Functional clustering methods for longitudinal data with application to electronic health records.

Area of Science:

• Biostatistics
• Health Informatics
• Machine Learning

Background:

• Longitudinal data analysis is crucial for understanding disease progression and patient trajectories.
• Current methods for patient phenotyping and outcome identification often rely on diagnosis codes or prescriptions, which may not capture the full clinical picture.
• Electronic health records (EHRs) offer rich longitudinal data, but effective analytical methods are needed to extract meaningful insights.

Purpose of the Study:

• To develop a novel statistical method for estimating subject-level trajectory functions from longitudinal data.
• To enable patient phenotyping, feature extraction, and outcome identification using laboratory test results.
• To improve the prediction of disease status by integrating longitudinal outcomes and baseline covariates.

Main Methods:

• A joint model for continuous longitudinal outcomes and baseline covariates was developed using an enriched Dirichlet process prior.
• The model decomposes into semiparametric linear mixed models for outcomes and marginal models for covariates.
• A nonparametric enriched Dirichlet process prior was applied to regression coefficients, error variance, and predictor space parameters, facilitating patient clustering.

Main Results:

• The proposed method enables prediction of outcomes at unobserved time points for existing and new subjects.
• Improved prediction accuracy was observed compared to mixed models with Dirichlet process priors, especially with a large number of covariates.
• The method was demonstrated on EHR data of patients initiating second-generation antipsychotic medications, predicting laboratory values indicative of diabetes.

Conclusions:

• The developed method provides a robust framework for analyzing longitudinal data for patient phenotyping and outcome identification.
• The approach enhances disease prediction accuracy by leveraging detailed patient trajectories and laboratory values.
• This methodology holds promise for early detection and management of conditions like diabetes in patient populations.