Adaptive modeling of longitudinal viral load data

Background. We have recently developed methods for adaptively modeling repeated measures data. These methods use likelihood cross-validation along with heuristic search through fractional polynomial models to identify nonparametric regression models for expected outcomes while accounting for correlation for repeated outcome measurements. We describe these methods and demonstrate them with an example analysis of HIV viral load in terms of antiretroviral adherence measured electronically using MEMS caps.

Methods. Viral loads were obtained from medical records and matched in time to interview dates. Adherence prior to an interview was measured by percent prescribed administrations taken based on MEMS cap openings up to that interview. Data consisted of 643 measurements for 160 subjects over 7 time points. Log (base 10) viral load over time was adaptively modeled in terms of prior adherence.

Results. Adherence ranged from 0.7-100% and log viral loads from 1.30 to 5.88 log copies/mL. Mean log viral load decreased nonlinearly with increased adherence, but nearly logarithmically suggesting that there was little impact to high adherence on improving viral load levels. Similar results held for post-baseline viral load conditioned on baseline viral load. However, baseline viral load adaptively grouped into low and high initial viral load levels with a nearly linear impact to increasing adherence for low levels and a constant impact for high levels. The distinct impact of adherence for low levels was masked by the constant impact for high levels when analyzed together.

Conclusions. Adaptive modeling can provide novel insights into repeated measures data like longitudinal viral load data.