Premium
P3‐241: ONE‐YEAR OBSERVATION IN FDG‐PET IMPROVES PREDICTION OF PROGRESSIVE MCI TO ALZHEIMER'S DISEASE WITHIN 5 YEARS: LONGITUDINAL RESULTS FROM SEAD‐J
Author(s) -
Ishii Kazunari,
Takahashi Ryuichi,
Fujiwara Ken,
Kato Takashi,
Ito Kengo,
Washimi Yukihiko
Publication year - 2014
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1016/j.jalz.2014.05.1332
Subject(s) - medicine , posterior cingulate , precuneus , area under the curve , alzheimer's disease , cardiology , nuclear medicine , disease , cognition , psychiatry
Background: Recently, we introduced metrics for Alzheimer’s disease (AD) risk assessment called AD pattern similarity (AD-PS) scores. These are class-conditional probabilities generated using regularized logistic regression (RLR) estimated in a high-dimensional voxel space. Here we report exploratory analyses designed to determine if the AD-PS scores sensitivity would improve when (a) we use measures of cortical thickness (CT) and/or regional volumetric (Vols) measures as predictors instead of voxels, or (b) RLR is replaced by Random Forest (RF), a highly nonlinear classifier that also generates conditional probabilities. Methods: Baseline data from 359 participants of the AD Neuroimaging Initiative study were analyzed: 188 were cognitively normal (CN) and 171 had AD. FreeSurfer (FS) measures of CT and volumes and gray matter images were used to create four different sets of predictors (CT, Vols, combined CT and Vols (CT + Vols) and voxels) for both classifiers. Thus, we compared eight different versions of these classifiers when discriminating CN from AD. We varied the sample size from 20 to 280. Mean classification performance was estimated using 30 random data partitions in training and testing datasets. GLMNET and randomForest libraries versions of RLR and RF were used in our implementation. Results: For larger samples, RF and RLR relative classification accuracy was similar for the sets of predictors based on FS measures. RF clearly performed worse for the high-dimensional voxel space. For larger sample sizes, RLR estimated in the voxel space and both RLR and RF estimated in the combined CT+Vols sets of predictors, produced similar classification accuracy and were the best three performers. Both methods trained on the CT+Vols predictors often performed better than their counterparts trained independently on CT or Vols. Conclusions: When distinguishing CN from AD, given a sufficiently large sample size, we did not find evidence that using sets of predictors of lower dimensionality instead of voxel space improved the classification performance of RLR. RF non-linearity also did not yield advantages. These analyses suggest that RF or FS measures will not improve performance of the AD-PS scores. However, they do provide useful information about the impact of dimension and sample size on these two machine learning methods.