Risk-Adjusted Incidence Modeling on Hierarchical Survival Data with Recurrent Events

TL;DR

This paper develops a risk-adjusted hierarchical survival model for recurrent infection events in cystic fibrosis patients, addressing data complexities like missingness and multilevel correlations to improve infection rate monitoring.

Contribution

It extends the Andersen-Gill model with a step-by-step pipeline that accounts for missing data and hierarchical structures, validated through simulations and real registry data.

Findings

Model coverage was close to the desired confidence level.

Recent years provided better predictions of future infection rates.

The pipeline effectively monitors recurrent infection events in healthcare data.

Abstract

There is a constant need for many healthcare programs to timely address problems with infection prevention and control (IP&C). For example, pathogens can be transmitted among patients with cystic fibrosis (CF) in both the inpatient and outpatient settings within the healthcare system even with the existing recommended IP&C practices, and these pathogens are often associated with negative clinical outcomes. Because of limited and delayed data sharing, CF programs need a reliable method to track infection rates. There are three complex structures in CF registry data: recurrent infections, missing data, and multilevel correlation due to repeated measures within a patient and patient-to-patient transmissions. A step-by-step analysis pipeline was proposed to develop and validate a risk-adjusted model to help healthcare programs monitor the number of recurrent events while taking into account missing data and the hierarchies of repeated measures in right-censored data. We extended the mixed-effect Andersen-Gill model (the frailty model), adjusted for important risk factors, and provided confidence intervals for the predicted number of events where the variability of the prediction was estimated from three identified sources. The coverage of the estimated confidence intervals was used to evaluate model performance. Simulation results indicated that the coverage of our method was close to the desired confidence level. To demonstrate its clinical practicality, our pipeline was applied to monitor the infection incidence rate of two key CF pathogens using a U.S. registry. Results showed that years closer to the time of interest were better at predicting future incidence rates in the CF example.