Multivariate prediction of Hippocampal atrophy in Alzheimer's disease

Background: Hippocampal atrophy (HA) is one of the biomarkers for Alzheimer's disease (AD). Objective: To identify the best biomarkers and develop models for prediction of HA over 24 months using baseline data. Methods: The study included healthy elderly controls, subjects with mild cognitive impairment, and subjects with AD, obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI 1) and the Australian Imaging Biomarkers and Lifestyle Flagship Study of Ageing (AIBL) databases. Predictor variables included cognitive and neuropsychological tests, amyloid-, tau, and p-Tau from cerebrospinal fluid samples, apolipoprotein E, and features extracted from magnetic resonance images (MRI). Least-mean-squares regression with elastic net regularization and least absolute deviation regression models were tested using cross-validation in ADNI 1. The generalizability of the models including only MRI features was evaluated by training the models with ADNI 1 and testing them with AIBL. The models including the full set of variables were not evaluated with AIBL because not all needed variables were available in it. Results: The models including the full set of variables performed better than the models including only MRI features (root-mean-square error (RMSE) 1.76-1.82 versus 1.93-2.08). The MRI-only models performed well when applied to the independent validation cohort (RMSE 1.66-1.71). In the prediction of dichotomized HA (fast versus slow), the models achieved a reasonable prediction accuracy (0.79-0.87). Conclusions: These models can potentially help identifying subjects predicted to have a faster HA rate. This can help in selection of suitable patients into clinical trials testing disease-modifying drugs for AD.