Fast Variational Bayesian Inference for Correlated Survival Data: An Application to Invasive Mechanical Ventilation...

Area of Science:

• Biostatistics
• Clinical Epidemiology
• Health Informatics

Background:

• Correlated survival data are common in clinical research, particularly in intensive care units (ICUs).
• Patients within the same ICU share characteristics, leading to correlated mechanical ventilation durations.
• Existing statistical models may not fully capture intra-cluster correlation in survival data.

Purpose of the Study:

• To develop and evaluate a statistical model for analyzing correlated survival data in the context of invasive mechanical ventilation.
• To introduce a novel, computationally efficient variational Bayes (VB) algorithm for parameter inference in shared frailty models.
• To investigate the impact of ICU-specific factors on mechanical ventilation duration.

Main Methods:

• A shared frailty log-logistic accelerated failure time model with a cluster-specific random intercept was employed.
• A novel, fast variational Bayes (VB) algorithm was developed for parameter estimation.
• Simulation studies were conducted to assess algorithm performance under varying cluster numbers and sizes.
• The VB algorithm's performance was compared against the h-likelihood method and a Markov Chain Monte Carlo (MCMC) algorithm.

Main Results:

• The proposed VB algorithm demonstrated satisfactory performance in parameter estimation.
• The VB algorithm showed significant computational efficiency compared to the MCMC algorithm.
• The analysis of ICU ventilation data revealed a significant ICU-site random effect on ventilation duration.

Conclusions:

• The shared frailty log-logistic accelerated failure time model effectively accounts for intra-cluster correlation in survival data.
• The novel VB algorithm provides an efficient and accurate method for analyzing such data.
• The findings highlight the importance of accounting for site-specific effects in multi-center ICU studies.