Application of Structural and Functional Connectome Mismatch for Classification and Individualized Therapy in Alzheimer Disease

Abstract

While machine learning approaches to analyzing Alzheimer disease connectome neuroimaging data have been studied, many have limited ability to provide insight in individual patterns of disease and lack the ability to provide actionable information about where in the brain a specific patient's disease is located. We studied a cohort of patients with Alzheimer disease who underwent resting state functional magnetic resonance imaging and diffusion tractography imaging. These images were processed, and a structural and functional connectivity matrix was generated using the HCP cortical and subcortical atlas. By generating a machine learning model, individual-level structural and functional anomalies detection and characterization were explored in this study. Our study found that structural disease burden in Alzheimer's patients is mainly focused in the subcortical structures and the Default mode network (DMN). Interestingly, functional anomalies were less consistent between individuals and less common in general in these patients. More intriguing was that some structural anomalies were noted in all patients in the study, namely a reduction in fibers involving parcellations in the right anterior cingulate. Alternately, the functional consequences of connectivity loss were cortical and variable. Integrated structural/functional connectomics might provide a useful tool for assessing AD progression, while few concerns have been made for analyzing the mismatch between these two. We performed a preliminary exploration into a set of Alzheimer disease data, intending to improve a personalized approach to understanding individual connectomes in an actionable manner. Specifically, we found that there were consistent patterns of white matter fiber loss, mainly focused around the DMN and deep subcortical structures, which were present in nearly all patients with clinical AD. Functional magnetic resonance imaging shows abnormal functional connectivity different within the patients, which may be used as the individual target for further therapeutic strategies making, like non-invasive stimulation technology.

Keywords: Alzheimer's disease; brain connectivity; brain parcellation; diffusion tractography imaging; functional MRI; machine learning.

Figure 1

Workflow for the research. From the upper left to the right of this flowchart: the research starts with a standard atlas warped onto the brain, the boundaries are smooth because it is not machine learning–based. Then using the constrained spherical deconvolution–based tractography to adjust the atlas to personalize it. Process the rsfMRI to a functional matrix and structural MRI to a structural matrix by taking parcellation of atlas. The final step will be utilizing a training set in machine learning to make an anomaly matrix of structural and functional connectivity for further analysis.

Figure 2

Fiber tracts and fMRI-based brain network. (A) Parcellations and fiber tracts–based brain network pulled out from the machine learning algorithms. Three-dimensional rendering of parcellations and tractography-based MRI images for identified set of seven canonical brain connectivity networks that Only shows tracts within areas of the network. (B) Example submatrices of structural anomalies for the same patient based on affiliation in the same brain-network with (A). Normal or high variances (excluded areas) were indicated in white. Dots represent areas with less diffusion tractography fibers traces between them and normal, age-similar subjects. These maps provided a network-by-network fingerprint. CEN, central executive network; DAN, dorsal attention network; DMN, default mode network; VAN, ventral attention network.

Figure 3

A visual depiction of structural anomaly burden in these 21 subjects. This is a set of 377 bar graphs representing the total fractions of anomalies noted in each of the cortical parcellations and subcortical regions of interest expressed as a total % of possible anomalies. This gives a sense of which connections are most consistently abnormal compared to normal age-similar but healthy controls in non-variable areas. Note there are two inflection points in this graph that demonstrate steep transitions in the data. Areas to the left of the first inflection point are mostly subcortical structures, including the putamen, caudate, and thalamus, among others, and areas 10v, right 9M, bilateral 8BM, and right area 8BL. Areas between the two inflection points mainly include regions within the anterior cluster of the Default mode network. Most other areas have a lower anomaly burden and are to the right of the second inflection point.