New Results

A machine-learning approach for detection of local brain networks and marginally weak signals identifies novel AD/MCI differentiating connectomic neuroimaging biomarkers

Yanming Li, Jian Kang, Chong Wu, Ivo D. Dinov, Jinxiang Hu, Prabhakar Chalise, Jonathan D. Mahnken, for the Alzheimer’s Disease Neuroimaging Initiative
doi: https://doi.org/10.1101/2021.07.29.454368

Abstract

Introduction A computationally fast machine learning method is introduced for uncovering the wholebrain voxel-level connectomic spectra that differentiates different status of Alzheimer’s disease (AD). The method is applied to the Alzheimer’s Disease Neuroimaging Initiative (ADNI) Fluorinefluorodeoxyglucose Positron Emission Tomography (FDG-PET) imaging and clinical data and identified novel AD/MCI differentiating connectomic neuroimaging biomarkers.

Methods A divide-and-conquer algorithm is introduced for detect informative local brain networks at voxel level and whole-brain scale. The connection information within the local networks is integrated into the node voxels, which makes detection of the marginally weak signals possible. Prediction accuracy is significantly improved by incorporating the local brain networks and marginally weak signals.

Results Brain connectomic structures differentiating AD and mild cognitive impairment (MCI), AD and healthy, and MIC and healthy were discovered. We identified novel AD/MCI-associated neuroimaging biomarkers by integrating local brain networks and marginally weak signals. For example, networkbased signals in paracentral lobule (p-value=6.1e-5), olfactory cortex (p-value=4.6e-5), caudate nucleus (1.8e-3) and precentral gyrus (1.8e-3) are informative in differentiating AD and MCI. Connections between calcarine sulcus and lingual gyrus (p-value=0.049), between parahippocampal gyrus and Amygdala (p-value=0.025), between rolandic opercula and insula lobes (p-values=0.0028 and 0.0026). An overall prediction accuracy of 95.3% was achieved by integrating the selected local brain networks and marginally weak signals, compared to 84.0% by not considering the inter-voxel connections and using marginally strong signals only.

Conclusion (i) The connectomic structures differentiating AD and MCI are significantly different to that differentiating MCI and healthy, which may indicate different neuronal etiology for AD and MCI. (ii) Many neuroimaging biomarkers exert their effects on the outcome diseases through their connections to other markers. Integrating such connections can help identify novel neuroimaging biomarkers and improve disease prediction accuracy.

1. Introduction

Alzheimer’s disease (AD) ranks the third as a cause of death for people 75 and older1,2. Every 65 seconds, someone in the US develops AD3 and it afflicted around 5.8 million Americans in 20204. Overall, AD is considered the most expensive disease5 in US, with the overall cost of the disease being $305 billion in the U.S. in the year 20204. Before the onset of AD, individuals may experience an intermediate cognitive deterioration known as mild cognitive impairment (MCI) that are not severe enough to interfere with their daily activities. As the earliest detectable clinical stage toward AD, MCI provides an attractive checkpoint of disease-modifying intervention6,7. Compared to AD, MCI is more subtle and difficult to prognosis8,9. Early and accurate prediction of MCI and differentiating neuropathology of MCI and that of AD is crucially important for successful treatment development and precision prevention10,11,12,13.

Many cognitive tests have been developed for AD diagnosis14. However, these tests have certain limitations. For example, cognitive tests are often powerless in discriminating subtle differences between AD and MCI and between MCI and healthy. Also, cognitive tests cannot be used for novel neuroimaging biomarker discovery that may provide insights on underlying neuropathology.

High-resolution neuroimaging scans nowadays provide unparallel precision for discovering AD (MCI) associated neuroimaging biomarkers. For example, Positron Emission Tomography (PET) images15,16,17 have been successfully used to understand the neurodegenerative mechanisms18. Most of current methods focus on identifying the locales and individual values of imaging biomarkers, such as important voxels, hotspots or brain regions associated with a disease19,20,21, while ignoring the connectivity between markers. The number of potential connections between tens of thousands of voxels in a neuroimage is often of astronomical scales. For example, each ADNI PET scan used in our study consists of more than 185,000 voxels. Jointly inferring all potential connections between these voxels accounts to inverting a covariance matrix of dimension 185,000×185,000, which has a computational complexity of O(185, 000 2.375)22. Even with very powerful computing tools, it is infeasible to invert such a large-dimensional matrix.

Even though extremely challenging, detection of voxel-level brain connectomes associated with AD has attracted much interest. First, both AD and MCI can be viewed as a connectomic disorder neuropathologically, in the sense that their onsite is often accompanied with not only the loss of brain matters itself, but also the reduction of inter-neuron fiber connectivity required for healthy cognitive functioning23,24. Brain connectomics, which models the whole brain as a spectrum of network circuits, provides a systematic view to AD neuropathology and have been increasingly used to link the diseases with structural and functional neuroimages25. Most existing brain connectomic networks are regionbased functional networks, which aggregate neurons into priori-defined functionally related or spatially circumscribed regions of interest (ROIs)26,27,28,29. While the current neuropathological theories, such as beta-amyloid initiated senile plagues and tau-protein initiated neurofibrillary tangles, indicate that developments of AD and MCI are more directly related to breaks-down of connectivity at neuronal level, other than at regional level. Such connectomic patterns are more reflected in the voxel-level connections30,31. Region-based methods often lead to a loss of information and inferior prediction performance.

Another advantage of detecting inter-voxel connections is about prediction. As we will demonstrate shortly, a large portion of the predictiveness for AD is embedded in the connections between voxels. In fact, AD-risk imaging biomarkers currently identified explain only a small proportion of the disease variation32,33,34. More novel neuroimaging biomarkers are yet to be discovered. Integrating connections between voxels into the disease prediction makes the detection of marginally weak signals possible. The marginally weak signals have small power in differentiating the disease status and are not detectable by themselves. They are, therefore, usually ignored in contemporary neuroimaging association studies. However, when taken into consideration their connections with other signals, marginally weak signals could exude strong predictive effects35,36.

Such a case is illustrated in Figure 1. The left panel depicts a local brain-network in cerebellum crus on the left hemisphere consisting of twelve voxels. The network contains two marginally detectable voxels “vox 94316” and “vox 98031”. All other ten voxels are marginally weak and undetectable. The right-panel table in Figure 1 gives the marginal and local-network-adjusted mean differences between the AD and MCI groups. The first column lists the marginal mean differences (divided by the marginal standard deviations) without considering the connections between the voxels. The second column lists the localnetwork-adjusted mean differences. The local-network-adjusted mean difference for a voxel integrates its connective information with other voxels, such as the number of edges connected to it, the strength of these connections and marginal differentiating power of its connected voxels. The discriminant power of most of the ten marginally undetectable signals are significantly boosted up by incorporating their connective information. All the ten marginally weak voxels become more powerful in differentiating AD and MCI than the two marginally strong signals after adjusting for the local network structure.

Figure 1:
Figure 1:

Left panel: a local brain-network in cerebellum Crus on the left hemisphere of the brain. Red: marginally detectable (strong) signal. Yellow: marginally undetectable (weak) signals. Solid line: positive connections. Dashed line: negative connections. Line width represents strength of connections. Right panel: Mean differences between the AD and MCI groups. First column: marginal mean differences (divided by the marginal variances). Second column: local network adjusted mean differences. Integers in the brackets are the ranks of the absolute mean differences for the 12 voxels, ranked from the greatest to the smallest.

In this work, we analyzed the ADNI Fluorine-fluorodeoxyglucose (FDG)-PET imaging and clinical data by detecting local brain networks and marginally weak signals36. To the best of our knowledge, this is the first whole-brain voxel-level connectivity study in the literature. The identified connectomic signatures were then integrated into AD/MCI prediction. Our approach avoids ultrahigh-dimensional precision matrix calculation by disassembling the whole brain connectome into disjoint local brain networks. It is highly efficient in computation. By integrating the voxel-level connectivity and marginally weak signals, the prediction accuracy has been significantly improved. Moreover, meaningful biological interpretation about identified network-based signatures can be made based on our findings, which might help to advance our understanding on the mechanisms of MCI and AD.

The rest of the paper is organized as following. Section 2 introduces the methods for detection of predictive local brain networks and marginally weak signals, and prediction rules for classifying the diseases. Section 3 introduces the ADNI PET imaging datasets. Section 4 gives the analysis results and biological annotations for our findings. The paper is concluded by Section 5, where a brief discussion of relevant issues is provided.

2. Methods

Figure 2 depicts the major steps of our analysis pipeline. Details for the methods used are elaborated next.

Figure 2.
Figure 2.

Flow chart of the proposed analysis pipeline.

2.1. Notations

Let X = (X1,…,Xp)′ be the stacked vector for the intensity scores of all p voxels in an imaging scan. Denote by Xi = (Xi,1,…,Xi,p)′ the observed X vector from subject i. Let Yi be the class indicator (coded as 0, 1, 2 for healthy, MCI and AD, respectively) for subject i, i = 1,…,n, where n is the total number of subjects. Denote by Embedded Image the n × p data matrix. Denote by nk the size of class k, k = 0, 1, 2. Denote by G = (V, E) the whole brain connectomic network, with a vertex set V ≡ {1,…,p} and an edge set E. A connection (or an edge) (j, j0) exists between two voxels j and j0 if and only if voxels Xj and Xj0 are conditionally dependent given all other voxels. Note that connection (or an edge) between two voxels is essentially different than the correlation between them. The former is a joint concept depending on all the other voxels, while the latter is a marginal concept depending only on the pair of voxels under consideration.

2.2. Detection of marginally strong signals

Marginal two sample t-tests are used to select marginally strong signals, which each differentiates a pair of classes by itself. Specifically, for each feature j, j = 1,…, p, we calculate Embedded Image for a class pair (k, k′), k ≠ k′, in {0,1,2}. Here Embedded Image and Embedded Image. The first τ ≤ n features with the highest Embedded Image values are selected as the marginally strong signals. Here τ is a tuning parameter and can be selected by data-driven procedures such as cross-validation.

2.3. Detection of predictive local brain networks

Under the assumption that the p voxels follow a multivariate normal distribution with a mean vector μ and a covariance matrix Σ, The network G can be characterized by the precision matrix Ω = Σ-1. An edge (j, j′) exists if and only if the (j, j′)th entry of Ω, denoted by Ωjj′, is nonzero and the strength of the edge equals to the magnitude of Ωjj. As stated before, jointly estimating the whole Ω matrix of all 185,000 voxels is computationally prohibitive. Here we employed a “divide-and-conquer” algorithm introduced in Li et al.36, to detect the local brain networks. The local networks are much smaller in size and thus their corresponding precision matrices are much easier to calculate. Assume that each local brain network contains at least one marginally strong voxel. These marginally strong voxels serve as hubs in the local networks. For each marginally strong signal detected, we look for the set of voxels connected to it, either directly through an edge or indirectly through a path consisting of a series of edges. That is, to find the connected component in Ω containing it. However, it is impossible to detect such connected components without knowing Ω. Utilizing a statistical property which states that the connected component structure of Ω can be asymptotically recovered by that of the thresholded sample correlation matrix, Embedded Image)37,38, we can detect the connected components in Embedded Image instead. Here Embedded Image is the estimated sample correlation matrix and Embedded Image is its (j, j′)th entry and 1 is the indicator function. The thresholding parameter 0 < α < 1 controls the sparsity of Embedded Image. The computational complexity of Embedded Image is orders of magnitude lighter than that of estimating Ω. Moreover, since the correlations can be estimated pairwise, they can be calculated in a parallel way on multi-core computer clusters.

Recursive labeling algorithm39 is employed in detecting the connected component in Embedded Image for each marginally strong signal. Denote by Cl, l = 1,…,B, the B connected components identified. Each Cl indexes a local brain network. Let Embedded Image be the sub-sample covariance matrix corresponding to the set Cl. The precision matrices Embedded Image characterize the topology for the corresponding local networks. Sizes of Cl s can be controlled by carefully choosing the thresholding parameter α.

2.4. Detection of marginally weak signals

Local-network-adjusted effects are then calculated for voxels within each Cl. Specifically, for each Cl, the following network-adjusted statistics vector is calculated: Embedded Image

For each voxel j in Cl, its network-adjusted statistic is the corresponding entry in Embedded Image. Specifically, Embedded Image

Compared to the marginal statistics Embedded Image in (1), instead of standardizing by the marginal variation (measured in sΔ), Embedded Image also adjusts for the local network connections for feature j (estimated in Embedded Image) and the marginal differential effects of voxels connected to feature j (measured in Embedded Image for j′ ≠ j in Cl).

Not all the voxels in Cls are necessarily predictive. To reduce false positives, we further select predictive marginally weak signals within Embedded Image. The top ranked v voxels with the greatest Embedded Image are selected, beside of the marginally strong signals. Here v is a tuning parameter controlling the size of marginally weak signal set. The predictive signals (both marginally strong and marginally weak), together with the local brain networks they are residing in, form the AD (or MCI) predictive brain connectome. Figure 3 gives a toy example about incorporating a local brain network into marginally weak signal detection.

Figure 3.
Figure 3.

A toy example of a local brain network with five voxels. (a) Topology of the network. numbers on each edge are for the edge strength. Negative numbers mean the two voxels are negatively connected. (b) The corresponding precision matrix ΩCl. (c) Marginal differentiating effects between AD and MCI groups. Red: marginally strong signals. (d) Local network adjusted differentiating effects between AD and MCI groups. Yellow: marginally weak signals.

2.5. Assigning p-values to selected signals and local brain networks

A non-parametric permutation test is used to evaluate significance of selected signals40,41. Specifically, the p-value for a selected voxel j is calculated through the following procedure:

  1. Calculate Embedded Image and Embedded Image using the original data. If voxel j was not selected, then Embedded Image.

  2. Randomly permute the class label vector (Y1,…,Yn)′ M times for some large number M and generate M permuted datasets.

  3. For each feature j, calculate (2) using the mth permuted dataset, denoted by Embedded Image.

  4. Assign p-value to a selected voxel j to be Embedded Image

To access significance of the network-based signatures identified, we assign network p-values to the selected local brain networks using the following Hotelling’s T-squared distribution Embedded Image where, c is the number of voxels in Cl and Fc,nk+nk′−c is the F-distribution with parameters c and nk + nk′ −c.

2.6 Disease status prediction

Let Xnew and Znew be the predictor vector and adjusting covariate vector for a new subject in the test dataset, respectively, then the class of the new subject can be predicted by the following rule. Embedded Image where Xnew,Cl is the subvector of Xnew indexed by Cl.

3. ADNI datasets

During the last two decades, large amount of neuroimaging, genetic and clinical data have been acquired in ADNI studies42,43,44. In this study, we used the ADNI phase-I FDG-PET baseline scans with 54 AD patients, 131 MCI patients and 72 healthy subjects to detect brain connectomic structures for AD and MCI prediction.

3.1 Data processing

Standard preprocessing steps including co-registration, normalization and spatial smoothing (8 mm full width at half maximum) were applied to the PET dataset. Each image is registered to a template with 185,405 active voxels embedded in a 91 x 109 x 91 3D-array. We further grouped the 185,405 voxels into 116 regions of interest (ROI) segmented by the automated anatomical labeling atlas45. ROI names and numbers are listed in the Table A1 in the Appendix.

4. Analysis and results

Pairwise correlations and covariances between all voxels were first calculated on parallel threads using a Hewlett-Packard workstation with 16 core Intel Xeon processors and a 256GB memory size. Correlation matrix was then thresholded at 0.8 to retain only highly correlated voxels. To further reduce the computational burden, we set all entries within a thresholded correlation block between two ROIs to zeros if the average L2 norm (L2 norm divided by the number of nonzero entries in the block) of that block is ≤ 0.4. Thus, only highly correlated ROIs were be considered. For example, For ROI 5, only its closely connected ROIs: ROI 6, 9, 15, 25, 26, 27 and 28, were considered in our analysis (see Figure 4-a). This correlation block structure represents an overarching region-level connectivity structure (Figure 4-b).

Figure 4.
Figure 4.

(a) ROI block-wise thresholded sample correlation matrix. A cell represents the weighted L2 norm of the correlation block between a pair of ROI’s (weighted by Embedded Image where v1 and v2 are the numbers of voxels in the two ROIs, respectively) in the voxel-level correlation matrix. Raw voxel-level correlations are thresholded at 0.8. The weighted L2 norms are further thresholded at 0.8, which is the median of the 6,670 off-diagonal ROI block L2 norms. All blocks with L2 norm ≤ 0.4 are set to zero. The black squares represent the zero blocks. The lighted squares are for non-zero blocks. Lighter color indicates greater L2 norm values. (b) 3D ROI connection topology viewed from sagittal (left), axial (middle) and coronal (right) directions. Each dot represents a central point of an ROI. An edge exists between two ROIs if the corresponding thresholded correlation block is nonzero.

Variable selection was performed separately within each highly correlated ROI clusters. Parallel computing was implemented for this step. We used τ = 5, α = 0.8 and v = 20 on each ROI cluster to select local-brain networks, and marginally strong and weak signals that differentiated a pair of classes. When searching for the connected component of a marginally strong signal, we focused on the voxels that are connected to a marginally strong signal with a connection path less than 10 in length. We used M = 10,000 in the non-parametric permutation tests. In total, 2,686 informative voxels were selected. Among them, 836 are marginally strong and 1,850 are marginally weak. There were 661 (302), 695 (306) and 610 (309) were marginally weak (marginally strong) signals for differentiating AD and MCI, AD and Healthy, and MCI and Healthy, respectively. There were 47 (61) marginally weak (marginally strong) signals that differentiate both AD-to-MCI and AD-to Healthy, 3 (0) marginally weak (marginally strong) signals that differentiate both AD-to-MCI and MCI-to-Healthy, 67 (20) marginally weak (marginally strong) signals that differentiate both AD-to Healthy and MCI-to-Healthy, and 1 marginally weak signal that differentiate all three pairs. The selected informative signals differentiating AD and MCI, AD and healthy, and MCI and healthy are sitting in 168, 129 and 122 local brain networks, respectively. Table 1 lists the top selected voxel signals that are informative in differentiating at least one pair of disease status. Figure 5 shows the top selected marginally weak signals, along with their marginal and network-adjusted effects. Note that some regions identified containing mostly marginally weak signals have either not been reported to be associated with AD or MCI, or just been discovered recently, such as middle temporal gyrus46, olfactory bulbs47, lingual gyrus48 and amygdala69. These novel findings demonstrate the power of our method in identifying novel neuroimaging biomarkers. Table 2 lists the top selected brain networks with a significant unadjusted p-values. Most of selected voxels would still be significant after adjusting for multiple tests. The topologies of these networks are depicted in Figure 6. For the sake of presentation, we only included 12 the most closely connected voxels to a marginally strong signal in each depicted network. The whole brain connectomic structure that differentiate AD and MCI is presented in Figures 7 and 8. Each cluster of arcs in Figure 7 represent connections within an ROI. Arcs run across different clusters represent connections between different ROIs. Figure 8 shows, on the other hand, the positions and overarching topologies of the detected connectome. The corresponding figures for connectomes that differentiate AD and healthy, MCI and healthy are provided in the Appendix.

Figure 5:
Figure 5:

The symbol Embedded Image, defined in equation (3), represents the local network adjusted mean difference between a pair of classes for a voxel. The symbol Embedded Image, defined in equation (1), represents the marginal mean difference between a pair of classes for a voxel. The triplet (x, y, z) represents the coordinates of the top voxel signal highlighted in yellow circle. ROI gives the regions of interest for the top voxel signal.

Figure 6:
Figure 6:

Example local brain networks that differentiate AD and MCI groups. Yellow: marginally weak signals. Red: marginally strong signals. Blue: other selected connecting signals that differentiate other pairs. Node sizes are proportional to the local-network-adjusted mean differences. Edge widths are proportional to the absolute partial correlations. The top selected voxels within each network and the ROI(s) the network resided in are indexed.

Figure 7:
Figure 7:

Local brain network connectome differentiating AD and MCI. Each dot represents a voxel and each cluster of arcs represents a local network detected. Different colors represent different ROIs. A network connecting dots in different colors indicates that the network is crossing multiple ROIs.

Figure 8:
Figure 8:

Overlaying of the brain network connectome differentiating AD and MCI on the brain. All local networks are projected to a coronal slice of the brain at midline. Red voxel: marginally strong signal. Yellow voxel: marginally weak signal. Blue voxel: non-selected connecting voxel in a selected local brain network (might differentiate other pair of classes).

View this table:
Table 1.

Top selected voxel signals*.

View this table:
Table 2:

Top selected local brain networks*.

For many of the local networks identified, associations between their residing ROI and AD status are supported by literature evidence. These threads of evidence are listed in the last column of Table 2. Our analysis showed that connections within paracentral lobule (p-value=6.1e-5), olfactory cortex (p-value=4.6e-5), caudate nucleus (1.8e-3) and precentral gyrus (1.8e-3) are informative in differentiating AD and MCI; Connections within cerebellum crus (p-value=0.03), lingual gyrus and calcarine sulcus (p-value=3.8e-6) and amygdala (p-value=1.6e-8) are informative in differentiating AD and the healthy; Connections within precuneus gyrus (p-value=2.7e-7), middle temporal gyrus (p-value=1.6e-3), middle orbitofrontal cortex (p-value=0.014) and inferior parietal lobule (p-value=0.038) are informative in differentiating MCI and the healthy.

Connections across different ROIs that contribute to differentiating a pair of disease status are of particular interest, as they may indicate the functional connection that contributes to AD etiology. The following cross-ROI connections were identified for differentiating AD and MCI: connections between calcarine sulcus and lingual gyrus (network p-value=0.049), connections between parahippocampal gyrus and amygdala (network p-value=0.025), connections between rolandic opercula and insula lobes (network p-values=0.0028 in the left hemisphere and =0.0026 in the right hemisphere). The connections between calcarine sulcus and lingual gyrus (network p-value=7.5e-6) and connections between parahippocampal gyrus and amygdala (network p-value=1.7e-4), were also identified to differentiate AD and the healthy. Other significant cross-ROI networks that differentiate AD and the healthy, MCI and the healthy are listed in Table S2 in the Appendix.

The patterns of the selected connectomes differentiating the three disease pairs (AD-to-MCI, AD-to-healthy, and MCI-to-healthy) are significantly different. See Figure 8 and Figures A3 and A6 in the Appendix. While connections between calcarine sulcus and lingual gyrus, between parahippocampal gyrus and amygdala are informative for differentiating both AD-to-MCI and AD-to-healthy; connections between cingulum posterior, left hemisphere and cingulum posterior, right hemisphere, between cingulum middle, left hemisphere and cingulum middle, right hemisphere are informative for differentiating both AD-to-healthy and MCI-to-healthy. Note that there is no selected local network that differentiates both AD-to-MCI and MCI-to-healthy, which indicates that the development of MCI and the progression from MCI to AD may have different neuropathology pathways.

For prediction, we used five-fold cross-validation to divide the whole dataset into five parts with about the same sample sizes. Each time, we combined four folds into a training set and set the rest fold as the test set. We applied our analysis procedure on the training dataset to select predictive voxels and networks. Same tuning parameters τ = 5, α = 0.8 and v = 20 were used on each ROI cluster. The selected voxels on 116 ROIs were then pooled together and used to predict the disease status on the test dataset, along with covariates age and sex. We repeated this procedure till each fold has been used as the test set once. The overall classification errors were then computed by summing over all five test sets. The prediction results are summarized in Table 3. For comparison, we also reported prediction results from (i) marginal linear discriminant analysis (LDA), which assumes all voxels are independent (unconnected), (ii) sure independent screening (SIS)49,50, and (iii) iterative sure independent screening (ISIS)50,51. For SIS and ISIS, logistic regressions were applied on each pair of classes. The class with the highest average predicted probability was assigned to be the final predicted class. Numbers of misclassification are given in Table 3. Our approach gives the smallest prediction errors overall and within each class.

View this table:
Table 3:

Numbers of mis-classified subjects by different methods.

5. Discussion

The computational burden of the proposed analysis pipeline is comparable to that from the marginal approaches49,50. The major computational cost comes from calculating the whole brain correlation matrix. However, this can be alleviated by parallel computing. The connected component searching could also be computation-intensive if the number of connected voxels to a marginally strong signal is large. In our application, we restricted the length of any connection path in a connected component to be less than 10. This way, only the most closely connected voxels were maintained in a large-sized connected component. The analysis source codes were packed into an R package and can be found at https://github.com/lyqglyqg/mLDA

Strengths and limitations

The strengths of this study are two-fold. First, it integrates inter-voxel connectivity into neuroimaging signal selection, and makes it possible to select both network-based and marginally weak signals. Secondly, different to many principal component analyses in neuroimaging association studies, our approach can select neuropathologically meaningful biomarkers - the local brain networks, while achieving high prediction accuracy.

Accurate estimation for the covariance structures requires greater sample sizes compared to for the mean structures. Breaking down the overall brain connectomic structures onto separate ROI clusters (see Figure 4) may cause loss of information. It is recognized that the inter-voxel connectivity may not fully reveal the connection pattern between neurons.

Disclosure of Potential Conflicts of Interest

No potential conflicts of interest were disclosed by the authors.

References

  1. ↵
    James, B. D., Leurgans, S. E., Hebert, L. E., Scherr, P. A., Yaffe, K., and Bennett, D. A. (2014). Contribution of Alzheimer disease to mortality in the United States. Neurology, 82(12):1045–1050.
    OpenUrlCrossRefPubMed
  2. ↵
    Xu, J., Murphy, S. L., Kochanek, K. D., Bastian, B., and Arias, E. (2018). Deaths: Final data for 2016. National Vital Statistics Reports, 67(5).
  3. ↵
    Alzheimer’s Association report (2018). 2018 Alzheimer’s disease facts and figures. Alzheimer’s Dementia, 14(3):367–429.
    OpenUrl
  4. ↵
    Alzheimer’s association report (2020). 2020 Alzheimer’s disease facts and figures. Alzheimer’s Dementia., 16:391–460.
    OpenUrl
  5. ↵
    Hurd, M. D., Martorell, P., Delavande, A., Mullen, K. J., and Langa, K. M. (2013). Monetary costs of dementia in the united states. New England Journal of Medicine, 368(14):1326–1334.
    OpenUrlCrossRefPubMedWeb of Science
  6. ↵
    Gauthier, S., Reisberg, B., Zaudig, M., Peterson, R. C., Ritchie, K., Broich, K., et al. (2006). Mild cognitive impairment. The Lancet, 367(9518):1262–1270.
    OpenUrl
  7. ↵
    Markesbery, W. R. (2010). Neuropathologic alterations in mild cognitive impairment: a review. Journal of Alzheimer’s Disease, 19(1):221–228.
    OpenUrl
  8. ↵
    Mitchell, A. J. and Shiri-Feshki, M. (2009). Rate of progression of mild cognitive impairment to dementia - meta-analysis of 41 robust inception cohort studies. Acta Psychiatrica Scandinavica, 119(4):252–265.
    OpenUrlCrossRefPubMedWeb of Science
  9. ↵
    Vos, S. J., Verhey, F., Frolich, L., Kornhuber, J., Wiltfang, J., Maier, W., Peters, O., Ruther, E., Nobili, F., Morbelli, S., et al. (2015). Prevalence and prognosis of Alzheimer’s disease at the mild cognitive impairment stage. Brain, 138(5):1327—1338.
    OpenUrlCrossRefPubMed
  10. ↵
    A Sanchez-Catasus, C., N Stormezand, G., Jan van Laar, P., P De Deyn, P., Alvarez Sanchez, M., and AJO Dierckx, R. (2017). FDG-PET for prediction of AD dementia in mild cognitive impairment. a review of the state of the art with particular emphasis on the comparison with other neuroimaging modalities (MRI and perfusion SPECT). Current Alzheimer Research, 14(2):127–142.
    OpenUrl
  11. ↵
    Brodmann, K. (2010). Brodmann’s Localisation in the Cerebral Cortex. Springer.
  12. ↵
    McKhann, G. M., Knopman, D. S., Chertkow, H., Hyman, B. T., Jack Jr, C. R., Kawas, C. H., Klunk, W. E., Koroshetz, W. J., Manly, J. J., Mayeux, R., et al. (2011). The diagnosis of dementia due to Alzheimer’s disease: Recommendations from the national institute on aging-Alzheimer’s association workgroups on diagnostic guidelines for Alzheimer’s disease. Alzheimer’s, Dementia, 7(3):263–269.
    OpenUrl
  13. ↵
    Mueller, S. G., Weiner, M. W., Thal, L. J., Petersen, R. C., Jack, C. R., Jagust, W., Trojanowski, J. Q., Toga, A.W., and Beckett, L. (2005). Ways toward an early diagnosis in Alzheimer’s disease: the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Alzheimer’s, Dementia, 1(1):55–66.
    OpenUrl
  14. ↵
    Albert, M. S., DeKosky, S. T., Dickson, D., Dubois, B., Feldman, H. H., Fox, N. C., Gamst, A., Holtman, D. M., Jagust, W. J., Petersen, R. C., Snyder, P. J., Carrillo, M. C., Thies, B., and Phelps, C. H. (2011). The diagnosis of mild cognitive impairment due to Alzheimer’s disease: recommendations from the national institute on aging-Alzheimer’s association workgroups on diagnostic guidelines for Alzheimer’s disease. The Journal of the Alzheimer’s Association, 7(3):270–279.
    OpenUrl
  15. ↵
    Garibotto, V., Herholz, K., Boccardi, M., Picco, A., Varrone, A., Nordberg, A., Nobili, F., Ratib, O., for the Roadmap, G. T. F., et al. (2017). Clinical validity of brain uorodeoxyglucose positron emission tomography as a biomarker for Alzheimer’s disease in the context of a structured 5-phase development framework. Neurobiology of Aging, 52:183–195.
    OpenUrl
  16. ↵
    Hoffman, J. M., Welsh-Bohmer, K. A., Hanson, M., Crain, B., Hulette, C., Earl, N., and Coleman, R. E. (2000). FDG PET imaging in patients with pathologically verified dementia. J Nucl Med, 41:1920–1928.
    OpenUrlAbstract/FREE Full Text
  17. ↵
    Wee, C.-Y., Yap, P.-T., Shen, D., and Alzheimer’s Disease Neuroimaging Initiative (2013). Prediction of Alzheimer’s disease and mild cognitive impairment using cortical morphological patterns. Human Brain Mapping, 34(12):3411–3425.
    OpenUrl
  18. ↵
    Jack Jr, C. R., Bennett, D. A., Blennow, K., Carrillo, M. C., Dunn, B., Haeberlein, S. B., et al. (2018). Nia-aa research framework: Toward a biological definition of Alzheimer’s disease. Alzheimer’s Dementia, 14(4):535–562.
    OpenUrl
  19. ↵
    Frisoni, G. B., Fox, N. C., Jack, C. R., Scheltens, P., and Thompson, P. M. (2010). The clinical use of structural MRI in Alzheimer disease. Nature Reviews Neurology, 6:67–77.
    OpenUrl
  20. ↵
    Lowe, V. J., Bruinsma, T. J., Wiste, H. J., Min, H.-K., Weigand, S. D., Fang, P., Senjem, M. L., Theneau, T. M., Boeve, B. F., Josephs, K. A., Pandey, M. K., Murray, M. E., Kantarci, K., Jones, D. T., Vemuri, P., Gra_-Radford, J., Schwarz, C. G., Machulda, M. M., Mielke, M. M., Roberts, R. O., Knoman, D. S., Petersen, R. C., and Jack, C. R. (2019). Cross-sectional associations of tau-pet signal with cognition in cognitively unimpaired adults. Neurology, 93(1):29–39.
    OpenUrl
  21. ↵
    Scheltens, P., Fox, N. C., Barkhof, F., and De Carli, C. (2002). Structural magnetic resonance imaging in the practical assessment of dementia: beyond exclusion. The Lancet Neurology, 1:13–21.
    OpenUrl
  22. ↵
    Coppersmith, D. and Winograd, S. (1990). Matrix multiplication via arithmetic progressions. Journal of Symbolic Computation, 9(3):251.
    OpenUrlCrossRefWeb of Science
  23. ↵
    Delbeuck, X., Van der Linden, M., and Collette, F. (2003). Alzheimer’s disease as a disconnection syndrome? Neuropsychology Review, 13(2):79–92.
    OpenUrlCrossRefPubMedWeb of Science
  24. ↵
    Zhou, J., Gennatas, E. D., Kramer, J. H., Miller, B. L., and Seeley, W. W. (2012). Predicting regional neurodegeneration from the healthy brain functional connectome. Neuron, 73(6):1216–1227.
    OpenUrlCrossRefPubMedWeb of Science
  25. ↵
    Contreras, J. A., Goni, J., Risacher, S. L., Sporns, O., and Saykin, A. J. (2015). The structural and functional connectome and prediction of risk for cognitive impairment in older adults. Current Behavioral Neuroscience Reports, 2(4):234–245.
    OpenUrlCrossRef
  26. ↵
    Craddock, R. C., James, G. A., Holtzheimer III, P. E., Hu, X. P., and Mayberg, H. S. (2012). A whole brain fMRI atlas generated via spatially constrained spectral clustering. Human Brain Mapping, 33(8):1914–1928.
    OpenUrlCrossRefPubMedWeb of Science
  27. ↵
    Li, Y., Yao, Z., Zhang, H., and Hu, B. (2018). Indirect relation based individual metabolic network for identification of mild cognitive impairment. J Neurosci Methods, 309:188–198.
    OpenUrl
  28. ↵
    Parker, D. B. and Razlighi, Q. R. (2019). Task-evoked negative bold response and functional connectivity in the default mode network are representative of two overlapping but separate neurophysiological processes. Sci Rep, 9(1):14473.
    OpenUrl
  29. ↵
    Shen, X., Tokoglu, F., Papademetris, X., and Constable, R. T. (2013). Groupwise whole-brain parcellation from restingstate fMRI data for network node identification. NeuroImage, 82:403–415.
    OpenUrlCrossRefPubMedWeb of Science
  30. ↵
    Guest, O. and Love, B. C. (2017). What the success of brain imaging implies about the neural code. eLife, 6:e21397.
    OpenUrlCrossRef
  31. ↵
    Kriegeskorte, N., Cusack, R., and Bandettini, P. (2010). How does an fMRI voxel sample the neuronal activity pattern: compact-kernel or complex spatiotemporal filter? NeuroImage, 49(3):1965–1976.
    OpenUrlCrossRefPubMedWeb of Science
  32. ↵
    Lee, G., Nho, K., Kang, B., et al. (2019). Predicting Alzheimer’s disease progression using multi-modal deep learning approach. Sci Rep, 9:1952.
    OpenUrl
  33. ↵
    Li, J., Zhang, Q., Chen, F., Yan, J., Kim, S., Wang, L., Feng, W., Saykin, A. J., Liang, H., and Shen, L. (2015). Interactions explain variance in cingulate Amyloid burden: An AV-45 PET genome-wide association and interaction study in the ADNI cohort. BioMed research international, 647389.
  34. ↵
    Ridge, P., Mukherjee, S., Crane, P., Kauwe, J., and Consortium, A. D. G. (2013). Alzheimer’s disease: analyzing the missing heritability. PLoS ONE, 8(11):e79771.
    OpenUrlCrossRefPubMed
  35. ↵
    Li, Y., Hong, H. G., Ahmed, E., and Li, Y. (2019a). Weak signals in high-dimensional regression: Detection, estimation and prediction. Appl. Stochastic Models Bus., Ind., 35(2):283–298.
    OpenUrl
  36. ↵
    Li, Y., Hong, H. G., and Li, Y. (2019b). Multiclass linear discriminant analysis with ultrahigh-dimensional features. Biometrics, 75(4):1086–1097.
    OpenUrl
  37. ↵
    Bickel, P. and Levina, E. (2008). Covariance regularization by thresholding. Ann. Statist., 36(6):2577–2604.
    OpenUrl
  38. ↵
    Fan, J., Liao, Y., and Min, M. (2011). High-dimensional covariance matrix estimation in approximate factor models. Ann. Statist., 39(6):3320–3356.
    OpenUrl
  39. ↵
    Shapiro, L. and Stockman, G. (2002). Computer Vision. Prentice Hall.
  40. ↵
    Subramanian, A., Tamayo, P., Mootha, V. K., Mukherjee, S., Ebert, B. L., Gillette, M. A., Paulovich, A., Pomeroy, S. L., Golub, T. R., Lander, E. S., and Mesirov, J. P. (2005). Gene set enrichment analysis: A knowledge-based approach for interpreting genome-wide expression profiles. Proceedings of the National Academy of Sciences, 102(43):15545–15550.
    OpenUrlAbstract/FREE Full Text
  41. ↵
    Wu, M. C., Zhang, L., Wang, Z., Christiani, D. C., and Lin, X. (2009). Sparse linear discriminant analysis for simultaneous testing for the significance of a gene set/pathway and gene selection. Bioinformatics, 25(9): 1145—1151.
    OpenUrlCrossRefPubMedWeb of Science
  42. ↵
    Jack, C. R. J., Bernstein, M. A., Fox, N. C., Thompson, P., Alexander, G., Harvey, D., et al. (2008). The Alzheimer’s disease neuroimaging initiative (ADNI): MRI methods. journal of magnetic resonance imaging. JMRI, 27(4):685–691.
    OpenUrl
  43. ↵
    Moon, S.W., Dinov, I. D., Hobel, S., Zamanyan, A., Choi, Y. C., Shi, R., Thompson, P. M., Toga, A. W., and Initiative, A. D. N. (2015). Structural brain changes in early-onset Alzheimer’s disease subjects using the LONI pipeline environment. Journal of neuroimaging: official journal of the American Society of Neuroimaging, 25(5):728–737.
    OpenUrlCrossRef
  44. ↵
    Weiner, M. W. and Veitch, D. P. (2015). Introduction to special issue: Overview of Alzheimer’s disease neuroimaging initiative. Alzheimer’s, Dementia, 11(7):730–733.
    OpenUrl
  45. ↵
    Tzourio-Mazoyer, N., Landeau, B., Papathanassiou, D., Crivello, F., Etard, O., Delcroix, N., Mazoyer, B., and Joliot, M. (2002). Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the mni MRI single-subject brain. NeuroImage, 15(1):273–289.
    OpenUrlCrossRefPubMedWeb of Science
  46. ↵
    Li T, Liao Z, Mao Y, Hu J, Le D, Pei Y, Sun W, Lin J, Qiu Y, Zhu J, Chen Y, Qi C, Ye X, Su H, Yu E. Temporal dynamic changes of intrinsic brain activity in Alzheimer’s disease and mild cognitive impairment patients: a resting-state functional magnetic resonance imaging study. Ann Transl Med 2021;9(1):63. doi: 10.21037/atm-20-7214.
    OpenUrlCrossRef
  47. ↵
    Bathini, P., Mottas, A., Jaquet, M., Brai, E., Alberi, L. (2019). Progressive signaling changes in the olfactory nerve of patients with Alzheimer’s disease. Neurobiology of Aging, 76:80–95.
    OpenUrl
  48. ↵
    Yang, H., Xu, H., Li, Q., Jin, Y., Jiang, W., Wang, J., Wu, Y., Li, W., Yang, C., Li, X., Xiao, S., Shi, F., & Wang, T. (2019). Study of brain morphology change in Alzheimer’s disease and amnestic mild cognitive impairment compared with normal controls. General psychiatry, 32(2), e100005. https://doi.org/10.1136/gpsych-2018-100005.
    OpenUrlAbstract/FREE Full Text
  49. ↵
    Fan, J. and Lv, J. (2008). Sure independence screening for ultrahigh dimensional feature space (with discussion). Journal of Royal Statistical Society B, 70:849–911.
    OpenUrl
  50. ↵
    Fan, J. and Song, R. (2010). Sure independence screening in generalized linear models with np-dimensionality. The Annals of Statistics, 38(6):3567–3604.
    OpenUrl
  51. ↵
    Zhong, W. and Zhu, L. (2015). An iterative approach to distance correlation-based sure independence screening. Journal of Statistical Computation and Simulation, 85(11):2331–2345.
    OpenUrl
  52. Yao, Z., Zhang, Y., Lin, L., Zhou, Y., Xu, C., Jiang, T., and et al. (2010). Abnormal cortical networks in mild cognitive impairment and Alzheimer’s disease. PLoS Comput Biol, 6(11):e1001006.
    OpenUrlCrossRefPubMed
  53. Li, Y., Wang, Y., Wu, G., Shi, F., Zhou, L., Lin, W., Shen, D., and Alzheimer’s Disease Neuroimaging Initiative (2012). Discriminant analysis of longitudinal cortical thickness changes in Alzheimer’s disease using dynamic and network features. Neurobiol Aging, 33(2):427.e15—30.
    OpenUrlCrossRefPubMed
  54. Chen, X., Zhang, H., Zhang, L., Shen, C., Lee, S. W., and Shen, D. (2017). Extraction of dynamic functional connectivity from brain grey matter and white matter for MCI classification. Hum Brain Mapp., 38(10):5019–5034.
    OpenUrlCrossRefPubMed
  55. Li, Y., Liu, J., Gao, X., Jie, B., Kim, M., Yap, P.-T., Wee, C.-Y., and Shen, D. (2019). Multimodal hyper-connectivity of functional networks using functionally-weighted lasso for MCI classification. Medical Image Analysis, 52:80–96.
    OpenUrl
  56. Madsen, S., Ho, A., Hua, X., Saharan, P., Toga, A., Jack, C., Weiner, M., and Thompson, P. (2010). 3d maps localize caudate nucleus atrophy in 400 alzheimer’s disease, mild cognitive impairment, and healthy elderly subjects. Neurobiology of Aging, 31(8):1312–1325.
    OpenUrlCrossRefPubMedWeb of Science
  57. Liu, Y., Paajanen, T., Zhang, Y., Westman, E., Wahlund, L.-O., Simmons, A., Tunnard, C., Sobow, T., Mecocci, P., Tsolaki, M., Vellas, B., Muehlboeck, S., Evans, A., Spenger, C., Lovestone, S., and Soininen, H. (2011). Combination analysis of neuropsychological tests and structural MEI measures in differentiating AD, MCI and control groups – the addneuromed study. Neurobiology of Aging, 32(7):1198–1206.
    OpenUrlPubMed
  58. Elshafey, R., Hassanien, O., Khalil, M., Allah, M. R., Saad, S., Baghdadi, M., and El Zayady, M. (2014). Hippocampus, caudate nucleus and entorhinal cortex volumetric MRI measurements in discrimination between Alzheimer’s disease, mild cognitive impairment, and normal aging. The Egyptian Journal of Radiology and Nuclear Medicine, 45(2):511–518.
    OpenUrl
  59. Vasavada, M. M., Wang, J., Eslinger, P. J., Gill, D. J., Sun, X., Karunanayaka, P., and Yang, Q. X. (2015). Olfactory cortex degeneration in Alzheimer’s disease and mild cognitive impairment. J. Alzheimer’s Dis., 45(3):947–958.
    OpenUrlPubMed
  60. Zou, Y. M., Lu, D., Liu, L. P., Zhang, H. H., and Zhou, Y. Y. (2016). Olfactory dysfunction in Alzheimer’s disease. Neuropsychiatric disease and treatment, 12:869–875.
    OpenUrl
  61. Murphy, C. (2019). Olfactory and other sensory impairments in Alzheimer disease. Nature Reviews Neurology, 15:11—24.
    OpenUrl
  62. Wang, M., Yan, Z., and Jiang, J. (2019). Brain metabolic connectome classify mild cognitive impairment into alzheimer’s dementia. Annu Int Conf IEEE Eng Med Biol Soc., 32—35.
  63. Singh, V., Chertkow, H., Lerch, J. P., Evans, A. C., Dorr, A. E., and Kabani, N. J. (2006). Spatial patterns of cortical thinning in mild cognitive impairment and Alzheimer’s disease. Brain, 129(11):2885–2893.
    OpenUrlCrossRefPubMedWeb of Science
  64. Thomann, P. A., Schlafer, C., Seidl, U., Santos, V. D., Essig, M., and Schroder, J. (2008). The cerebellum in mild cognitive impairment and Alzheimer’s disease - a structural MRI study. Journal of Pschiatric Research, 42(14):1198–1202.
    OpenUrl
  65. Liang, P., Xiang, J., Liang, H., Qi, Z., Li, K., and Alzheimer’s Disease NeuroImaging Initiative (2014). Altered amplitude of low-frequency uctuations in early and late mild cognitive impairment and Alzheimer’s disease. Curr Alzheimer Res., 11(4):389–398.
    OpenUrlCrossRefPubMedWeb of Science
  66. Willette, A. A., Calhoun, V. D., Egan, J. M., and Kapogiannis, D. (2014). Prognostic classification of mild cognitive impairment and Alzheimer’s disease: MRI independent component analysis. 8Psychiatry Research: Neuroimaging, 224(2):81–88.
    OpenUrl
  67. Liu, X., Bai, F., Yue, C., Shi, Y., Yu, H., Luo, B., and Zhang, Z. (2014). The association between TOMM40 gene polymorphism and spontaneous brain activity in amnestic mild cognitive impairment. Journal of Neurology, 261:1499–1507.
    OpenUrl
  68. Long, X., Zhang, L., Liao, W., Jiang, C., Qiu, B., and Alzheimer’s Disease Neuroimaging Initiative (2013). Distinct laterality alterations distinguish mild cognitive impairment and Alzheimer’s disease from healthy aging: statistical parametric mapping with high resolution MRI. Hum Brain Map, 34(12):3400–3410.
    OpenUrlCrossRefPubMedWeb of Science
  69. ↵
    Iaccarino, L., Sala, A., Caminiti, S-P, Presotto, L., Perani, D., and Alzheimer’s Disease Neuroimaging Initiative (2020) In Vivo MRI Structural and PET Metabolic Connectivity Study of Dopamine Pathways in Alzheimer’s Disease. Journal of Alzheimer’s Disease, 7(3):1003–1016.
    OpenUrl
  70. Bailly, M., Destrieux, C., Hommet, C., Mondon, K., Cottier, J.-P., Beauls, E., Vierron, E., Vercouillie, J., Ibazizene, M., Voisin, T., Payoux, P., Barrse, L., Camus, V., Guilloteau, D., and Ribeiro, M.-J. (2015). Precuneus and cingulate cortex atrophy and hypometabolism in patients with Alzheimer’s disease and mild cognitive impairment: MRI and 18f-FDG PET quantitative analysis using freesurfer. BioMed Research International. doi: 10.1155/2015/583931.
    OpenUrlCrossRef
  71. Rombouts, S. A., Barkhof, F., Goekoop, R., Stam, C. J., and Scheltens, P. (2005). Altered resting state networks in mild cognitive impairment and mild Alzheimer’s disease: an fMRI study. Hum Brain Mapp, 26(4):231–239.
    OpenUrlCrossRefPubMedWeb of Science
  72. Ikonomovic, M. D., Klunk, W. E., Abrahamson, E. E., J. Wuu, C. A. M., Scheff, S. W., Mufson, E. J., and DeKosky, S. T. (2011). Precuneus amyloid burden is associated with reduced cholinergic activity in Alzheimer disease. Neurology, 77(1):39–47.
    OpenUrlCrossRefPubMed
  73. Mattioli, P., Pardini, M., Fama, F., et al. (2021). Cuneus/precuneus as a central hub for brain functional connectivity of mild cognitive impairment in idiopathic rem sleep behavior patients. Eur J Nucl Med Mol Imaging. 48(9):2834–2845. doi: 10.1007/s00259-021-05205-6.
    OpenUrlCrossRef
  74. Gupta, Y., Kim, J.-I., Kim, B. C., and Kwon, G.-R. (2020). Classification and graphical analysis of Alzheimer’s disease and its prodromal stage using multimodal features from structural, diffusion, and functional neuroimaging data and the APOE genotype. Frontiers in Aging Neuroscience, 12:238.
    OpenUrl
  75. Greene, S. J. and Killiany, R. J. (2010). Subregions of the inferior parietal lobule are affected in the progression to Alzheimer’s disease. Neurobiology of Aging, 31(8):1304–1311.
    OpenUrlPubMed
  76. Owen, J. B., Di Domenico, F., Sultana, R., Perluigi, M., Cini, C., Pierce, W. M., and Butterfield, D. A. (2009). Proteomics-determined differences in the concanavalin-a-fractionated proteome of hippocampus and inferior parietal lobule in subjects with Alzheimer’s disease and mild cognitive impairment: implications for progression of ad. J Proteome Res., 8(2):471–482.
    OpenUrlCrossRefPubMedWeb of Science
  77. Convit, A., de Asis, J., de Leon, M. J., Tarshish, C. Y., De Santi, S., and Rusinek, H. (2000). Atrophy of the medial occipitotemporal, inferior, and middle temporal gyri in non-demented elderly predict decline to Alzheimer’s disease. Neurobiol Aging, 21(1):19–26.
    OpenUrlCrossRefPubMedWeb of Science
  78. De Santi, S., de Leon, M. J., Rusinek, H., Convit, A., Tarshish, C. Y., Roche, A., Tsui, W. H., Kandil, E., Boppana, M., Daisley, K., Wang, G. J., Schlyer, D., and Fowler, J. (2001). Hippocampal formation glucose metabolism and volume losses in MCI and AD. Neurobiology of Aging, 22(4):529–539.
    OpenUrlCrossRefPubMedWeb of Science
Back to top
Posted July 30, 2021.
Download PDF

Supplementary Material

Data/Code
Email
A machine-learning approach for detection of local brain networks and marginally weak signals identifies novel AD/MCI differentiating connectomic neuroimaging biomarkers
Yanming Li, Jian Kang, Chong Wu, Ivo D. Dinov, Jinxiang Hu, Prabhakar Chalise, Jonathan D. Mahnken, for the Alzheimer’s Disease Neuroimaging Initiative
bioRxiv 2021.07.29.454368; doi: https://doi.org/10.1101/2021.07.29.454368
Twitter logo Facebook logo LinkedIn logo Mendeley logo
Citation Tools

Subject Area