Mixture models for longitudinal viral load trajectories with heterogeneity and skewness in the multicenter AIDS cohort study

Background: Viral load (plasma HIV-1 RNA copies) trajectories which play an important role in reflecting the efficacy of antiretroviral therapy, are observed with the following features: (i) patients may come from a heterogeneous population with more than one mean trajectories; (ii) this longitudinal measure may suffer from a serious departure of normality. However, most of HIV dynamic modelling ignored these features which may lead invalid inference. We applied a novel statistical modelling to identify the different trajectories of viral load, and to estimate unobserved membership probabilities at both population and individual levels.

Methods: Data from the multicenter AIDS cohort Study (MACS), in which HIV+ subjects have viral load measured repeated visits, is used. A flexible finite mixture of nonlinear mixed-effects models with skew distributions that allows estimates of both parameters and class membership probabilities through various scenarios, is applied to MACS data under Bayesian framework.

Results: The simulation showed finite mixture of nonlinear mixed-effects models with skewed-normal or skewed t-distribution fitted the data better than models with other non-skewness distributions. Three different patterns of viral load trajectories were found. It is an ongoing study, more results will come out.

Conclusion: The proposed method may fit the data better than traditional homogeneous linear mixed-effect models, in the presence of heterogeneity and skewness of viral load responses. This appropriate statistical inference for HIV dynamics provide evidence for making reliable clinical decisions and predicting disease prognosis.