Rotate to view desktop version ... Statistical ModelingLinear Mixed Effects Models
Linear mixed effects (LME) models are optimally suited for the analysis of longitudinal data. They can deal well with missing data, drop-outs or different time intervals. In a collaboration with Dr. Sabuncu we have introduced LME models to neuroimaging applications [3] and extended it for the analysis of spatio-temporal signals on the cortical surface [2]. Proximity regions on the cortex demonstrate similar correlation patterns, a fact that can be exploited to increase statistical power as well as improve run-time of the LME models. We furthermore extended the model to time-to-event analysis in [1], by sampling missing data points from the LME fit, allowing inclusion of time varying covariates that were not observed in all subjects at all event times.
Non-linear Modeling of Disease Trajectories
Another topic of interest is the non-linear modeling of disease trajectories. While atrophy rates are often modeled with a linear approach, there is evidence that rates increase around the time of conversion and potentially reach a floor effect at advanced progression [4,5]. Due to this non-linear behavior, atrophy rates may help predict time-to-disease-onset or allow selection of candidates into promising disease-modifying therapies at early pre-symptomatic stages.
Selected Publications (full list):
[1] Event Time Analysis of Longitudinal Neuroimage Data (NI 2014)
[2] Spatiotemporal Linear Mixed Effects Modeling... (NI 2013) [3] Statistical Analysis of Longitudinal Neuroimage Data with Linear Mixed Effects Models (NI 2012) [4] Longitudinal FreeSurfer for Reliable Imaging Biomarkers (MICCAI NIBAD 2012) [5] The dynamics of cortical and hippocampal atrophy in Alzheimer's disease (Arch.Neurol. 2011) |