Clustering functional data with application to electronic medication adherence monitoring in HIV prevention trials

Maintaining high medication adherence is essential for achieving desired efficacy in clinical trials, especially prevention trials. However, adherence is traditionally measured by self-reports that are subject to reporting biases and measurement error. Recently, electronic medication dispenser devices have been adopted in several HIV pre-exposure prophylaxis prevention studies. These devices are capable of collecting objective, frequent, and timely drug adherence data. The device opening signals generated by such devices are often represented as regularly or irregularly spaced discrete functional data, which are challenging for statistical analysis. In this paper we focus on clustering the adherence monitoring data from such devices. We first pre-process the raw discrete functional data into smoothed functional data. Parametric mixture models with change-points, as well as several non-parametric and semi-parametric functional clustering approaches are adapted and applied to the smoothed adherence data. Simulation studies were conducted to evaluate finite sample performances, on the choices of tuning parameters in the pre-processing step as well as the relative performance of different clustering algorithms. We applied these methods to the HIV Prevention Trials Network(HPTN) 069 study for identifying subgroups with distinct adherence behavior over the study period.