Demographic characteristics of the subjects whose CSF samples were used in this study are given in . A summary of results from CSF measurements for 183 subjects in the control, AD, and PD groups is given in . Patients with PD had mild to moderate impairment with a median Hoehn and Yahr Staging Score of 2 (range, 1–3).15
Patients with AD had mild to moderate cognitive impairment with a median clinical dementia rating of 1 (range, 0.5–2).16
When considering each measurement separately, as expected from numerous previous studies, τ was increased and Aβ42
decreased when comparing control with AD but not control with PD groups (reviewed by Galasko17
). BDNF, IL-8, VDBP, apoAII, and apoE were significantly different between the control and both neurode-generative groups but not between neurodegenerative groups; of these, IL-8 and VDBP were increased in neurodegenerative diseases, whereas the remainder were decreased compared with control samples, again consistent with our previous quantitative proteomic studies.1,4
-microglobulin was significantly increased only between the control and PD groups, whereas apoAI and haptoglobin were not different among the 3 groups. Of the proteins identified in previous proteomic work as potential biomarkers for AD or PD, all but haptoglobin were confirmed in the present study. The AD group was older on average, and the control group had more women.
Demographic Characteristics of Groups
Cerebrospinal Measurements and RF Analysis*
These 12 variables (10 CSF proteins, age, and sex) were entered into an RF analysis.12–14
shows the 12 variables in decreasing order of relative contribution to discriminating among the 3 groups, as measured by mean decrease in accuracy. For each tree, the prediction accuracy for the cases not used to construct the tree was recorded, and this process repeated after permuting the values of each predictor variable. The mean decrease in accuracy was computed by averaging the differences between these 2 accuracy estimates for all trees and normalizing by the SE.
lists the RF classification rates for each group using all 12 variables or subsets of the top-ranked variables. The top 8 variables offered optimal classification.14
It is important to note that these top 8 excluded age, sex, apoAI (measured simply because it was contained in a prefabricated kit), and haptoglobin (the one unconfirmed protein). By using the top 8 variables for classification, all misclassified control subjects were incorrectly grouped as having AD, the 12 misclassified patients with AD were incorrectly grouped as 10 control subjects and 2 patients with PD, and the 2 misclassified patients with PD were incorrectly grouped as 1 control subject and 1 patient with AD.
Random Forest Classification*
In addition to the 183 control, AD, and PD CSF samples, we also analyzed available CSF samples that met our inclusion and exclusion criteria from 12 patients with MCI and 7 with FTD. By using all 12 variables or only variables 1 through 8, 9 of 12 patients with MCI were classified as control subjects and the other 3 were classified as having AD; none were classified as having PD. Again, by using all 12 or only the top 8 variables, 5 of 7 patients with FTD were classified as control subjects and the other 2 were classified as having AD; none were classified as having PD.