Multiple polyexponentials and quasipolynomials as empirical nonlinear regression models: a case study with HIV viral load data.

Measurements of HIV1-RNA plasma concentrations are an important method of assessing patient response to anti-HIV1 treatment, and in most clinical trials of such treatments HIV1-RNA levels are assessed at regular intervals of time. HIV1-RNA levels in successfully treated patients tend to follow a standard pattern of biphasic decline-a rapid early decline in viral load, followed by a period of slower decline or a steady level. Fitting nonlinear regression models to these patterns of declining HIV1-RNA levels can be of value in comparing different treatment regimes and in predicting treatment outcome. Simple exponential-decline models can give an adequate fit to the typical pattern of HIV1-RNA decline, but we have explored the extent to which curve-fitting can be improved by using two novel nonlinear model forms. Specifically, we describe the fitting of multiple polyexponential and quasipolynomial forms to longitudinal HIV1-RNA plasma data collected in two recent trials of the novel anti-HIV1 treatment Fuzeon. We comment on the practicalities of fitting these nonlinear models, and compare the fit using various criteria.