Bayesian latent multi-state modeling for nonequidistant longitudinal electronic health records

Area of Science:

• Biostatistics
• Longitudinal Data Analysis
• Health Informatics

Background:

• Longitudinal health records offer insights into disease progression but are often sparse and irregularly observed due to patient-system interactions.
• Existing models struggle to capture the continuous-time latent processes underlying these irregular health observations.
• Accurate modeling is crucial for dynamic health monitoring and understanding disease trajectories.

Purpose of the Study:

• To develop a novel continuous-time hidden Markov model (CT-HMM) for analyzing longitudinal health data with irregular observations.
• To account for time-varying covariates and different types of observations within the model.
• To implement efficient Bayesian inference methods for practical application to large datasets.

Main Methods:

• Development of a continuous-time hidden Markov model tailored for irregular longitudinal health records.
• Utilization of a specific missing data likelihood formulation for computational efficiency.
• Application of Bayesian inference facilitated by expectation-maximization and Markov chain Monte Carlo algorithms.

Main Results:

• The proposed CT-HMM effectively analyzes longitudinal data characterized by irregular visits and diverse observation types.
• Simulation studies confirm the efficient implementation of Bayesian inference methods for large-scale datasets.
• The model successfully applied to a chronic obstructive pulmonary disease cohort, analyzing drug counts using covariates.

Conclusions:

• The developed continuous-time hidden Markov model provides a robust framework for analyzing complex longitudinal health data.
• The Bayesian inference approach ensures efficient and accurate modeling, even with large and sparse datasets.
• This methodology enhances the understanding of disease progression and can inform clinical decision-making in chronic conditions.