The human functional connectome in neurodegenerative diseases: relationship to pathology and clinical progression

ABSTRACT

Introduction

Neurodegenerative diseases can be considered as ‘disconnection syndromes,’ in which a communication breakdown prompts cognitive or motor dysfunction. Mathematical models applied to functional resting-state MRI allow for the organization of the brain into nodes and edges, which interact to form the functional brain connectome.

Areas covered

The authors discuss the recent applications of functional connectomics to neurodegenerative diseases, from preclinical diagnosis, to follow up along with the progressive changes in network organization, to the prediction of the progressive spread of neurodegeneration, to stratification of patients into prognostic groups, and to record responses to treatment. The authors searched PubMed using the terms ‘neurodegenerative diseases’ AND ‘fMRI’ AND ‘functional connectome’ OR ‘functional connectivity’ AND ‘connectomics’ OR ‘graph metrics’ OR ‘graph analysis.’ The time range covered the past 20 years.

Expert opinion

Considering the great pathological and phenotypical heterogeneity of neurodegenerative diseases, identifying a common framework to diagnose, monitor and elaborate prognostic models is challenging. Graph analysis can describe the complexity of brain architectural rearrangements supporting the network-based hypothesis as unifying pathogenetic mechanism. Although a multidisciplinary team is needed to overcome the limit of methodologic complexity in clinical application, advanced methodologies are valuable tools to better characterize functional disconnection in neurodegeneration.

KEYWORDS:

• Functional connectome
• fMRI
• neurodegenerative diseases
• connectomics
• graph analysis

1. Introduction

Neurodegenerative diseases are an intriguing and challenging topic in neurology, as the mechanisms underlying pathology onset and development, spreading and diffusion across the central nervous system are still to be comprehensively elucidated. Answering these open questions will be crucial for the development of effective drugs and treatments, currently lacking. The main neurodegenerative diseases (including Alzheimer’s disease [AD], Parkinson’s disease [PD], the spectrum of frontotemporal lobar degeneration [FTLD] and motor neuron disease [MND]) are nowadays considered as ‘disconnection syndromes,’ i.e. pathologies where a communication breakdown occurs between distinct brain regions, leading to cognitive and/or motor dysfunctions [1–3]. This view has been supported by a growing body of scientific evidence, especially in the field of neuroimaging, which provides a wide array of computational techniques for the study of these diseases in vivo [4–7].

Recently, the translation of mathematical concepts based on the graph and its properties to the field of magnetic resonance imaging (MRI) has allowed to investigate and better understand the topological changes in the structural and functional brain network organization in humans [8], and, thus, to find evidence supporting the ‘disconnection syndrome’ hypothesis [9]. The brain is mathematically modelled as a set of ‘nodes,’ which intuitively represent the cortical and subcortical grey matter regions, and ‘edges’ that link pairs of nodes, representing either anatomical (i.e. structural) connections or inter-regional time series correlations from functional data. Thereby, such collection of nodes (brain regions or neurons) and edges (connections) defines a very comprehensive graph, known as the ‘brain connectome.’ If the edges of the graph represent the functional co-activation between two regions, we refer to the ‘functional brain connectome,’ which describes and maps the brain networks involved in complex motor, cognitive and behavioral functions. In fact, such approaches based on graph theory – i.e. connectomics – not only offer analytically powerful methods for brain network analysis and modeling, but also a comprehensive theoretical framework for understanding the biological basis of brain function and its alterations in pathological conditions, including neurodegenerative disorders [10,11].

The analysis of functional brain networks has consistently identified pivotal regions, called ‘hubs,’ whose integrity is critically important for enabling efficient neuronal signaling and communication, ensuring brain network properties such as segregation and integration [8,12]. Functional segregation refers to highly clustered connections that identify ‘modules.’ Modules are characterized by high density of connectivity among regions of the same module and low density of connectivity toward members of other modules. The most common graph metrics that describe functional segregation are the clustering coefficient, local efficiency or the participation coefficient [13]. The clustering coefficient indicates how well a node tends to cluster and so to be segregated in the network or in the brain, while the local efficiency refers to the efficiency of information transfer among neighboring, or directly connected, nodes. Finally, the participation coefficient measures the distribution of a node’s edges among the modules of a graph, if a node’s edges are entirely restricted to its module, the participation coefficient is low and identifies ‘provincial’ hubs, as their edges are not distributed widely among modules [14]. On the contrary, if edges are evenly distributed among all modules, the participation coefficient is high and identifies ‘connector’ hubs, providing functional integration, which represents the level of global communication and the ability of the network to distribute the information [14]. Among the most common graph metrics, the shortest path length provides a useful index of brain network integration properties, measuring the smallest number of edges that must be crossed to get from one node to another [13].

To date, the application of the abovementioned novel approaches and the investigation of the brain functional connectome have brought to light how there is a co-localization between the alterations of the functional network organization and the pattern of atrophy in many neurodegenerative diseases [15–19]. These findings have raised the hypothesis that alterations in the topography of brain functional connectivity (FC) architecture might precede and, therefore, govern region-specific structural changes leading to brain atrophy. This ‘network-based degeneration hypothesis’ suggests that neurodegenerative diseases might be characterized by the spread of molecular neuropathological features along the neuronal brain networks, as the most accredited hypothesis [20,21]. This implies that the different vulnerability of brain regions to pathological changes might correlate with the functional connectional proximity to the origination site (i.e. primarily affected regions) and strength of neuronal connections between the affected areas. However, many other hypotheses have been offered to explain network degeneration mechanisms, including the ‘nodal stress,’ in which brain regions are as exposed to degeneration as their network activity is heavy, the ‘shared vulnerability,’ in which brain regions linked to each other show a common protein expression signature showing a shared susceptibility to degeneration and, finally, the ‘failed trophism,’ in which brain regions losing trophic factors due to a connectivity interruption seem to be the more vulnerable to degeneration [21].

Through this review, we are aiming to show that the evaluation of the functional connectome allows not only to explore and validate models of pathological propagation, but also to answer fundamental clinical questions and needs in the context of the main neurodegenerative diseases. In contrast with the established use of brain volumetric or diffusivity measures to assess brain structural alterations in neurodegenerative disease [22], in this review we will demonstrate the great, yet mostly unexpressed potential of functional connectomic techniques in this field. First, we will highlight that studying the brain functional connectomic architecture has proven its utility to describe characteristic alterations that may support the clinical diagnosis of these conditions. Secondly, evidence suggesting the use of such techniques for monitoring and tracking in vivo changes along the disease course will be reviewed, as well as studies assessing potential use for prognostic stratification and prediction of treatment response. Finally, we will give space to the most recent and innovative connectomic techniques to investigate more in depth the functioning and the organization of the functional connectome, in order to build more accurate disease spreading models. Reaching these goals will help in developing tailored and targeted therapies for highly prevalent neurodegenerative diseases that, to date, are still unrelenting.

Although there are several additional technologies to evaluate the brain functional architecture, such as magnetoencephalography, electroencephalography or positron emission tomography, which assess electric neural activity or biochemical changes in the brain, here we focused on MRI, considering its wide availability in the clinical practice and high spatial resolution, with the aim to provide a comprehensive overview for such non-invasive technique.

We searched PubMed with terms including (‘neurodegenerative diseases’) AND (‘fmri’) AND (‘functional connectome’ OR ‘functional connectivity’) AND (‘graph analysis’ OR ‘graph metrics’ OR ‘connectomics’) for manuscripts published covering 20 years prior to our search, performed on 22 September 2022. Only original research manuscripts written in English were reviewed.

2. Pathophysiological definition to support clinical diagnosis

The subdivision of brain activity into functional networks is a very useful framework to understand the pathophysiological underpinnings of each neurodegenerative disease, as the identification of specific vulnerable networks and consequent rearrangements in global brain architecture through brain MRI holds the promise to suggest early functional biomarkers of ongoing aberrant protein accumulation and neurodegeneration, possibly before recurring to second-level, more invasive examinations such as positron emission tomography (PET) imaging or cerebrospinal fluid (CSF) analysis.

2.1. Alzheimer’s disease

Neurodegenerative dementia is most frequently caused by AD, which is characterized by an early and prominent memory impairment, followed by a gradual deterioration of all cognitive abilities. Several studies have evaluated the brain architectural reorganizations across the presentations of the AD continuum, i.e. not only in patients with full-blown AD dementia, but also in those with mild cognitive impairment (MCI) or even in the preclinical phases of the disease.

Assessing the topology of the functional connectome as a whole, most studies [17,23–25] found a significant loss in the ‘small-world’ organization of the functional brain networks centered on temporal and posterior brain regions in AD. As a ‘small-world’ network is characterized by highly clustered nodes with a limited number of global shortcuts between distant clusters – a configuration which is thought to favor functional synchronization within networks [26] –, a disorganization of such arrangement is reflected by decreased clustering and modularity measures [17,23–25]. A multicenter study [27] confirmed that relative to healthy controls, AD patients had a reduced FC in the default-mode network (DMN), basal ganglia and cingulate gyrus, along with increased connectivity in the prefrontal lobe and hippocampus. These alterations were correlated with the degree of cognitive impairment [27]. One early study [28] showed a higher clustering coefficient in AD patients compared with controls, with concomitant higher average shortest path length and in general a lower global efficiency of brain networks, suggesting a compromission in both integration of long-distance information and a decreased ability in transferring information, although counterbalanced by a stronger capacity of local processing (as indicated by the higher clustering coefficient). The difference in FC within the DMN was also capable of discriminating normal ageing from AD in another study (Figure 1) [29]. Indeed, although aging brains presented a lower global efficiency in antero-posterior interactions of the DMN hubs, such decreased integration was significantly more pronounced in AD patients compared to elderly controls. Divergent findings among studies (e.g. regarding the pattern of increased/decreased clustering coefficient) might derive from inhomogeneities between sample populations and computational techniques.

Figure 1. Functional connectivity matrices in AD patients and healthy controls. Thresholded correlation matrices (threshold = 0.197; p < 0.05) for young controls (YC), elderly controls (EC), and patients with Alzheimer’s disease (AD), with regions outlined to indicate the sub-systems; from upper left to bottom right: frontal, parietal, temporal, and hippocampal sub-systems. Abbreviations: MFG = middle frontal gyrus; SFGR = right superior frontal gyrus; SFGL = left superior frontal gyrus; PCC = precuneus-posterior cingulate; LPR = Lateral parietal right; LPL = lateral parietal left; MTLR = middle temporal lobe right; MTLL = middle temporal lobe left; HipR = Right Hippocampus; HipL = Left Hippocampus. Reprinted from [29] with permission from Elsevier.

In patients affected by MCI, consistently reduced clustering coefficient, reduced global efficiency and increased path length were found [30,31]. Furthermore, a general reduction in node degree was demonstrated, which indicates a reduced local density of connections [30]. A recent study observed similar rearrangements in the functional connectome of MCI patients, although these did not reach statistical significance, in contrast with a more severe, significant damage to the structural network topology [17].

AD causes large-scale functional disconnections even before the onset of symptoms [25,32]. In fact, subjects with CSF evidence of AD pathology who were cognitively unimpaired already showed reduced values of functional segregation and functional modularity, with increasing severity of the same pattern after symptom onset [25]. To further support the hypothesis that AD pathology tends to spread via functional networks since the preclinical phase, a study [32] demonstrated that PSEN1 mutation carriers exhibited reduced segregation and integration in posterior DMN regions and in the retrosplenial cortex compared to non-carriers. These alterations were associated with greater tau burden at tau PET and worse episodic memory in cognitively unimpaired carriers, strengthening the correlation between functional alterations, pathology and memory impairment [32].

Functional connectomics has also been employed to study atypical variants of AD. Posterior cortical atrophy patients compared to controls showed decreased small-worldness, as well as longer mean path length in parietal, temporal, occipital, sensorimotor areas and basal ganglia, both at the intra- and inter-hemispheric levels, even exceeding the pattern of structural and metabolic brain damage [33].

Graph theoretical analysis was also used to discriminate AD from the behavioral variant of frontotemporal dementia (bvFTD) [34]. AD patients had lower integration in the DMN and frontoparietal networks, while bvFTD exhibited disintegration in the salience network. Interestingly, AD and bvFTD had the highest and lowest degree of integration in the thalamus, respectively. Furthermore, aberrations in network topology were related to worse attention deficits and greater severity in neuropsychiatric symptoms across syndromes. When comparing patients affected by early-onset AD (EOAD) and bvFTD [35], EOAD patients showed severe global functional network alterations, which were relatively preserved in patients with bvFTD. FC breakdown in the posterior brain nodes differentiated EOAD from bvFTD. Ultimately, while EOAD was associated with widespread loss of both intra-hemispheric and inter-hemispheric correlations, bvFTD showed a preferential disruption of the intra-hemispheric connectivity.

The role of microglia activation in the progression of cognitive decline in AD has recently emerged. A recent study [36] found a correlation between the uptake of ligand at [¹¹C]PK11195 PET (which reflects microglia activation) and diffuse disruption of the functional brain connectome in AD, which was itself linked to cognitive performance. Henceforth, the disruption of connectivity linked to neuroinflammation is possibly a mediator of cognitive deficits in AD patients.

2.2. Parkinson’s disease

Connectomic studies performed in patients with PD have demonstrated a complex de-segregation of functional networks. In one study, PD patients showed increased intra-modular path length within medial prefrontal cortex, salience network, somatomotor network and fronto-parietal network along with a decreased inter-modular FC and decreased intra-modular clustering [37]. Disrupted modular organization was confirmed in drug-naïve early-stage PD patients, who showed disrupted global efficiency and small-worldness in temporal-occipital and sensorimotor regions [38]. These findings were also consistent with another study [39], in which early-stage PD patients exhibited lower global efficiency, along with an alteration of modular structure and hub distribution correlating with symptom severity. A decreased functional segregation and integration (both globally and locally) was observed in PD patients [40]. In another series [41] PD patients also showed reduction in local network metrics and a decreased modularity in DMN, dorsal attention and sensorimotor network.

When examining functional alterations in PD patients according to their clinical phenotypes, the involvement of different networks emerges. Non-tremor PD patients showed a decreased efficiency in the basal-ganglia motor pathway [42]. PD patients with resting tremor instead had a widespread increase of FC, along with an increased centrality in the frontal, parietal, and occipital regions, and decreased centrality in the cerebellum and thalamus [43]. However, this result was refuted by another study, in which tremor PD patients presented increased connectivity in the bilateral thalami, correlating with tremor severity [44]. Freezing of gait, which was known to be associated with executive and attentive deficits, was found to be associated with lower global efficiency at the level of the dorsal attentional network [45].

PD patients also present a distinctive topological pattern of functional disconnection based on their cognitive status. In fact, compared with cognitively unimpaired PD patients, PD MCI were found to display a loss of connections (i.e. nodal degree) in the salience network, possibly counterbalanced by increased betweenness centrality in the DMN [46]. Another study confirmed a global tendency towards increased segregation and modularity in PD MCI [47]. Graph theoretical measures were able to train a model that could discriminate MCI from non-MCI PD patients with an accuracy of 80% [48]. For what concerns the presence of neuropsychiatric symptoms, depressed PD patients showed significantly decreased local efficiency compared with controls, and nodal degree of the right inferior occipital gyrus was positively correlated with depression severity [49].

2.3. Frontotemporal lobar degeneration

A first study assessing functional connectomic alterations in bvFTD patients [50] showed a global decrease of brain network efficiency and a substantial loss of cortical hubs in the frontal lobes. Clinically, altered global network properties correlated with executive dysfunction [50]. Graph analysis has shown its utility in the differential diagnosis between bvFTD and EOAD, as previously described (see Paragraph 2.1 Alzheimer’s disease), revealing a preservation of inter-hemispheric connectivity as a signature of bvFTD [35]. In another work, different disruption of networks emerged with disintegration of the DMN in AD and disruption of salience network in bvFTD [34]. Furthermore, at the level of the thalamus, the highest and the lowest degree of integration were seen in AD and bvFTD, respectively.

Divergent patterns of network disruption could be appreciated in the different clinical variants of frontotemporal dementia [51]. Indeed, bvFTD was characterized by the compromise in hubs of the limbic system and basal ganglia, while non-fluent variant progressive primary aphasia (nfvPPA) presented a major disconnection in the operculum and parietal inferior regions [51]. In another study, while all PPA variants presented with a loss of functional hubs in the left superior frontal and parietal regions, new hubs were recruited according to the PPA variant [52]. Both logopenic and semantic variants (but not nfvPPA) were characterized by new hubs recruited in the right hemisphere, which could not be explained by atrophy [52]. Semantic variant PPA (svPPA) patients were instead characterized by greater disconnection compared to controls along with a strongly left-lateralized loss of hubs and reduced nodal degree in the inferior and ventral temporal regions and occipital cortices [53].

Functional connectomics have also been implemented in the study of carriers of mutations known to be associated with FTD, such as progranulin (GRN). One study [54] analyzed the pattern of FC in presymptomatic GRN carriers, prodromal MCI GRN carriers and controls, showing an increased FC in preclinical GRN carriers associated with hyperconnectivity between cortical hubs and the thalamus, while symptomatic carriers featured instead decreased connectivity.

2.4. Motor neuron disease

The study of functional brain organization in MND has shed light on the pattern of upper motor neuron disconnection, as well as extra-motor network derangements, which eventually cause both motor and extra-motor symptoms in MND patients.

A multicenter study [6] has analyzed network properties in the three variants of MND: amyotrophic lateral sclerosis (ALS), progressive muscular atrophy (PMA) and primary lateral sclerosis (PLS). ALS patients presented increased FC in frontal regions, while patients with PLS featured increased connectivity in the sensorimotor, basal ganglia and temporal network. These alterations were correlated with executive disfunction and behavioral derangements. Consistent with pure lower motor neuron damage, patients with PMA did not show significant functional alterations. Another series of ALS patients only evidenced an increase in FC in a subset of nodes located in the frontal, temporal, parietal and subcortical regions [55]. At a local level, a reduced connection in frontal region was counterbalanced by an increased connectivity in occipital areas. Functional disability was correlated with changes in frontal connectivity [55]. The functional network degeneration was found to overlap considerably with the most affected structural connections in ALS, suggesting that the pathogenic process affects the brain connectome both functionally and structurally (Figure 2) [56].

Figure 2. Overlap between changes in structural and functional connectivity in ALS. Connectomic representation of the brain network in which the top 10% most impaired SC connections (left) and the top 10% most impaired FC connections (right) as well as the overlapping connections are colored. The locations of homologous nodes, based on anatomy, have been symmetrized. Reprinted from [56] with permission of John Wiley and Sons.

Recent studies assessed the functional connectomic underpinnings of cognitive deficits in MND [57,58]. When evaluating brain network modifications in patients classified into ALS-cn (cognitive normal), ALS-ci/bi (cognitive impairment/behavioral impairment) and ALS-FTD, ALS-ci/bi showed a distinctive increased FC within sensorimotor regions and decreased connectivity within frontotemporal nodes, whereas ALS-FTD had widespread FC loss closely resembling a bvFTD-like pattern (Figure 3) [57]. The presence of semantic impairment in ALS was found associated with a reduced degree centrality in the right lingual/fusiform gyrus with a concomitant reduction in connectivity with regions devoted to word perception, semantic processing, or speech production, and an increased degree centrality in the left inferior/middle temporal gyrus [58].

Figure 3. Alterations in functional connectivity in patients with ALS and patients with bvFTD relative to healthy controls and each other. Altered functional connections are represented per each significant contrast, respectively (p < 0.05). Comparisons were adjusted for age, sex, and education. The node color represents its belonging to specific macro-areas (frontal, sensorimotor, basal ganglia, parietal, temporal, and occipital). The node size is proportional to the number of affected connections (the higher the number of disrupted connections, the larger the node). A = anterior; ALS = amyotrophic lateral sclerosis; ALS-ci/bi = amyotrophic lateral sclerosis with cognitive or behavioral impairment; ALS-cn = amyotrophic lateral sclerosis with motor impairment only; ALS-FTD = amyotrophic lateral sclerosis with frontotemporal dementia; bvFTD = behavioral variant of frontotemporal dementia; FA = fractional anisotropy; HC = healthy controls; P = posterior. Reprinted, from [57] with permission of Wolters Kluwer Health, Inc.

3. Tracking disease progression

Functional graph metrics have been used to describe the progressive changes in brain network organization, which occur in different neurodegenerative processes as pathological aggregates accumulate.

As mentioned previously, the network-based degeneration hypothesis is currently among the most accredited theories to describe such pathological progression, as initially suggested by some seminal seed-based RS fMRI studies [19,59]. The first evidence supporting this notion through the analysis of brain functional architecture was provided in 2012, when it was demonstrated that brain regions affected by atrophy in patients with five different neurodegenerative conditions (including AD, bvFTD, svPPA, nfvPPA, and corticobasal syndrome) were also those with higher connectional flow and shorter functional paths to the respective epicenters in the healthy functional connectome [21]. Among four competing models, the one fitting best the authors’ findings was a trans-neuronal spread model from the epicenter, although other hypotheses explaining nodal vulnerability to neurodegeneration – such as nodal stress due to increased metabolic demands of damaged networks, or shared vulnerability of neural regions with similar epigenetic backgrounds – could not be ruled out either [21].

In recent years, other cross-sectional studies have suggested a similar overlap between the spatial patterns of atrophy and epicenter-based intrinsic networks in patients with AD [17,60], PD [61], and presentations of the FTLD spectrum [62]. A recent study provided even neuropathological evidence suggesting a direct correlation between brain connectivity propagating from PD-related pathology epicenters in the brainstem and the distribution gradient of postmortem α-synuclein histopathology (Figure 4) [15]. Using tau PET as an indicator of pathological protein deposition, a correlation between tau burden and loss of connectivity has been demonstrated in AD, as increasing tau levels were mirrored by a progressive weakening of connectivity within the same nodes, in terms of decreasing weighted degree, clustering coefficient and local efficiency, resulting in weaker small-world properties [63]. In the same study, an opposite trend was demonstrated in progressive supranuclear palsy (PSP) patients, as regions that accumulated more tau displayed weaker connectivity but were also those with higher clustering coefficient and local efficiency, suggesting divergent mechanisms of network-based pathology propagation in the two tauopathies [63]. Another study has demonstrated a strong correlation between FC and tau levels, especially in inferior temporal ‘hotspots’ of tau accumulation, which was retained not only in AD but also in cognitively unimpaired elderly participants and patients affected by vascular dementia, supporting the view that tau spreading is facilitated by neural activity [16]. A similar correlation between highly functionally connected regions such as precuneus and cingulate and tau burden in AD patients was described in another study [64], which further corroborated the network-based degeneration hypothesis. In atypical variants of AD, such as posterior cortical atrophy and logopenic variant PPA (lvPPA), a strong correlation was demonstrated between the level of tau pathology in different brain regions and the strength of functional connection to the disease epicenters [65].

Figure 4. Brain co-localization of in-vivo SFC patterns and Allen gene expression data to detect PD vulnerable pathways and genetic/pathological signatures. (A) Scatterplot between in-vivo spreading connectivity pattern (bottom cortical maps) and SNCA (left cortical maps). (B) Relationship between SFC and α-synuclein-immunoreactive density scores. Abbreviations: Ant = anterior; cSCM = combined stepwise connectivity map; Inf = inferior; GO = gene ontology; L = left; Mid = middle; R = right; sc = spatial correlations; SFC = stepwise functional connectivity. Reprinted from [15] with permission of Elsevier.

To our knowledge, there is a lack of longitudinal studies describing the trajectory of functional brain connectomic alterations in AD. Instead, graph theory has been employed to track longitudinal abnormalities of FC in PD [5,66,67]. It has been observed that at baseline PD patients showed an increased connectivity between the cerebellum and the somatomotor network, in contrast with decreased connectivity between motor regions and the cingulate cortex [66]. At 1.5 years follow-up, a similar pattern of altered connectivity was appreciated, although the cerebellum presented additional hyperconnectivity within itself and to the caudate nucleus, thalamus and amygdala, correlating with scores of motor functional impairment [66]. This cerebellar hyperconnectivity could be interpreted as an ongoing attempt of recovery and a compensatory mechanism for lost function as PD unfolds. By contrast, the progression of cognitive deficits in PD patients has been associated with a loss of functional architecture, reflected in a decreased functional segregation and integration of the brain networks [67]. Indeed, clustering coefficient, mean path length and local efficiency progressively decreased with increasing severity of cognitive decline [67]. Another study classified a large cohort of PD patients in disease subtypes by using cluster analysis, describing different functional MRI trajectories at longitudinal follow-ups [5]. At baseline, moderate-to-severe PD exhibited decreased FC in sensorimotor network, basal ganglia circuit, temporal and parietal cortices compared to mild motor-predominant. Different longitudinal functional changes were appreciated in diverse PD subgroups. As the disease progressed, moderate-to-severe patients presented more severe cognitive and extra-motor symptoms, while at fMRI they showed a progressive alteration of global properties (decreased mean nodal strength, local efficiency and clustering coefficient, longer path length), which mild PD did not report. At follow up, alterations of brain functional topology were exhibited by both mild and moderate-to-severe PD cases with increase or decrease in FC. Hyperconnectivity prevailed in mild cases, while hypoconnectivity prevailed in later stages of the disease, possibly due to an initial compensation followed by neurodegeneration [5].

4. Prediction and prognostic stratification

In the context of neurodegenerative disorders, it is fundamental to elaborate an accurate stratification of patients into prognostic groups, not only for the clinical impact at an individual patient level, but also to allow testing for efficacy of upcoming disease-modifying drugs in a homogeneous population. As a matter of fact, an incorrect selection of the target population is believed to have contributed to the failure of several clinical trials in neurology [68,69]. The development of predictive models of pathology progression taking into account MRI quantitative features is a ‘hot topic’ for research in the neuroimaging field. The paradigm of the brain functional connectome is an ideal tool to prove the prognostic implications of the network-based degeneration hypothesis, through the demonstration of a close correlation between (a) the ‘healthy’ connectomic architecture of brain networks functionally inter-linked with the disease epicenter of each clinical syndrome, and (b) the subsequent progression of neurodegeneration in patients. In fact, critical proof validating the trans-neuronal spreading model in neurodegenerative diseases can be provided only in a longitudinal setting, through the demonstration that network connectivity can predict the progressive spread of neurodegeneration at the individual patient level. However, to date, only few studies have performed this kind of analysis.

In patients with amyloid-positive defined AD, the regional longitudinal increase of tau-PET tracer uptake could be predicted based on a model combining baseline tau levels, FC and topological distance between brain regions, supporting the view of tau transmission across neuronal connections (Figure 5) [16]. Similarly, the assessment of multicenter MRI data from the Parkinson’s Progression Markers Initiative (PPMI) [70] demonstrated that cortical regions with greater exposure in terms of FC to a ‘disease reservoir’ in critical subcortical regions were also those showing greater progression of atrophy over a one-year follow-up period in PD patients [71]. It has been shown that the definition of individualized seeds representing the characteristic disease epicenter of each patient might yield significantly better predictions regarding the future patterns of regional atrophy, compared with the use of a static seed defined at a group level (e.g. the hippocampal formation for AD patients) [72]. A similar individualized prediction model of longitudinal atrophy was also tested in patients with bvFTD and svPPA, demonstrating that the functional connectional distance from disease epicenters identified in each patient could be used to model accurately the whole-brain accumulation of atrophy in most patients [73]. Of note, the individual predictions were less accurate for bvFTD patients carrying a pathogenic genetic variant (e.g. the C9orf72 expansion) [73], possibly suggesting an interplay between functional connectomic architecture and (epi)genetic background of neural populations in different brain networks, which are expected to vary according to the underlying pathological substrate.

Figure 5. Prediction of longitudinal tau-PET change in amyloid-positive defined AD. a. Hypothetical network spreading model of tau pathology. Each node within the network represents a brain region, where color indicates local tau pathology, distance between regions indicates connection length (i.e. Euclidean distance) and edge thickness indicates functional connectivity strength. Example formulas for models 1–3 illustrate how we computed tau-weighted distance (Model 1), tau-weighted functional connectivity (Model 2) or tau- & distance-weighted functional connectivity (Model 3) that were used to model group-mean annual tau-PET change in the 53 Aβ + ADNI (b–d). Adapted from [16].

Besides providing information about the subsequent anatomical spread of pathological aggregates, functional connectomics has also been employed to develop prognostic models of relevant clinical milestones. For example, a number of studies have explored the influence of functional brain architecture on the rate of cognitive deterioration in AD [32,74,75]. Higher levels of functional network segregation shown by rs-fMRI were found to mitigate the effect of tau pathological burden on the decline of global cognition and episodic memory in AD [74], suggesting a possible mechanistic theory underlying cognitive reserve, in line with the significant correlation between education levels and preserved global FC in frontal networks of patients with AD [75]. A similar mechanism of resilience through preserved functional brain architecture was also suggested in presymptomatic subjects with autosomal dominant AD (ADAD) [74,75], which might be responsible for preserved cognition despite an estimated shorter time to symptom onset. Taken from a different perspective, a similar approach can be applied across the AD continuum to predict proximity to the onset of cognitive symptoms or the development of full-blown dementia. For example, a decrease in functional segregation and integration of key hubs of the DMN (i.e. the precuneus and the medial prefrontal cortex) was strongly correlated with increased regional tau burden in preclinical ADAD – and, only weakly, with amyloid-beta –, suggesting that dysconnectivity of these regions might serve as an early indicator of imminent disease progression [32], although longitudinal evidence supporting such statement is still lacking. A few studies have assessed the role of functional connectomic disruption as a means to predict clinical conversion from MCI to AD dementia [17,76], but their statistical power was limited by relatively small sample sizes.

There is a scarcity of literature on prognostic applications of functional MRI connectomics in the context of non-AD neurodegenerative conditions. In a recent study evaluating PD patients, decreased parieto-occipital nodal strength and global increased path length – indicating loss of network integration across widespread brain regions – were found to be independent predictors of cognitive progression over a 2-year follow-up, even in the absence of significant grey matter atrophy [77]. Similar to observations in ADAD, presymptomatic carriers of FTD-related genetic mutations showed substantial maintenance of brain functional network efficiency until years close to the expected symptom onset, possibly accounting for preserved cognition despite the development of subtle grey matter atrophy and the loss of functional connections [78]. The marked decline of functional network efficiency observed in early symptomatic genetic FTD suggested that such measures might serve as indicators of imminent clinical onset, although longitudinal evidence for this statement is still lacking [78].

5. Response to treatment

Considering the dynamic nature of FC, studying architectural rearrangements in the brain connectome of patients with neurodegenerative diseases in response to pharmacological or non-pharmacological treatments is an ideal, non-invasive way to test for efficacy and assess the mechanism of action of such interventions. This approach has been mostly tested in patients with PD [79–81], a population characterized by significant motor and cognitive fluctuations and a wide range of available treatment options, to date.

Although dopaminergic treatment has been used for decades in the treatment of parkinsonian syndromes, the underlying effect over brain networks is still unclear. Previous RS fMRI studies comparing PD patients in ON and OFF states suggested that dopaminergic treatment might generally act through the normalization of aberrant functional patterns [79]. Recent studies assessing the functional connectome of PD patients allowed a better definition of such rearrangements, showing a more complex interplay between brain networks. For example, injection of apomorphine in patients with tremor-dominant PD was found to significantly reduce tremor symptoms by rescuing the overall connectivity strength and the modular structure of basal ganglia and fronto-striatal networks, with increased participation coefficients indicating greater integration of motor-premotor regions [80]. Regarding the effects on cognitive networks, another study assessing PD patients treated with levodopa both in ON and OFF states showed that dopaminergic treatment specifically maximized increased connectivity between salience and default-mode networks, with a possibly maladaptive ‘excessive’ hyperconnectivity that could be used to indicate those patients who would later develop dementia within a follow-up period of 10 years [81].

Another well-established treatment option for PD patients with invalidating symptoms is the implant of a deep brain stimulation (DBS) device modulating aberrant motor pathways within the basal ganglia. DBS of the subthalamic nucleus was found to act by increasing coupling between motor thalamus and motor cortex, at the same time reducing striatal connectivity with pallidal, subthalamic and cerebellar structures, with an overall normalization effect compared with healthy controls [82]. A recent study suggested the use of functional connectomics to allow an early identification of PD patients who would soon be candidate to DBS implantation, as these patients were characterized by distinctive decreased connectivity in basal ganglia and sensorimotor brain regions as well as hyperconnectivity of parieto-occipital regions up to four years before satisfying criteria for surgery [83].

Despite the fact that several studies have explored brain network topology rearrangements following cognitive training in young healthy subjects [84–86], these approaches have been applied rarely to evaluate the effects of cognitive training in patients with PD or other neurodegenerative conditions. One study did not demonstrate significant effects over large-scale brain network metrics following an 8-week cognitive training intervention in PD patients, although more localized connectivity changes in the anterior cingulate and dorsolateral prefrontal cortices could be observed [87]. Preliminary evidence also suggests that cognitive training may specifically act in patients with AD and amnestic MCI through an up-regulation of FC within hubs of the DMN [88,89]. Although validation in larger cohorts is still needed, these findings support the use of brain network topology measures as outcomes for upcoming interventional trials.

6. Advanced methodologies: a turning point in neurology

Recently, research has been making a great effort to implement even more sophisticated connectomic techniques, trying to consider in the modelling the intrinsic dynamic nature of the brain and its topological and hierarchical functional architecture. Among these approaches, the most promising include the dynamic functional connectivity (dFC) analysis, stepwise functional connectivity analysis and effective connectivity.

6.1. Dynamic functional connectivity

Unlike the traditional and conventional approach of static FC analysis, which is obtained from the correlation between two time series averaged throughout the entire acquisition time, dFC aims to identify recurring functional patterns and to describe the dynamic brain properties in sub-portions of time-series [90]. We suggest referring to our previous review on this topic, delving into the techniques to obtain and investigate brain dynamics in the field of neurological neuroimaging [90]. The study of these functional recurring and transition states is known as the ‘chronnectome,’ which consists in the functional time-varying rearrangements of the brain connectome [91].

Such novel approach found great application for the assessment of AD, bringing to light results with high consistency. dFC studies revealed that the loss of brain dynamics, i.e. the ability to switch from states with low inter-network connectivity into more highly and specifically connected configurations, is a hallmark of AD dementia [92–94]. Patients with AD tend to spend more time in functional configurations that are sparsely connected and without strong connections compared to controls [92,94], pointing out a disruption of the temporal dynamics, already present in the early stages of disease, leading to a progressive reduction of the number of functional patterns across the AD continuum [93,95]. dFC has also allowed to characterize the functional disconnection between sensory networks in patients with mild dementia, providing a possible explanation of the pathophysiological mechanisms underlying conserved self-awareness in patients with AD [93]. A further strength of dFC lies on the identification of functional features as potential diagnostic biomarkers with the help of machine learning algorithms, which was not only useful for the stratification of patients along the AD continuum [96–98], but also for the differential diagnosis with other neurodegenerative conditions, such as Lewy body dementia [94] and FTD syndromes [99].

In PD, studies that investigated dFC suggested alterations which are present since early stages [100], linked to disease progression [101], and presence of cognitive deficits [102,103]. In particular, in patients with PD, two main configurations were found: a more frequent and sparsely connected within-network state (state I) and a less frequent and strongly interconnected between-network state (state II) [101,102]. It has been observed that patients with PD tend to shorten the dwell time in state I and to increase proportionally the dwell time in state II. This might be interpreted as the loss of functional segregation properties, which was correlated with the severity of clinical symptoms, and an increase in the global functional integration (i.e. state II) [101]. The shift toward a more integrated network might be considered as a compensation mechanism in the dopaminergic OFF state [104]. dFC might also help in better understanding the neural substrates of cognitive impairment and full-blown dementia in PD. This functional signature consists in a progressive increase of dwell time in the segregated state and a reduced number of transitions between states, concomitantly with the worsening of cognitive deficits [102]. This novel approach was also applied to explore functional alterations linked to specific neurocognitive deficits as the impulse control disorders (ICD). Patients with ICD, in contrast to patients without ICD and controls, were characterized by the absence of a sparsely connected state, at the expense of a strong within-network connection state in key nodes of salience network [105]. A longer maintenance in this state was positively correlated with the severity of impulsivity [105]. dFC analysis has also proven its value in identifying neural substrates in different phenotypes of PD, such as tremor dominant and non-tremor-dominant [106].

The analysis of dFC has demonstrated its predictive and anticipatory capability in FTD. Indeed, presymptomatic carriers of FTD-causing gene mutations showed altered dynamic functional properties relative to controls, in terms of reduced dynamic fluidity and restricted dynamic range, given by lower number of states, lower switching rate between states and a lower functional distance between states [107]. The assessment of such properties delineated an early functional signature of the closest phase to the clinical conversion [107]. Furthermore, FTD patients with higher cognitive reserve coped better with the ongoing pathological process and the accumulation of brain damage, quantified as brain flexibility in this study [108]. Finally, applying a dFC approach in ALS brought out that the functional dynamics features were found among the most discriminating biomarkers in patients with ALS relative to controls [109].

6.2. Stepwise functional connectivity

Stepwise functional connectivity (SFC) integrates the conventional approaches of connectomic, revealing not only the direct functional couplings between brain regions but also multistep associations (indirect connectivity) between a given region to the rest of the brain, reconstructing a hierarchical architecture of the functional connectome from a region of interest [110]. Even though it has been rarely applied in the context of pathological conditions, this technique has helped in better understanding the mechanisms of information flow in the brain. Indeed, it was possible to elucidate the functional organization of the main areas of multimodal integration (visual, auditory, somatosensory) and also to disentangle the complex communication transitions between primary and high-order cognitive elaboration systems of the brain [110,111]. In addition to describing the topological functional organization of the brain, SFC has been applied to reveal changes in the brain functional connectome in presence of neurodegenerative diseases, such as AD and PD. SFC allowed to characterize the functional disintegration of the brain network across the AD continuum [112] and to find subtype-specific changes in cerebellar connectivity across cortical and subcortical networks in PD patients with tremor-dominant and postural-instability-and-gait-disorder phenotypes [113]. SFC provided an interesting framework able to bring out the spatial convergence between the functional topological organization and the pathological protein accumulation. In AD, SFC was applied focusing on the spatial correlations between the medial temporal lobe (specifically hippocampus and parahippocampal gyrus) and the rest of the brain. Reconstructing the functional connectome from these regions of interest in a healthy young group, they revealed topological similarity with the amyloid-β deposit maps obtained from patients with AD [114]. In addition, the functional architecture originated from PD-related pathology epicenters (i.e. in the brainstem) followed the anatomical distribution of α-synuclein histopathology in postmortem data [15]. These findings support the hypothesis that brain FC architecture might govern region-specific structural changes in neurodegenerative diseases.

6.3. Effective functional connectivity

Brain is not only organized into a topological network architecture, but also into a causal functional configuration between its areas. In connectomics, effective connectivity (EC) is defined as the influence that a node exerts over another [115], to highlight the directionality of communication identifying the cause-and-effect relationship of alterations in the presence of pathological conditions, as neurodegenerative diseases. In AD, EC not only identified excitatory and inhibitory connectivity changes specific for disease stages [116], but also observed a hub reconfiguration within the DMN [117]. Moreover, EC provided evidence in support of molecular models suggesting that the spreading of pathological proteins is linked to the directionality of signaling pathways [118], with important implications in the therapeutic intervention. In this regard, it has been suggested as possible therapeutic intervention target the noradrenergic pathways in AD starting from brainstem, specifically locus coeruleus, which was found the critical focus of AD disruption [119]. Indeed, the observed EC alterations of locus coeruleus (i.e. the largest repository of noradrenergic neurons in the human brain) would cause a cascade of events starting from a decrease in norepinephrine levels, with consequent loss of its protective function against cognitive deterioration, the onset of inflammatory reaction and the reduction of β-amyloid clearance, whose accumulation is the initiating event of AD [119]. Even though the potentiality of EC is clear, to date, the application of EC methods remains mainly confined to AD, as only few studies have attempted to understand the complex effective connectivity of motor and non-motor regions in PD [120–122] and in ALS [123].

6.4. Machine learning

In recent years, there has been a great ‘hype’ for the use of machine learning in medicine, in general, and in neurology, specifically. The most common purpose in the application of such algorithms has been to find the most informative and useful features to answer clinical questions, yet unsolved. The main advantage in using machine learning algorithms is their ability to extract features and information by combining multimodal datasets with all possible sources of data. To date, there are few studies in the field of neurodegenerative diseases, restricted to the AD spectrum, which have applied classification algorithms exploiting the information obtained from functional connectomic approaches [76,124,125]. All these studies obtained a very good level of accuracy in discriminating patients with AD, MCI, and healthy controls using graph metrics based on the functional connectome. In addition, combining these metrics with structural data, the machine learning algorithm reached an extraordinary level of accuracy (97%) in discriminating MCI to AD converters from non-converters [125]. Such findings give us good expectations also in other pathological conditions.

We need to highlight that the innovative and novel techniques described in the previous paragraphs have proven to be not just academic exercises, but valuable tools in clinical practice. However, their greatest limitation lies in the complexity of implementation, which can be overcome with the ever-closer collaboration between physician and technical professional figures.

7. Conclusions

The present review depicts an up-to-date overview of fMRI connectomic applications in the wide spectrum of neurodegenerative conditions. Although some of these techniques are still in development, a relevant body of studies have already shown that brain network analysis provides strong evidence supporting a network-based progression of pathological protein aggregates in these conditions, therefore helping to elucidate the pathophysiology of neurodegeneration in these proteinopathies, which is still largely unknown. The consequent implications for allowing an earlier diagnosis, tracking pathological progression, predicting clinical evolution and providing quantitative measures of treatment response using a non-invasive, easily available tool such as MRI make further research in this field a priority, also in the view of the upcoming development of novel, effective treatment options. As the main limitation of these techniques lies in the complexity of fMRI image post-processing and lack of standardized protocols, future efforts should also be made to overcome such shortcomings.

8. Expert opinion

Considering the great pathological complexity and phenotypical heterogeneity of neurodegenerative diseases, the identification of a common framework to diagnose, monitor and elaborate accurate prognostic models in this field of clinical neurology is definitely a challenge. Structural MRI is a commonly available, non-invasive tool to identify patterns of atrophy suggestive of ongoing neurodegeneration, but its sensitivity to subtle alterations may not be as good as what provided by techniques exploring the functional rearrangements preceding neuronal death. Although resting-state functional MRI has been used for decades in the study of neurological conditions, the translation of mathematical concepts from graph theory to this kind of data has allowed to develop the novel concept of the functional brain connectome, which describes and maps the brain networks involved in the complex functions that are typically affected in each neurodegenerative disease. In this review, we have shown that graph analysis is able to describe the complexity of brain architectural rearrangements due to the accumulation of toxic protein aggregates in these diseases, highlighting the ever-increasing corpus of evidence supporting the network-based hypothesis as a unifying pathogenetic mechanism. Ideally, once a – more or less – stereotypical pattern of pathological propagation from each disease epicenter through the brain functional networks is elucidated by researchers, clinicians will be able to provide quantitative measures useful for patient staging and monitoring (see Paragraph 3), prognostic stratification (Paragraph 4), and evaluation of response to treatment (Paragraph 5), possibly aided by machine learning algorithms to enable the application of these techniques at a single patient level (as mentioned in Paragraph 6). However, the translation of this ideal road map to reality needs to overcome some important shortcomings, considering the inter-individual heterogeneity of pathology distribution and the complexity of implementation of these techniques into the everyday practice.

First of all, within each neurodegenerative syndrome, it is reasonable to expect that disease epicenters may not be the same across all patients, but instead vary according to individual factors, (genetic, epigenetic, neurodevelopmental, etc.). For this reason, some recent studies have suggested that disease epicenters should be determined at the individual level [72,73], in order to allow tailored monitoring and predictions of longitudinal pathological spreading. Of note, individual predictions based on connectivity patterns alone were less accurate in some subgroups with genetic background [73], suggesting that a ‘prion-like’ transneuronal spreading model might not account for all the interindividual heterogeneity in disease evolution. As a matter of fact, several studies in this field have considered a combination of other possible models [16,21,63], which might contribute to network-based degeneration, including the ‘nodal stress’ hypothesis (i.e. degeneration induced by maladaptive neural activity), a ‘shared vulnerability’ of networks induced by a common protein expression signature across neuronal populations, or a loss of trophic factors from disconnected regions. Of course, it also expected that different models may apply to different proteinopathies, even among different tauopathies such as AD and progressive supranuclear palsy [63].

In addition to interindividual heterogeneity, another relevant difficulty to be considered is the complexity of mathematical computing needed to implement connectomic techniques, with the consequent ‘disconnect’ between technical professional figures who elaborate computational algorithms and physician neurologists who can evaluate the clinical rationale and practical applications of research studies into the clinical practice. However, this pitfall can be easily overcome with the ever-closer collaboration within multidisciplinary teams of the neuroimaging research field.

To our knowledge, none of the studies currently available in the literature has performed a direct comparison of connectome-based and other neuroimaging measures in the context of diagnostic/prognostic implications in neurodegenerative diseases, nor has specifically suggested a single tool to be implemented for imminent clinical translation. It is imperative to bridge such knowledge gaps in the near future.

Article highlights

• The identification of rearrangements in vulnerable networks and global brain architecture through resting-state brain functional MRI holds the promise to suggest early biomarkers of aberrant protein accumulation, possibly before recurring to second-level, more invasive examinations such as PET imaging or CSF analysis.

• Functional graph-based analyses have also been used to describe the progressive changes in network organization, shedding light on the different hypotheses of disease progression through brain networks.

• Recent studies have applied the paradigm of the brain functional connectome to prove the prognostic implications of the network-based degeneration hypothesis, through the demonstration of a close correlation between the healthy brain functional architecture and the subsequent progression of neurodegeneration and protein accumulation.

• As shown mostly in patients with PD, functional connectomics may be used to evaluate and measure the response to treatment in terms of adjustments of the brain functional architecture.

• Advanced methodologies have proven to be valuable tools to better model the brain functional architecture and to characterize altered functional connectome properties in presence of neurodegenerative diseases, although the involvement of multidisciplinary professional figures is needed to overcome limitation deriving from the complexity of their implementation in the clinical practice.

Declaration of interest

M Filippi is the Editor-in-Chief of the Journal of Neurology as well an Associate Editor for Human Brain Mapping, Neurological Sciences, and Radiology. He has also received compensation for consulting services from Alexion, Almirall, Biogen, Merck and Co, Novartis, Roche and Sanofi and has performed speaking activities from Bayer, Biogen, Celgene, Chiesi Italia SpA, Eli Lilly and Company, Genzyme, Janssen Pharmaceuticals, Merck Serono, Neopharmed Gentili S.p.A, Novartis, Novo Nordisk, Roche, Sanofi, Takeda, and TEVA. M Filippi has also participated on advisory boards for Alexion, Biogen, Bristol-Myers Squibb, Merck and Co, Novartis, Roche, Sanofi, Sanofi, Sanofi-Genzyme and Takeda and provided scientific direction of educational events for Biogen, Merck, Roche, Celgene, Bristol-Myers Squibb, Eli Lilly, Novartis and Sanofi-Genzyme. Finally, he also receives research support from Biogen Idec, Merck Serono, Novartis, Roche, the Italian Ministry of Health, the Fondazione Italiana Sclerosi Multipla, and the Italian Research Foundation for Amyotrophic Lateral Sclerosis (ALS). Meanwhile, F Agosta is Associate Editor of NeuroImage: Clinical and has received speaker honoraria from Biogen Idec, Roche and Zambon. Furthermore, F Agosta receives or has received research supports from the Italian Ministry of Health, the Italian Research Foundation for Amyotrophic Lateral Sclerosis, the European Research Council and the Foundation Research on Alzheimer Disease. The authors have no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed.

Reviewer disclosures

Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Additional information

Funding

This manuscript was funded by the European Research Council (StG-2016_714388_NeuroTRACK) as well as by the Alzheimer Research Foundation.