Predictability Landscape of Alzheimer's Disease in Longitudinal Data

Abstract

In the absence of a cure for Alzheimer’s disease (AD), research efforts have been focusedon the early detection and prediction of individuals who may develop the disease based on demographics, cognitive assessments, genetics, and imaging biomarkers. Heterogeneity and the slow progression of the disease are two factors that often challenge the pathology of the disease. In longitudinal studies such as the Alzheimer’s Disease Neuroimaging Initiative (ADNI), missing data can arise due to reasons such as study design, lost to follow-up or subject non-compliance. We wanted to determine the predictability of risk of conversion in subjects over a five-year time horizon. To this end, we separated subjects from ADNI into two groups, namely those who were cognitively normal (CN) at their baseline visit, and those who had mild cognitive impairment (MCI) at their baseline visit. Missing data were imputed via Multiple Imputations by Chained Equations (MICE). We compared the predictability of conversion risk when using Elastic Net versus Siamese networks; when one versus two clinical visits were used as inputs to the model; when different categories of variables were included, and finally, when different sample sizes were used to train the model. Based on the area under the receiver operating characteristic (AUROC) curve, the linear models were often sufficient in that it either performed similarly or outperformed the non-linear model; there were moderate advantages to using all demographic, cognitive, genetic, and imaging variables as features in the models, particularly for the CN group; using two clinical visits improved predictive performance, particularly in the earlier timepoints, and that predictive performance increased as the number of subjects used during training increased.