This study investigated the diagnostic value of imaging biomarkers, obtained by integrating structural and functional images via high-dimensional pattern classification, for discrimination of cognitively normal individuals from individuals with mild cognitive impairment. Complex spatial patterns of brain atrophy and blood flow were identified and found to have very high diagnostic accuracy that reached 100%, which is 6% higher than the maximum classification rate achieved from MRI only. A more conservative estimate of generalization performance, obtained via a procedure called bagging in conjunction with leave-one-out cross-validation, was 90%, with an area under the curve equal to 0.978, which was also about 3% higher than the respective accuracy obtained via MRI alone, and much higher than accuracy achieved via PET alone. These results indicate an excellent diagnostic value of the SFBS, especially in view of the relatively limited sample size available in this study.
A strength of our prospectively assessed cohort is that we were able to study MCI at a relatively mild stage of cognitive decline and presumably of underlying pathology. The early identification of MCI in our sample is an important distinction between the current study and other studies that involve patients reporting to the clinic with memory complaints.
Evaluation of the spatial patterns of brain atrophy indicated quite widespread, yet often subtle, reduction of mainly GM, but also WM volumes in a variety of brain regions, including the hippocampus, inferior temporal cortical GM and WM, insular and orbitofrontal GM, posterior cingulate, and precuneus. Two regions of reduced blood flow were also measured, namely the posterior cingulate and a fairly extended area in the left parietal and parieto-temporal cortex. Most of these regions are known to be affected by AD (Ardekani et al., 2007
; Chetelat et al., 2002
; Convit et al., 1997
; Dickerson et al., 2001
; Firbank et al., 2007
; Fox and Schott, 2004
; Fox et al., 1996b
; Frisoni et al., 2007
; Ishii et al., 2005a
; Jack et al., 1999
; Karas et al., 2007
; Killiany et al., 2000
; Krasuski et al., 1998
; Matsuda et al., 2002
; Pantano et al., 1999
; Pennanen et al., 2005b
; Thompson et al., 2007
); however, it is their combination via high-dimensional pattern classification that leads to high diagnostic accuracy on an individual basis.
The pattern of reduced blood flow measured in our study was asymmetric, and in agreement with the findings of (Minoshima et al., 1994b
; Reiman et al., 1996b
; Scarmeas et al., 2004
) who used PET 15
O imaging in an AD sample. A variety of PET studies in AD have demonstrated asymmetrically reduced blood flow or metabolic activity in AD patients (Foster et al., 1983
; Ishii et al., 2005b
; Kawachi et al., 2006
; Koss et al., 1985
; Martin et al., 1986
) and in elderly individuals with cognitive decline (Hunt et al., 2007
); however the biological underpinnings of these asymmetries are not quite known. Notably, these asymmetries are often reversed, implying that different individuals might be affected by, or compensate for, AD pathology differently. The pattern of reduced blood flow in association with MCI indicated decreases in posterior cingulate rCBF. Decreased neural activity in the posterior cingulate has been described in AD (e.g.(Minoshima et al., 1994a
) and in unaffected individuals who carry the Apolipoprotein E e4 risk factor for AD (Minoshima et al., 1994a
; Reiman et al., 1996a
Our findings suggest relatively symmetric brain atrophy, in contrast to some previous findings using VBM (Karas et al., 2004
; Thompson et al., 2001
). Although methodological differences in image analysis could also account for the differences of our findings and the ones in (Karas et al., 2004
), patient selection is also likely to be a factor. In particular, the study in (Karas et al., 2004
) was retrospective, whereas our study was prospective. Retrospective studies can be confounded by the fact that a relatively larger degree of right-hemisphere atrophy, compared to left, is likely to be tolerated before the patient reports to the clinic, since deficits in verbal memory are more likely to trigger a visit to the clinic relatively earlier. Prospective studies, however, tend to be more robust to such confounds. Moreover, our findings are consistent with histopathological studies of AD, which have found relatively symmetric patterns of brain atrophy (Braak et al., 1998
The results of the high-dimensional pattern classification are very encouraging, since they indicate that sufficient sensitivity and specificity can be achieved for these tools to have diagnostic value in the clinic. Although 100% cross-validated classification accuracy was achieved for certain parameter settings, a more realistic estimate of the generalization accuracy of this approach is 90%, as the results of indicated, which is based on automatic estimation of a single number of brain clusters from the training set, instead of evaluating performance over a range of values for the number of clusters. This small discrepancy is mainly due to the small sample size available to us in this study, which typically leads to the jittery performance curves shown in due to under-training of the classifier. Based on the curve of , we anticipate that more extensive training on larger sets of data will ultimately allow us to achieve stable classification performance close to 100%.
The set of ODC () tended to be in agreement with the voxel-based analysis of the MRI and PET images (). However, the ODC approach yielded a more parsimonious number of brain clusters. This is to be expected, since the classification methodology inherently seeks the minimal number of brain clusters necessary for classification. Parsimonious models are of great value in interpreting the data, since they allow us to focus on the patterns of structural and functional changes that are most distinctive of MCI, thereby achieving a higher level of robustness. Interestingly, the spatial pattern of fits with the known patterns of early AD pathology, spanning mainly the temporal lobe, the orbitofrontal cortex, and the posterior cingulate.
Although the majority of the neuroimaging literature of AD has focused on measuring regions that are most affected by the disease, such as the medial temporal lobe and posterior cingulate, our study suggests that patterns of structural and functional imaging characteristics must be evaluated to increase the accuracy of tools sufficient for diagnosis on an individual basis. These spatial patterns are complex and include many regions that are not known in advance. The methodology presented herein is built around this concept, namely that the set of brain measurements that optimally differentiates between the two groups cannot be known a priori, but is determined from the data. However, since leave-one-out cross validation was used, the optimal set of clusters was always tested on new, previously unseen scans. This is fundamentally important, and guards against over-fitting the data, a problem analogous to that of multiple comparisons in VBM analyses.
Important for achieving high classification rate was not only the combination of many brain regions, but also the combination of PET and MRI. MRI alone had far superior classification accuracy than PET when each modality was used alone. However, the combination of the two provided the best results, suggesting that PET offers additional and somewhat complementary information to MRI. The combination of the two modalities is important for an additional reason, namely that changes in blood flow can be attributed, in part, to brain atrophy. Although regional measurements of brain atrophy could be used to normalize the PET signal, the fully multivariate approach followed herein offers a more general and comprehensive way of examining the data.
While the current study aims to develop an imaging-based diagnostic tool, it only takes one step in that direction. Our study examines the ability to distinguish between amnestic MCI and cognitively normal groups but does not include all possible types of dementia. For example, we do not test the ability of this classifier to distinguish between early stages of frontotemporal dementia and early stages of AD. However, this type of multi-class classification problem is generally a relatively straightforward extension of the two-group problem (Duda et al., 2001
). In particular, multi-class classification is often achieved by applying a number of pair-wise two-class classifiers, and then combining the results using voting. Multi-class support vector machine classifiers are also available.
In summary, this study combined structural MRI and blood flow PET images in a high-dimensional pattern classification framework, which achieved up to 100% accuracy in classifying individual scans of patients with relatively mild MCI in a prospective longitudinal study of aging. The results indicate that high-dimensional classification, partly built upon voxel-based multivariate analysis of the integration of structural and functional images, has the potential to serve as an early diagnostic tool for AD on an individual patient basis. Future studies on larger samples, as well as on healthy elderly individuals with cognitive decline, will further test whether these structural and physiological alterations seen in this MCI cohort are replicated robustly and perhaps at even earlier disease stages. These diagnostic tests have a great potential in complementing standard neurological examinations, especially in evaluating disease progression reliably and quantitatively.