|
Abstract:
|
Recently, medical studies have focused on the analysis of dense multivariate physiological data, such as Ventricular Assist Device (VAD) parameters, recorded over multiple time scales, e.g., both within a circadian cycle and longitudinally across multiple circadian cycles. The analytic challenge with this data structure is to model the population and subject specific cyclic behavior (fast time scale, t) as well as the longitudinal evolution (slow time scale, s) of the cycles. We propose a multiscale decomposition of these two independent time scales (t,s) and assume that the data is generated by two functions f(t) and g(s). The estimation of the f(t) and g(s) is done in the wavelet domain and broken down into two stages: a Functional Mixed Model is used for the fast time scale and a Functional Principal Components Analysis is used for the slow time scale. The time scales are linked through the multiscale decomposition. We demonstrate our method using data from a cohort of VAD patients. Simulation studies indicate that our proposed model effectively characterizes the population's average cardiac trajectory and the between subject variability.
|
