We introduced a new clustering-based method that can clearly define the reference clusters. Using this method, clusters with homogeneous functional connectivity changes between the aMCI and CN groups were identified. The distribution of these clusters, as well as their disconnected regions (the areas that had decreased connectivity to these clusters), resembled the altered memory network regions identified in task-fMRI studies. Since intrinsic functional architecture may predict task-induced BOLD activity (Mennes et al., 2010
; Vincent et al., 2007
), the results may provide a strong physiological basis that can explain the memory task performance differences in aMCI and CN subjects.
The clustering method can advance our knowledge about the complex relationship among diseases, behavior symptoms and the neural networks. As noted, the functional connectivity difference between healthy and diseased states can guide clustering, as demonstrated here. It also can be guided by the correlation between functional connectivity and behavior assessments or the correlation between functional connectivity improvement and an outcome measure responding to a treatment, etc. Furthermore, the method may be generalized to multiple groups to detect functional connectivity differences among subjects with Alzheimer’s disease, aMCI and CN status.
A list of nine homogeneity thresholds (0.1, 0.2, …, 0.9) were used to define clusters that were functionally homogeneous above each threshold to avoid potential bias by choosing a single-homogeneity threshold. The threshold-free cluster enhancement (TFCE, Smith and Nichols, 2009
) method may prevent the problem of choosing thresholds. Although the TFCE method requires height and extent parameters optimization, it may provide a way to simplify the clustering method in future studies.
To detect group differences, the Fisher transformation (Zar, 1996
) and the two-sample t
-tests were employed. The approach is only optimal when the data are perfectly Gaussian after the transformation. If they are not, other tests (e.g., the rank sum test) may yield higher power. To further validate the findings of the current study, we performed the rank sum test on the whole data set. The identified clusters were the same. However, the computation time was 25 times longer. Therefore, we did not use the rank sum test in the leave-one-out procedure or in the subject permutation test.
The proposed method has limitations. First, it increases the computation cost. In ICA or conventional clustering methods using time course similarity, the data size is proportional to nV (#voxels) x nT (#time points). In this proposed method, the data size to be analyzed is proportional to nV x nV. This can easily increase the computation cost by the order of hundreds (nV / nT). With the ever-increasing computation power, this difficulty can be alleviated. Still, it may affect the feasibility of the method. Second, the method may only be sensitive to pairwise-type correlations. The connectivity differences may be too small to be detected at this level. They may be more discernible if one uses the full multivariate information, as in the case with the ICA. However, the number of detectable ICs also depends on the data (Smith et al., 2009
). Similarly, different data-driven approaches may fail to detect connectivity differences at some level. These are important considerations that deserve a full-length study in the future.
In conclusion, this clustering-based method clearly defines the reference clusters for detecting functional connectivity differences between aMCI and CN subjects. The functional connectivity changes of all voxels within the reference clusters are guaranteed to be homogeneous above desired levels, which is not warranted in the current available methods. The distribution of the reference clusters, as well as their disconnected regions, resembled the altered memory network regions identified in task-fMRI studies. The connectivity indices obtained from the leave-one-out analysis were significantly different between aMCI and CN subjects. The method has the potential to identify brain connectivity biomarkers, which can be used to classify disease status, predict individual behavioral performance, and monitor the efficacy of disease-modifying therapies.