Related Experiment Video
Updated: Aug 9, 2026
06:19
Constructing and Visualizing Models using Mime-based Machine-learning Framework
Published on: July 22, 2025
Dynamic Individualized Prediction of Binary Clinical Outcomes Using Bayesian Generalized Linear Mixed Models With
Zixing Liu1, Adrianne Woods1, Laura K Hummers1
1Division of Rheumatology, Johns Hopkins University School of Medicine, Baltimore, Maryland, USA.
Statistics in Medicine
|August 7, 2026
Summary
A new Bayesian framework enables accurate, individualized predictions of critical binary health events using longitudinal electronic health record data. This approach dynamically updates risk estimates for better chronic disease management and precision medicine.
Area of Science:
- Biostatistics
- Medical Informatics
- Machine Learning
Background:
- Accurate prediction of critical events from longitudinal electronic health record (EHR) data is vital for precision medicine in complex chronic diseases.
- Existing prediction methods often lack transparency, ignore temporal dynamics, or are difficult to update, limiting clinical use, especially for binary outcomes.
- Current dynamic prediction methods are typically designed for continuous, not binary, outcomes common in clinical practice.
Purpose of the Study:
- To develop a novel dynamic prediction algorithm for binary outcomes using longitudinal EHR data.
- To create a generalizable framework for individualized, visit-by-visit risk estimation in chronic disease management.
- To improve the transparency, temporal dependence handling, and dynamic update capabilities of predictive models.
Main Methods:
- A Bayesian generalized linear mixed model approach was used to develop a cross-validated sequential prediction (CVSP) algorithm.
- The model integrates population-level fixed effects, lagged outcomes, and patient-specific random intercepts.
- Variable selection using bootstrap LASSO and Monte Carlo integration were employed for model parsimony and efficient prediction without refitting.
Main Results:
- The CVSP framework demonstrated superior discriminative performance (cross-validated AUC of 0.86) in predicting proximal muscle weakness in systemic sclerosis patients.
- The method showed good calibration and outperformed standard regression and machine learning approaches.
- Individualized predictions incorporating a patient's full medical history were efficiently generated, with variable selection preserving accuracy.
Conclusions:
- The Bayesian CVSP framework offers a generalizable method for individualized, visit-by-visit prediction of binary outcomes from longitudinal EHR data.
- Dynamic risk updates without model refitting support real-time clinical decision-making.
- This approach advances precision medicine for managing chronic diseases with evolving patient trajectories.
