Dynamic clinical prediction models for discrete time-to-event data with competing risks-A case study on the...

Area of Science:

• Biostatistics
• Clinical Epidemiology
• Infectious Disease Modeling

Background:

• Clinical prediction models are crucial for patient care, requiring careful selection of predictor variables.
• Existing Bayesian variable selection methods are established for linear, generalized linear, and Cox models.
• Discrete time-to-event data with competing risks present unique challenges for model development.

Purpose of the Study:

• To propose a novel methodology for developing a clinical prediction model for ventilator-associated pneumonia (VAP) risk.
• To address the complexities of discrete time-to-event data and competing risks in intensive care units.
• To enable dynamic selection of relevant predictors based on time at risk.

Main Methods:

• Utilizing a landmark approach for dynamic Bayesian variable selection.
• Applying cause-specific variable selection to determine the direct impact of variables on competing events.
• Focusing on predicting the daily risk of *P. aeruginosa* (PA) VAP.

Main Results:

• The developed methodology allows for time-dependent predictor relevance in clinical models.
• Identifies specific variables that impact the risk of PA VAP versus competing events.
• Provides a framework for dynamic risk assessment in intensive care settings.

Conclusions:

• The proposed dynamic Bayesian variable selection approach enhances prediction models for time-to-event data with competing risks.
• This method offers a more accurate and adaptive approach to predicting VAP risk in ICUs.
• Cause-specific variable selection clarifies the independent effects of predictors on distinct adverse outcomes.