We studied 25 patients with PCA (age [mean ± SD] 62.2 ± 7.2 years, 56% male, age at symptom onset [AAO] 58.2 ± 6.6 years), 15 patients with LPA (age 60.4 ± 6.1 years, 67% male, AAO 56.3 ± 5.3 years), 14 patients with tAD (age 60.8 ± 5.2 years, 43% male, AAO 56.3 ± 3.9 years), and 30 healthy control subjects (age 63.9 ± 6.7 years, 50% male). Of the patients, 6 with PCA (24%), 9 with LPA (60%), and 5 with tAD (36%) had postmortem confirmed AD pathology. Groups did not differ significantly in age, gender, or AAO, although control subjects were slightly older than subjects with tAD (p
= 0.05). A subset of 29 genotyped patients showed no significant disease by APOE
status association (p
> 0.17). All patients fulfilled the respective clinical criteria.2,4,5
Standard protocol approvals, registrations, and patient consents.
All clinically affected subjects had attended the Specialist Cognitive Disorders Clinic at the National Hospital for Neurology and Neurosurgery, London, UK. Informed consent was obtained from all subjects, and the study had local ethics committee approval.
T1-weighted volumetric MRI scans (124 contiguous 1.5-mm coronal slices) were acquired using an inversion recovery spoiled gradient recalled sequence on 3 identical 1.5-T General Electric Signa units (no significant group by scanner association, p
> 0.25). FreeSurfer 4.5.06
was used to extract and align the cortical surfaces, resulting in thickness measurements at approximately 300,000 points (vertices) on an average surface, which were smoothed to 20-mm full-width at half-maximum. Vertices in FreeSurfer's medial wall region were excluded from subsequent analysis. Two modifications to standard FreeSurfer processing were undertaken: a locally generated brain mask was used, and FreeSurfer ventricular segmentations were added to its white matter mask to improve segmentation.
Regional cortical thickness variations were assessed with a vertex-wise general linear model using SurfStat (http://www.nitrc.org/projects/surfstat
). Cortical thickness was modeled as a function of group, controlling for age, gender, and scanner. Two-tailed t
contrasts were thresholded to control familywise error at p
< 0.05. Intersection maps were produced, highlighting common atrophy in the patient groups (the conjunction of the 3 syndrome vs control contrasts).
Multivariate machine learning.
Considering each subject's cortical thickness measurements simultaneously at every vertex yields high-dimensional multivariate patterns. To visualize the distribution of subjects based on these cortical thickness profiles, we use multidimensional scaling (MDS) and a sophisticated nonlinear technique called t
-distributed stochastic neighbor embedding (t-SNE).7
Both MDS and t-SNE are data-driven or unsupervised methods (i.e., trained without group information) that preserve distances between pairs of subjects to reveal neighborhood relationships and clusters. Here, the distance between 2 subjects is defined as 1 − r
, where r
is the Pearson correlation between the subjects' cortical thickness profiles − vectors of the ~300,000 thickness values (after adjustment for age, gender, and scanner using the univariate general linear model).
MDS minimizes discrepancies of intersubject distances in a 2-dimensional representation with respect to the intersubject distances in the original high-dimensional space. Although intuitively appealing, this criterion can be dominated by larger distances, neglecting the local neighborhood structure that is important for clustering. Instead of directly matching distances, t-SNE matches the high- and low-dimensional distributions of the data probabilistically: the probability of 2 points relates to their proximity such that the criterion appropriately balances shorter and longer distances. We report both MDS and t-SNE because the latter is not guaranteed to converge to the global optimum of the more complicated probabilistic criterion; broadly similar visualizations using each technique would suggest that an acceptable optimum had been found.
After the visualizations, we quantify group separation by using a “kernel” matrix derived from the above intersubject distances in a supervised machine learning classifier (support vector machine [SVM]).8,9
One classifier was trained to separate control subjects from patients (pooling all patient groups), with resultant SVM scores subsequently labeled by group. Three separate classifiers were then trained specifically to discriminate each pair of patient groups.
We used the SVM approach because of its suitability for very high-dimensional data. With high dimensionality, a particular way of separating groups determined with training data might not successfully separate unseen test data; SVMs address this by finding the dimension along which the groups are separated with the widest margin, as this typically generalizes well to new data. Here we used a soft-margin SVM, which allows some of the training data to be misclassified to obtain an even wider margin. To optimize the tradeoff between soft-margin width and misclassifications without biasing the estimated performance, a second (inner) cross-validation is needed on the training data8
; because of the few subjects, we used a nested leave-one-out procedure to accomplish this efficiently.9
presents regional differences in cortical thickness between the patient groups and control subjects. In PCA, cortical thickness was most significantly reduced in posterior regions including bilateral parietal and occipital areas, as well as the posterior cingulate gyrus and precuneus (A). In contrast, lower cortical thickness in subjects with LPA compared with control subjects occurred predominantly in left hemisphere temporal and frontal lobe regions (B). The most significant cortical thickness reductions in the tAD group were in bilateral temporal and posterior parietal lobe regions, as well as in the precuneus and posterior cingulate (C). The greatest overlap for the 3 patient groups (D) was found in the left hemisphere, including parietal, inferior temporal, and middle frontal lobe regions, as well as precuneus, fusiform, and posterior cingulate. Right hemisphere overlap was also found, in medial and superior temporal lobes, precuneus, fusiform, and posterior cingulate.
Regional differences in cortical thickness between control subjects and subjects with (A) posterior cortical atrophy (PCA), (B) logopenic progressive aphasia (LPA), and (C) typical amnestic Alzheimer disease (tAD)
The visualizations in show a notable tendency for control subjects to separate naturally from patients, but for patient groups to be more interspersed. Results from MDS and t-SNE are qualitatively similar, but the more sophisticated method achieves better clustering. Among patient groups, the subjects with PCA and LPA are best separated and the subjects with tAD are distributed along a spectrum between these extremes (with several straying into the territory of the control subjects).
Two-dimensional visualizations of the distribution of subjects in terms of their high-dimensional cortical thickness profiles
Quantitative results from SVM classification analyses were consistent with the visualizations. Control subjects separated well from all patients, although no natural separation emerged among patient groups (A). However, separate classifiers trained directly to distinguish pairs of patient groups achieved significant discrimination for every pair (B), again best distinguishing PCA and LPA.
Supervised support vector machine (SVM) classification of control subjects and patients and of AD variant groups