (ML) methods.
• Existing ML models struggle to capture complex, nonlinear relationships in time-series EHR data.

Purpose of the Study:

• To present an unsupervised probabilistic model for analyzing longitudinal EHR data.
• To capture nonlinear relationships between variables over continuous time, accommodating arbitrary sampling patterns.
• To enable the evaluation of future health states using trained models.

Main Methods:

• Developed an unsupervised probabilistic model for continuous-time EHR data analysis.
• The model captures the joint probability distribution of variable measurements and time intervals.
• Inference algorithms were derived for likelihood evaluation of future outcomes.

Main Results:

• The model successfully handles arbitrary sampling patterns inherent in EHR data.
• Demonstrated application using United States Veterans Health Administration (VHA) data for diabetes and depression.
• Generated likelihood ratio maps to visualize depression risk stratification using Patient Health Questionnaire-9 (PHQ-9) scores.

Conclusions:

• The proposed unsupervised probabilistic model offers a robust approach for analyzing complex longitudinal EHR data.
• This method enhances the applicability of machine learning in precision medicine by addressing data heterogeneity.
• The model provides valuable tools for risk assessment and prediction in clinical settings, exemplified by depression risk analysis.