Group Comparisons via Voxel-based Analysis
The voxel-by-voxel analysis between two MCI sub-groups showed significant reduction of GM and WM in MCI-C compared to MCI-NC, at baseline. The results are shown in . Several regions of relatively reduced volumes of GM in MCI-C compared to MCI-NC are evident (red/yellow colors), including the hippocampus, amygdala, and entorhinal cortex, much of the temporal lobe GM and the insular cortex (especially the superior temporal gyrus), posterior cingulate and precuneous, and orbitofrontal cortex. Regions of increased periventricular WM tissue that appears gray, likely due to more pronounced Leukoaraiosis in MCI-C, relative to MCI-NC, were also evident and are shown in (blue colors), potentially indicating relatively more pronounced small-vessel disease in the former group. In agreement with this was the reduced WM in the periventricular frontal region which is shown in (red/yellow colors). WM was also relatively reduced in MCI-C in the perihippocampal temporal lobe region. As discussed in the Methods section, we also looked at group differences of the “beta” maps, i.e. of the rate of longitudinal change of brain tissue. The only findings were in GM RAVENS maps, and are displayed in . The most pronounced group difference in these beta maps was in the higher rate of periventricular Leukoaraiosis. Increased temporal cortical and anterior hippocampal atrophy was also measured.
Fig. 1 Maps of the t-statistics showing differences between MCI-C and MCI-NC. (a) and (b) show significantly more GM in MCI-NC relative to MCI-C (red/yellow), and areas of relatively increased periventricular WM tissue that appears gray in T1 images, likely (more ...)
The SPARE-AD scores of the AD and CN individuals were found to be in the expected range (mostly positive for the former and negative for the latter), therefore reconfirming the SPARE-AD as a marker of AD structural patterns. The average SPARE-AD scores were 0.69±0.50 and −0.80±0.43 for AD and CN groups, respectively, while they were 0.65±0.44 and 0.22±0.74 for MCI-C and MCI-NC. The histograms of the SPARE-AD scores of MCI-C and MCI-NC are shown in . Most MCI-C had positive scores, in fact their range of SPARE-AD values were indistinguishable from AD patients (t-test revealed no statistically significant differences), suggesting that significant atrophy has already occurred at MCI, for the people that are bound to convert to AD within the time-frame of this study. SPARE-AD scores of about 1/3 of the MCI-NC, however, were completely normal, indicating that a subgroup of MCI has normal brain structure, and that this subgroup doesn’t convert to AD in the time-frame of this study. However a majority of MCI-NC had sharply positive SPARE-AD scores, indicating significant atrophy similar to AD patients and to MCI-C. Voxel-based group comparison (images not shown) between the subgroup of MCI-NC with positive SPARE-AD scores and MCI-C showed a picture similar to the one of . Future follow-ups will determine whether these individuals convert or remain stable.
The histograms of baseline SPARE-AD scores for MCI-C (left) and MCI-NC (right).
The MCI-NC patients seemed to comprise 3 groups: people with low scores (well into the negative range centered around −1), people with scores around 0 (borderline cases), and people with high scores (well into the positive range centered around +1). We examined the longitudinal trajectories of these three subgroups separately, by obtaining SPARE-AD scores of all follow-ups. We used the following three ranges to subdivide the MCI-NC: SPARE-AD scores > 0.5 (part 1), SPARE-AD scores between −05 and 0.5 (part 2), and SPARE-AD scores below −0.5 (part 3). The average baseline and follow-up SPARE-AD scores are shown in , along with the trajectory of the average SPARE-AD score of the MCI-C group.
Trajectories of average SPARE-AD scores for MCI-C and sub-groups of MCI-NC. Scan #3 is on the average 12 months after Scan #1.
Integration of SPARE-AD and CSF in MCI
We only had a complete set of SPARE-AD and CSF biomarker measurements for a subset of 120 MCI patients with follow-up. jointly plots the CSF biomarkers data with SPARE-AD scores for MCI subjects. In order to evaluate the predictive value of combinations of imaging and CSF biomarkers, we used the weka software (http://www.cs.waikato.ac.nz/ml/weka/
), with input that features the SPARE-AD and various CSF biomarkers. A linear support vector machine (SVM)(Vapnik, 1998
) was used in a 5-fold cross-validation framework (20% of the data was left out, training was performed on the rest, and testing on the left-out patients; this procedure was repeated 5 times, with a different set of patients left out each time). The resulting classification accuracies and area under the curve (AUC) measures are shown in . For the SPARE-AD, we also repeated this experiment on all 239 patients (recall that we had CSF and SPARE-AD values for only a subset of 120 patients, whereas we had SPARE-AD values for all 239 patients).
Scatterplots of SPARE-AD against CSF markers Aβ42 and t-tau. The two oblique lines represent the SVM classifiers achieving two different levels of sensitivity, i.e. correct classification of MCI-C: ~82% (dotted line) and ~92% (solid line).
Classification of MCI-C vs. MCI-NC
In order to graphically show the joint value of SPARE-AD and CSF biomarkers, in Fig. 7 we plot the SPARE-AD against CSF tau and Aβ42. In the same plots we include two SVM classifiers determined for different levels of sensitivity (by varying the cost of misclassification of MCI-C relative to MCI-NC, one can create many such SVM separating lines, each of which corresponds to a different point on the ROC).