Regression Analysis and Prediction of Mini-Mental State Examination Score in Alzheimer's Disease Using Multi-granularity Whole-Brain Segmentations

摘要

We presented and evaluated three sparsity learning based regression models with application to the automated prediction of the Mini-Mental State Examination (MMSE) scores in Alzheimer's dis-ease(AD) using Tl-weight magnetic resonance images (MRIs) from 678 subjects, including 190 healthy control (HC) subjects, 331 mild cognitive impairment (MCI) subjects, and 157 AD subjects. The raw features were obtained from a validated multi-granularity whole-brain analysis pipeline, providing multi-level whole-brain segmentation volumes. We employed the ridge, lasso, and elastic-net as our regression algorithms, with the whole-brain volumes at each level being the independent variables and the MMSE score being the dependent variable. We used 10-fold cross-validation to evaluate the prediction performance and another 10-fold inner loop to estimate the optimal parameters in each model. According to our results, the combination of elastic-net and the second level of whole-brain segmentation volumes (a total of 137 volumes) worked the best compared to all other possible combinations. The work presented in this paper provides a potentially powerful and novel noninvasive biomarker for AD.