Clustering of recurrent events data applied to the re-admission of elderly people at hospital

Description

The recurrent admissions of elderly people at hospital can be modeled by a censored counting process. The intensity of this processe can be specified as functions of covariates as in the Andersen-Gill model. However, accounting for these covariates is often not sufficient to explain the observed inter-patient heterogeneity. We propose a mixture model (also called a latent class model) which takes into account this heterogenity and allows to perform the patients' clustering. Within each cluster, the recurrent events process intensity is modeled by a parametric baseline intensity (Weibull) adjusted by the effect of the covariates. The model parameters are estimated by maximum likelihood using the EM algorithm, and the BIC criterion is adopted to choose the optimal numbers of clusters. The behaviour of the model is studied on simulated data. The analysis of two real datasets is also performed. The first one (PAERPA cohort) contains administrative data on hospital admission dates, death and basic clinical/demographic variables for over 35000 older persons. The second one (DAMAGE cohort) contains more detailed data on over 3500 older persons: dates of hospital admissions, clinical, biological, socio-demographic variables.

Abstract

International audience

Additional details