The memory task we used in the current study is a modified version of the task we used previously (Miller et al., 2008). The Miller et al. paper utilized a pure event-related design, whereas the current paper uses a shorter mixed-block and event-related design that can be performed by more impaired subjects. So yes, one possibility for the lack of correlation with PIB and task performance is that the current task is not as difficult as the one in Miller et al., 2008. That one had 232 face-name pairs, whereas the Neuron task has only 84 novel face-name pairs. So we also may have less range of performance on the basis of task difficulty.

Several recent reports have also found no evidence of relationship between PIB and other memory measures among normal subjects (Aizenstein et al., 2008; Jack et al., 2008; Jack et al., 2009), so I am not too surprised that we didn't see a strong relationship,...  Read more

The memory task we used in the current study is a modified version of the task we used previously (Miller et al., 2008). The Miller et al. paper utilized a pure event-related design, whereas the current paper uses a shorter mixed-block and event-related design that can be performed by more impaired subjects. So yes, one possibility for the lack of correlation with PIB and task performance is that the current task is not as difficult as the one in Miller et al., 2008. That one had 232 face-name pairs, whereas the Neuron task has only 84 novel face-name pairs. So we also may have less range of performance on the basis of task difficulty.

Several recent reports have also found no evidence of relationship between PIB and other memory measures among normal subjects (Aizenstein et al., 2008; Jack et al., 2008; Jack et al., 2009), so I am not too surprised that we didn't see a strong relationship, either. There was a trend (p value of about .2). Also, we restricted this study sample to subjects without any objective memory impairment (within 1.5 SD), so we may have truncated the range even among a "generally normal" population.

We think cognitive reserve may have also played a role in allowing these subjects to perform well even with large amounts of amyloid deposition (see Roe et al., 2008). We are now conducting analyses to determine if cognitive reserve directly influences fMRI activity in the presence of amyloid.

Finally, we controlled for performance in this paper. That is, we only looked at successful encoding (High Confidence hits), because of our findings in Miller et al., 2008. We wanted to see if, controlling for performance, we still saw an effect of PIB on default network activity. If we had looked at all encoding trials, I suspect we would have again seen evidence of the relationship between deactivation and overall task performance.

Controlling for performance, we still found that the PIB+ subjects showed failure of deactivation even when they did encode the face-name pair successfully. Furthermore, at least in some subjects, we saw that successful encoding required increased hippocampal activity, which we speculate is compensatory, in the setting of both amyloid deposition and failure of default activity. I hypothesize that we will see evidence of memory decline in those subjects with high PIB retention and impaired default activity, but at least overall, at the time of this experiment, they were still performing pretty well. So at the moment, I would take our findings as evidence of early amyloid-related alterations that may convey vulnerability to eventual decline. It was striking how similar the pattern of paradoxical default network activation seen in the PIB+ adults was to previous reports in MCI and AD (Lustig et al., 2003; Petrella et al., 2007; Pihlajamaki et al., 2009), so I do think this is evidence that the memory systems are not working normally.

View all comments by Reisa Sperling

This article makes the very interesting observation that the “default-mode” network differs in healthy elders and patients with Alzheimer’s disease. It contributes to our understanding of the commonly observed changes in resting cerebral metabolism in Alzheimer’s disease. It also begins to point the way to an imaging approach that can reliably distinguish Alzheimer’s patients from normal, healthy elders.

The article is a strong demonstration of the utility of multivariate approaches such as ICA, PLS, or SSM that attempt to isolate covariance patterns. In contrast to more standard voxel-wise approaches, these multivariate approaches directly measure the relationship between functional changes in various brain areas. They often have increased sensitivity for detecting subtle perturbations that may be associated with disease processes such as early Alzheimer’s disease. Thus, in a study that uses a relatively small number of subjects, this technique provides relatively good separation between network expression (as measured by goodness of fit) in Alzheimer’s patients and...  Read more

This article makes the very interesting observation that the “default-mode” network differs in healthy elders and patients with Alzheimer’s disease. It contributes to our understanding of the commonly observed changes in resting cerebral metabolism in Alzheimer’s disease. It also begins to point the way to an imaging approach that can reliably distinguish Alzheimer’s patients from normal, healthy elders.

The article is a strong demonstration of the utility of multivariate approaches such as ICA, PLS, or SSM that attempt to isolate covariance patterns. In contrast to more standard voxel-wise approaches, these multivariate approaches directly measure the relationship between functional changes in various brain areas. They often have increased sensitivity for detecting subtle perturbations that may be associated with disease processes such as early Alzheimer’s disease. Thus, in a study that uses a relatively small number of subjects, this technique provides relatively good separation between network expression (as measured by goodness of fit) in Alzheimer’s patients and controls.

Another very strong feature of multivariate, covariance-based approaches is the ability to forward-apply results from one data set to another. In the current paper, the authors sought to find in the data of another group a default-mode network that they had identified in their own data. This ability to test network results across different data sets provides a very powerful method for validating results and provides an important empirical check of the untestable assumptions that usually enter the assessment of p-values in neuroimaging research.

Multivariate approaches are excellent for looking at inter-individual variability in network expression. Network expression, i.e., the degree to which a subject manifests the network, can typically be summarized with a single parameter (in the present case goodness of fit). Networks can be identified prior to the utilization of other variables of interest in an algorithmic, objective manner. The expression of these networks can then be tested for a relationship with those other variables. Because the information of these variables was not used in the identification of the networks in the first place, any association with a network gives additional credence to its validity. In this case, the authors related degree of network expression to disease state. Others have related network expression to various clinical characteristics or test performance measures that are derived independently from the network analysis. For example, in this case, degree of network expression might be associated with dementia severity. Our group has related differential network expression to differential performance on the activation task.

Thus, in principle, the approach taken here has great promise. There are a few practical limitations to the adaptation of the specific implementation utilized in the current paper to the problem of clinical diagnosis.

First, the approach utilized here relies on within-subject variability. The ICA analysis was performed on each subject’s motor task activation data, as collected over time (i.e., each subject’s 4D image). It would be useful to know whether a similar approach could be taken with more static images, such as PET or SPECT, which are the functional images that are most likely to be ubiquitously available at medical centers. The current approach would require each subject to undergo an fMRI study with a specific cognitive activation. This is a more labor-intensive and rarified procedure.

It is also important to keep in mind that the authors have simply demonstrated that the default-mode network identified in young subjects is replicated in elders, while the “fit” of activation data from AD patients to that seen in young subjects is poorer. From a practical point of view, what the analysis has done is taken 24 to 31 ICA components for each subject and attempted to match them to the default-mode network derived from young subjects. The analysis is in a difficult position of finding the one component for each subject that best matches the default-mode network and then calculating the degree to which it deviates from that network. This raises some important questions. Is there a point where the network is so much altered in Alzheimer’s patients that it cannot be reliably identified by the template, and there is no real default-mode network present anymore? Also, how specific is this deterioration in goodness of fit to Alzheimer’s disease? A gradual disappearance of the default-mode network might also be common to neurodegenerative disorders other than Alzheimer's. Thus, for the direct identification of patients with Alzheimer’s disease, it might be useful in the future to characterize a similar default-mode network in Alzheimer’s disease and then use similar methods to determine whether potential Alzheimer’s patients do or do not match that Alzheimer’s disease network.

All of these issues can easily be addressed in further research. These observations simply strengthen the point that this article brings to the fore the unique power of multivariate approaches to address these difficult problems of characterizing network change across diagnostic states.

View all comments by Yaakov Stern

As is usually the case with work from Buckner, Sperling, and Johnson, this is very interesting and innovative work. It’s similar in some respects to the 2005 J. Neurosci paper comparing the topography of the default mode network to amyloid deposition. Indeed, many of the hubs lie in this region, and hub activity may be at the root of default mode activity and, in turn, may exacerbate Aβ deposition. It’s not completely clear to me whether this hub-vulnerability is simply a function of activity level (of any type) or whether it’s more a function of some form of connectivity and activity that is unique to hubs and may be less dependent on the actual level of activity (as might be measured by fMRI or FDG, for example).

The implications to this work appear rather ominous to me. If there is this arrangement of hubs, it’s highly likely that this architecture and its normal functioning are essential to normal cognition. Therefore, it may be very difficult to affect the hub network without bad consequences. However, it may be that some people have hyperactive hubs and they may...  Read more

As is usually the case with work from Buckner, Sperling, and Johnson, this is very interesting and innovative work. It’s similar in some respects to the 2005 J. Neurosci paper comparing the topography of the default mode network to amyloid deposition. Indeed, many of the hubs lie in this region, and hub activity may be at the root of default mode activity and, in turn, may exacerbate Aβ deposition. It’s not completely clear to me whether this hub-vulnerability is simply a function of activity level (of any type) or whether it’s more a function of some form of connectivity and activity that is unique to hubs and may be less dependent on the actual level of activity (as might be measured by fMRI or FDG, for example).

The implications to this work appear rather ominous to me. If there is this arrangement of hubs, it’s highly likely that this architecture and its normal functioning are essential to normal cognition. Therefore, it may be very difficult to affect the hub network without bad consequences. However, it may be that some people have hyperactive hubs and they may benefit from some downregulation without noticing any untoward effects.

One thing is sure: this is an elegant and innovative hypothesis and could suggest novel avenues of treatment that would be most welcome in our current efforts against AD.

View all comments by William Klunk

This paper takes the previous associations between brain function and β amyloid deposition a step further. These investigators have previously noted the interesting similarity between regions of β amyloid deposition and the default mode network. Using a different computational approach they defined hubs as brain regions with unusually high connectivity, and they find that it is these areas that are particularly predisposed to β amyloid accumulation. The idea essentially parallels molecular studies that have shown how neural activity stimulates Aβ production.

So basically, the idea is that persistent, high levels of neural activity may be responsible for Aβ deposition. This is an attractive argument as it explains not only why β amyloid tends to occur in some regions but not others, and also because it might explain age-dependence of the disease and its ubiquity. On the other hand, it is somewhat difficult to reconcile with epidemiological data showing that cognitive activity reduces the risk of AD. The data do not also seem to fully correspond...  Read more

This paper takes the previous associations between brain function and β amyloid deposition a step further. These investigators have previously noted the interesting similarity between regions of β amyloid deposition and the default mode network. Using a different computational approach they defined hubs as brain regions with unusually high connectivity, and they find that it is these areas that are particularly predisposed to β amyloid accumulation. The idea essentially parallels molecular studies that have shown how neural activity stimulates Aβ production.

So basically, the idea is that persistent, high levels of neural activity may be responsible for Aβ deposition. This is an attractive argument as it explains not only why β amyloid tends to occur in some regions but not others, and also because it might explain age-dependence of the disease and its ubiquity. On the other hand, it is somewhat difficult to reconcile with epidemiological data showing that cognitive activity reduces the risk of AD. The data do not also seem to fully correspond to the deposition of β amyloid in subcortical structures and to some regions of prefrontal cortex. Nevertheless, as the authors point out, these data provide further testable approaches to exploring the pathogenesis of AD.

View all comments by William Jagust

Comment by Randy L. Buckner and Cindy Lustig
A major challenge to developing therapies for Alzheimer's disease is the availability of valid and robust diagnostic markers. Clinical assessment and cognitive testing have traditionally been the gold standard. Over the past decade, there has been an increasing emphasis on two categories of neuroimaging markers—those based on structural measures, and those based on metabolic measures. Greicius and colleagues, in their recent paper in the Proceedings of the National Academy of Sciences (2004), suggest a novel diagnostic marker for Alzheimer's disease, based on functional MRI measures.

Their work is based on the recent discovery of a "default network" that is ubiquitously observed in brain imaging studies of healthy, young participants (Raichle et al., 2001). Default network activity is observed during periods of rest and passive tasks that do not require targeted, effortful processing. Anticipating the work of Greicius and colleagues, it...  Read more

Comment by Randy L. Buckner and Cindy Lustig
A major challenge to developing therapies for Alzheimer's disease is the availability of valid and robust diagnostic markers. Clinical assessment and cognitive testing have traditionally been the gold standard. Over the past decade, there has been an increasing emphasis on two categories of neuroimaging markers—those based on structural measures, and those based on metabolic measures. Greicius and colleagues, in their recent paper in the Proceedings of the National Academy of Sciences (2004), suggest a novel diagnostic marker for Alzheimer's disease, based on functional MRI measures.

Their work is based on the recent discovery of a "default network" that is ubiquitously observed in brain imaging studies of healthy, young participants (Raichle et al., 2001). Default network activity is observed during periods of rest and passive tasks that do not require targeted, effortful processing. Anticipating the work of Greicius and colleagues, it is noteworthy that default network in young adults, which prominently includes regions in posterior cingulate and lateral parietal cortex, overlaps anatomically with those regions showing metabolic differences in Alzheimer's disease measured with FDG PET.

Employing a sophisticated analytic procedure that explores brain activity across networks of regions, Greicius et al. optimized the identification of the default network in elderly individuals with and without the earliest signs of Alzheimer's disease. Sensitivity and specificity of discrimination were 85 percent and 77 percent, respectively. As noted in the news story by Hakon Heimer, these numbers are promising and in the range considered clinically relevant.

Greicius and colleagues' observations are important from the perspectives of both clinical and basic science. First, the demonstrated discrimination between demented and nondemented groups using this functional MRI measure holds promise for developing a novel biomarker of Alzheimer's disease that may complement FDG PET metabolic measures. The relation between the changes in default network activity reported here and the common metabolic changes typically measured using FDG PET requires further exploration, but the possibility that the two are strongly related and that functional MRI measures may provide a complementary assessment in the early stages of dementia is intriguing.

Second, the results of their network analysis suggest a functional link to the medial temporal lobe structures that show early pathology in Alzheimer's disease. This may help to resolve the longstanding puzzle of how the pathological changes in medial temporal regions relate to the metabolic changes in parietal and posterior cingulate cortex as measured by PET. The Greicius et al. data suggest the tentative possibility that they are functionally linked and that posterior cortical changes, particularly within posterior cingulate cortex, may arise from anatomic projections between the medial temporal lobe and these regions.

In addition to the specific findings of their study, a social-scientific milestone was achieved in their paper. The data used for their discoveries were not their own. The data were downloaded from a freely available, online archive of raw functional imaging data of previously published manuscripts. The original authors who collected the data had not conceived of the form of analysis Greicius would later employ (we can speak to this point firsthand as original authors of the data). Thus, as intriguing as the results is also the process by which the discovery was made. The work of Greicius and colleagues directly illustrates the potential of open data sharing.

Reference:
Lustig C, Snyder AZ, Bhakta M, O'Brien KC, McAvoy M, Raichle ME, Morris JC, Buckner RL. Functional deactivations: change with age and dementia of the Alzheimer type. Proc Natl Acad Sci U S A. 2003 Nov 25;100(24):14504-9. Abstract

View all comments by Randy Buckner

These findings seem consistent with how the neurons of various brain loci communicate. The parietal lobe has been found in recent studies utilizing PET-PIB scans to be a prominent figure in early effects of amyloid deposition and shows high correlation with hippocampus atrophy.

fMRI studies further make a more significant correlation as well.

View all comments by Jacob Mack

This is fascinating information. I believe we are just beginning to see the potential of fMRI in the field of neuroscience. Much of my current research involves fMRI in the study of mirror neurons, and I encourage new scientists to consider further exploration of fMRI in AD research applications.

View all comments by J. Lucy Boyd

This can be regarded as a milestone in the search for conceptual patterns underlying different types of neurodegenerative disorders. In a single study, this group addressed a whole number of questions. The authors were able to demonstrate that the patterns of cerebral atrophy typically found in different neurodegenerative disorders indeed follow the pathways of pre-existing functional intrinsic connectivity networks (ICNs), which can be identified in healthy subjects. This work is all the more impressive as the authors actually used the foci of maximum atrophy detected in the different neurodegenerative disorders as a seed region for identification of the functional ICNs in healthy subjects (i.e., they identified regions in the brain which are functionally interrelated with this seed region). The similarity of the ICNs as identified in healthy subjects by this approach with the pattern of atrophy as detected in patients is striking.

These results do indeed strongly support the so-called network degeneration hypothesis, which implies that different types of...  Read more

This can be regarded as a milestone in the search for conceptual patterns underlying different types of neurodegenerative disorders. In a single study, this group addressed a whole number of questions. The authors were able to demonstrate that the patterns of cerebral atrophy typically found in different neurodegenerative disorders indeed follow the pathways of pre-existing functional intrinsic connectivity networks (ICNs), which can be identified in healthy subjects. This work is all the more impressive as the authors actually used the foci of maximum atrophy detected in the different neurodegenerative disorders as a seed region for identification of the functional ICNs in healthy subjects (i.e., they identified regions in the brain which are functionally interrelated with this seed region). The similarity of the ICNs as identified in healthy subjects by this approach with the pattern of atrophy as detected in patients is striking.

These results do indeed strongly support the so-called network degeneration hypothesis, which implies that different types of neurodegeneration follow distinct patterns of functionally associated neuronal populations in the brain. This notion has existed for a long time and is even implicitly found in terms describing neurodegenerative disorders (e.g., “system/multisystem degeneration”). However, in-vivo proof for this hypothesis so far has been sparse.

Furthermore, it is highly remarkable that the authors were able to detect network-associated atrophy patterns in different groups of neurodegeneration including Alzheimer disease and syndromes belonging to the frontotemporal lobar degenerative disorders (such as semantic dementia or frontotemporal dementia), because these disorders are typically based on different types of underlying causal pathologies (i.e., β amyloid, tau- or TDP-43 aggregation pathology). Another important finding is the detected interrelation between structure and function in healthy subjects, as demonstrated by the observed overlap between the ICNs and structural covariance networks (SCNs).

Although a number of questions are answered by the current work, it immediately raises new questions: For most neurodegenerative disorders (except for Alzheimer’s), it is not known yet if changes of functional connectivity do actually occur and, if so, if the atrophy in a specific network results in a change of functional connectivity within this network or vice versa. Possibly, it will also remain difficult to rule that measurements of functional connectivity are affected by regional atrophy (i.e., no activity can be measured where no tissue is present). Other issues to be addressed are the relation of white matter changes to the observed phenomena and the influence of developmental factors.

Furthermore, it remains to be clarified why the mentioned networks show specific susceptibility to basic underlying pathologies, also, why identical causal pathologies may result in different patterns of atrophy/neuronal dysfunction in different people, or why different causal pathologies may result in similar patterns of atrophy/dysfunction. In conclusion, this study will definitely stimulate further research in this direction and will serve as an important basis for subsequent experiments.

View all comments by Alexander Drzezga