Progressive Disintegration of Brain Networking from Normal Aging to Alzheimer Disease: Analysis of Independent Components of ¹⁸F-FDG PET Data

DISCUSSION

ICA identified spatially distinct clusters of voxels with meaningful pathophysiologic value in each group, allowing us to uncover the spatial distribution of the independent components in each cognitive state and to assess the accuracy of ICA in discriminating between groups. The hypothesis of a progressive disintegration of brain connectivity during the course of the disease was proven in terms of the generation of new local circuits that progressively replaced the brain-as-a-whole mean field detected in a previous study (10).

With increasing disease severity, the shared portion of brain metabolism (λ₁) loses relative importance with respect to local distinct neural circuits, which increase in both number and extent. Degradation of functional connectivity as AD progresses is a loss of integration: the brain becomes increasingly segmented into independent metabolic areas.

Disintegration of the intrinsic networks that characterize normal aging was progressively more severe with the transition from nonconverting MCI to AD, with the extent of the independent components being inversely proportional to time from conversion. Neurodegenerative processes affect modular networks rather than isolated regions (16). Long-distance interregional metabolic correlations are impaired by the anatomo-functional progression of neuronal and white matter fiber degeneration (17), causing local compensatory networks to take over and increasing the anatomic and functional segregation of brain processes.

The progressive disintegration of the giant component that accounts for the main whole-brain connectivity present in normal aging was disclosed by ICA, which identified in all groups a specific set of ¹⁸F-FDG PET–derived components whose pattern agreed with the metabolic decay from normal aging to MCI to AD.

Independent components, including the primary visual cortex and cerebellum, are present in all groups. Therefore, they are not implicated in the functional pattern of cognitive decline. This confirms well-known findings and supports the suitability of choosing the cerebellum as a normalizing factor in radionuclide imaging of brain function in AD (18).

The medial frontal gyrus and dorsolateral frontal cortex (representing the anterior portion of the default-mode network), as well as the supplementary motor cortex, are incorporated in independent components peculiar to normal aging. The anterior portion of the default-mode network is thus disintegrated in all patient groups, including MCI nonconverters. The disappearance of the anterior default-mode network in MCI nonconverters is intriguing and may point to the involvement of anterior frontal regions in a variety of conditions—ranging from late-life depression to cerebrovascular disease—that affect this network. However, Di et al. (19) reported that the absence of the anterior part of the default-mode network in PET investigations but not in functional MRI investigations was due to differences in the temporal scales of PET and MRI.

On the other hand, among the main components of the posterior portion of the default-mode network, the posterior cingulate cortex and the inferior parietal lobule belong to the same independent component in normal aging and nonconverting MCI. This independent component is disintegrated in MCI converters and in AD patients, in whom the posterior cingulate cortex stands as an independent component alone or in association with the precuneus. This agrees with the disappearance of the default-mode network in AD and with the knowledge that metabolic changes in the posterior cingulate cortex and precuneus are indeed markers of AD pathology (20).

In line with this interpretation, the dorsolateral frontal cortex, either alone or in association with the superior parietal lobule or temporal cortex, is embedded in independent components found in all MCI patients, either converting or not. Thus, the dorsolateral frontal cortex may be considered a correlate of cognitive impairment not specifically linked to AD pathology.

Surprisingly, the temporal cortex close to the sylvian fissure also was embedded in an independent component in all patients with cognitive deficit, including MCI nonconverters, and hence could be considered a correlate of memory deficit independent of AD pathology, possibly resulting from an aspecific temporal lobe atrophy shared by different conditions (21).

The temporal cortex and inferior parietal lobule in the right hemisphere are aggregated in an independent component in all groups progressing to AD and in AD itself and thus are more related to the specific AD signature, independent of the severity of the disease.

Three independent components characterize the late-MCI stage, namely those including the lateral occipital cortex bilaterally, the left superior parietal cortex, the temporal cortices, the precuneus, and the dorsolateral frontal cortex. The spreading of functional deficit to these association cortices follows the progressive-disconnection hypothesis as proposed by the classic model of Braak and Braak (22). Furthermore, the posterior cingulate cortex (either with or without the precuneus), ventrolateral frontal cortex, superior parietal lobe, and medial temporal lobe are the independent components specific to the late stage of the disease (MCI converters within 2 y and AD).

The final metabolic signature of the definitive conversion to AD are two independent components including the left temporal cortex. However, the full-AD stage includes several independent components: right and left precuneus/posterior cingulate cortex, medial temporal lobe, ventrolateral frontal cortex, inferior parietal lobule, temporal sylvian cortex, and right temporal and occipital lateral cortex.

Beyond the pathophysiologic significance of these independent components in AD, two further observations have to be made. The first is that this study can be considered prospective because all 3 groups of MCI patients were investigated by ¹⁸F-FDG PET at a time when the neuropsychological assessments showed similar levels of cognitive deficit (baseline), with an expected decline in memory performance from MCI nonconversion to MCI conversion within 2 y. Thus, the metabolic pattern at baseline was predictive of conversion to AD and was associated with the conversion time. The inclusion of both converting and nonconverting MCI patients adds clinical value to the analysis since these intermediate classes are the most relevant targets for a proper and timely diagnosis and for subsequent management. In the present study, the voxel-based analysis found a highly significant metabolic difference in the posterior cingulate and temporal and parietal cortices between these two groups (Fig. 1). Furthermore, this difference was confirmed by SVM analysis, which found that analogous regions separated nonconverting MCI from all AD with an accuracy of 83.5% (Table 3). The independent component aggregating the posterior cingulate cortex and the inferior parietal lobule, representing the posterior default-mode network, is considered to be a hub of functional connectivity sensitive to synaptic disconnection (23) and was found in an MRI ICA study to best predict conversion from stable MCI to progressing MCI (24).

Another observation that has to be made is that ¹⁸F-FDG PET uptake in the independent components of AD patients showed fairly good power in discriminating normal aging and nonconverting MCI from AD. This observation confirms the utility of such data-driven methodology to uncover correlations that have pathophysiologic meaning (Table 3). Several studies have implemented automated image-based classification methods to differentiate AD and MCI patients from controls and have found such methods to have a statistical accuracy of 90% in discriminating AD patients from controls (25–27). When MCI patients who later converted to AD were investigated by unimodal biomarkers, the reported ability to discriminate from controls ranged from 80% (28) to 91% (29). In the present study, the independent components derived by the AD dataset showed an accuracy of 90% in discriminating controls from the entire cohort with AD pathology. Moreover, the weak discriminatory power (area under the curve, 0.65) between controls and MCI nonconverters indicates a strong similarity between these two groups, providing functional support for the lack of conversion observed clinically in the latter group.

When the analyses were modeled on the two extreme classes (normal aging and AD), the patients of the intermediate groups were progressively assigned to the AD class according to disease severity. The same held true when the analyses were modeled on nonconverting MCI and AD (Table 3), confirming the reliability of the present approach.

Our results are consistent with an MRI study performed by Pereira et al. (30) that found in two large, multicenter cohorts a progressive degradation in the connectivity of brain functional networks in patient groups having the same characteristics as in our study, including MCI converters followed up for 1 and 3 y. The most striking similarity with our study was the finding of a progressive increase in modularity with severity of the disease, suggesting a degradation of whole-brain networks into segregated components. This consistency is even more remarkable if we consider that Pereira et al. used a different data analysis strategy (graph analysis) and, more important, a different biologic observable (MRI).

Some of the independent components found in the present study could be superimposed with those found by Laforce et al. (31), who performed one of the few investigations implementing ICA in ¹⁸F-FDG PET. In that study, independent components including the frontotemporal cortex, right and left occipitoparietal cortex, posterior cingulate–precuneus–superior parietal cortex, and right and left cerebellum were identified in 54 AD patients. In contrast, only one study attempting to use independent components to distinguish between controls and AD patients (who had been recruited within the Alzheimer Disease Neuroimaging Initiative) could discriminate the two groups with an accuracy of 91% (32).

Unlike multicenter studies, the diagnostic procedures in this investigation were uniform since they were performed by the same clinical group. Additionally, all ¹⁸F-FDG PET scans were performed using the same camera, minimizing the likelihood that inhomogeneous subject samples and camera acquisitions could have increased the variability of the data or reduced the robustness of the results. Another strength of this study was the long follow-up time (minimum, 5 y) available for MCI nonconverters, making it unlikely that our group included any late converters.

The progressive increase in the number of independent components with disease severity is the functional consequence of the loss of relevance of the brain-as-a-whole component that we observed in our previous work (10). The degradation of the average correlation field, encompassing the entire brain, resulted in creation of local (and largely autonomous) networks. The progressive loss of connectivity of brain metabolism caused by AD can thus be interpreted not as a loss of local correlation inside a functional network but as a loss of coordination among networks themselves. Furthermore, in neurodegenerative disease, ICA-derived patterns not only can be of high diagnostic utility but also can give mechanistic insight into the etiology of the various disorders (13).

From a methodologic point of view, it is worth noting that the present results were achieved thanks to the complementary properties of PCA and ICA. PCA emphasizes the correlation between different variables (brain regions, bins of a spectrum) in time or space. A latent physical process is considered a global-order parameter (in physical terms, a force) imposing a correlation on otherwise independent pieces of information. The contemporaneous presence of independent forces shaping the observed system allows us to consider each variable as a weighted sum of the same hidden-order parameters (principal components) acting on the different parts of the system. For ICA, the signature of an elemental signal is its multimodality. For PCA, the strength of each elemental signal is the degree of correlation it imposes on the whole system. This difference makes ICA dissect the system into separate local correlation circuits (networks), whereas PCA tries to recollect the complexity of the system into the minimum number of orthogonal components acting on all the elements (even though with different strengths). This implies that although ICA is, by definition, masked to global mean field effects (size components) shared by the entire system, PCA is by far less sensitive in catching the presence of local networks. Here, we have shown the added value of comparing the ICA and PCA pictures to disclose the connectivity dynamics of brain metabolism.