Introduction
Amyotrophic lateral sclerosis (ALS) is an intractable progressive neurodegenerative disease characterized classically by neurodegeneration and loss of upper motor neurons of the corticospinal tract and lower motor neurons of the brainstem and spinal cord anterior horns [
1]. While symptoms such as muscular weakness, spasticity and hyperreflexia may initially be manageable, the progressive loss of respiratory muscle innervation can lead to respiratory failure, typically within 2–4 years of symptom onset [
2]. There is increasing evidence that ALS also affects multiple neural systems beyond the motor cortex and corticospinal tract [
3] and there is an urgent need to identify reliable biomarkers for ALS progression in clinical practice and pharmacological trials [
4].
Typically, decreased fractional anisotropy (FA) within focal brain regions including motor, frontal and prefrontal areas are found in white matter (WM) studies in ALS using diffusion tensor imaging (DTI) tractography [
5] and tract‐based spatial statistics [
6‐
8]. DTI metrics are sensitive markers for WM change [
9], and have been recommended for ALS diagnosis [
9] and assessment of disease progression [
10,
11]. However, the direct correlations between focal magnetic resonance imaging (MRI) metrics and neuropsychological measures are questionable, because the motor, cognitive and behavioral functions are mediated by multisynaptic brain networks [
12]. Therefore, the notion of selective anatomic vulnerability [
13] is being supplemented and to some extent replaced by the syndrome-specific network vulnerability notion [
14], which is supported by concepts such as network-wise degeneration [
15], circuit-specific vulnerability [
16] and disease progression along structural connectivity patterns [
17]. DTI studies based on graph theory offer a valuable tool to analyze the topological organization of brain networks and inter-regional connections [
18], which may be indicators of ALS progression [
19]. In the ‘connectomics’ analysis, cortical and subcortical brain regions are parcellated into nodes, and the WM metrics of tracks between them taken as the edges of a mathematical graph. Such studies show that the brain networks have a “small-world” organization [
18], intermediate between random networks, whose shorter overall path length is associated with a low level of local clustering, and regular networks or lattices, whose high level of clustering is accompanied by a long path length [
18]. The small-world network architecture reconciles relatively independent functioning (i.e. segregation) with fast information transfer (i.e. integration) [
20]. Application of this powerful approach to the brain structural and functional connectome in ALS has, however, yielded inconsistent results [
19,
21‐
23].
Incorporating the interaction information across the whole brain, the single-subject network approach has value in characterizing objective brain features to distinguish patients from healthy individuals [
24] as well as for predicting clinical outcomes after drug treatment [
25]. Several clinical prognostic factors have been identified for ALS, including age, site of onset, functional and respiratory status, cognitive function, noninvasive ventilation, some genetic mutations [
26], and clinical phenotypes [
27]. In addition, there are some biological markers proposed as related to the ALS outcome, including dyslipidemia [
28,
29], uric acid [
30,
31], creatinine, albumin [
4], and granulocyte count [
32]. However, it is still unclear whether the WM network parameters have predictive value for the ALS baseline progression rate.
In the present study, we applied graph theoretical analysis to DTI data to compare the topological properties of brain WM networks between ALS patients and healthy controls (HC) at the global, regional and connection levels, and also evaluated the predictive value of the DTI-based connectome for the baseline ALS progression.
Discussion
In the present study, we found significant changes in the topological architecture of brain structural network at different levels. At the global level, the whole‐brain WM network showed decreased small-worldness in patients with ALS, reflected by lower σ, and decreased segregation reflected by lower Cp and λ. At the regional level, several nodes located mainly in the frontal, temporal and subcortical regions showed decreased topological centralities in patients with ALS. At the connection level, we found decreased WM connections between the nodes with decreased centralities. In addition, the machine leaning model showed that the single-subject structural connection network can be used as a biomarker to predict the ALS progression rate, which may inspire further clinical practice.
The WM structural networks in patients with ALS showed weaker small-worldization, evidenced by decreased clustering coefficients and small-worldness index. The small‐world organization reflects an optimal balance between network segregation (reflected by
Cp,
γ, or
Eloc) and network integration (reflected by
Lp,
λ, or
Eglob) of information processing [
20], and the balance can be measured as
σ [
52]. Despite having an overall small‐world architecture qualitatively similar to HC, patients with ALS showed lower
Cp and
γ, resulting in a lower small-worldness index
σ. These alterations of small-worldness were positively correlated with ALSFRS, suggesting a clinical relevance.
These results are consistent with a recent multicenter study that reported altered global structural brain network properties in patients with ALS [
53]. In contrast, several early studies on WM connectomics have reported no global topological alterations in patients with ALS [
21,
22,
54]. Although these studies utilized similar MRI sequences and tracking methods, their sensitivity to these topological changes may have been limited by relatively small sample sizes and/or relatively low numbers of non-collinear diffusion directions in the DTI sequence. Zhang et al. have reported that patients with ALS show a consistent rearrangement towards a regularized architecture evidenced by increased path length and clustering coefficient [
23]. This difference may result from different network definitions, as they used the structural covariance networks, in which the connections were defined by the Pearson correlation coefficients between two regions of interest in gray matter. However, rearrangement toward a regularized network itself reflects a breakdown of the original optimal small-world network architecture. No doubtfully, different modalities can provide different perspectives on network abnormalities. In future connectome studies on ALS, different modalities can be combined to include functional MRI, diffusion MRI and gray matter MRI in larger sample sizes.
In addition to the global topological abnormalities, we found topological alterations in several brain regions. Consistent with the recent multicenter study [
53], we found decreased nodal centralities in ALS patients in both motor and nonmotor networks including the secondary motor regions, the prefrontal regions, the temporal regions (superior temporal pole), the basal ganglia regions and the parietal region.
The cortical motor system is a distributed network of areas involved in different aspects of specific motor execution. Even simple movements are associated with activation of multiple cortical areas of the primary motor cortex and the secondary motor regions [
55‐
57], including the supplementary cortex, the premotor cortex, the paracentral lobule and the superior parietal motor areas, which are highly inter-connected, converging on the primary motor cortex [
58]. Our results suggest that deficits of the secondary motor regions may be an important trait in ALS. Consistent with this, earlier neuroimaging studies have reported decreased cortical thickness and gray matter volume of the secondary motor regions in patients with ALS [
59‐
61].
The pathological hallmarks of ALS are tau-negative and ubiquitin-positive intraneuronal inclusions, and the 43-kDa TAR DNA-binding protein (pTDP-43) is a major component of the inclusions specific for frontotemporal lobar degeneration and ALS [
62]. Initially, the TDP-43 burden is greatest in the agranular motor cortex and brainstem motor nuclei [
63]. As the disease progresses, the pTDP-43 lesions increasingly involve the prefrontal (gyrus rectus and orbital gyri), striatum, amygdala and temporal lobe along axonal pathways [
63,
64]. Consistently, neuroimaging studies in ALS have also confirmed the spread of atrophy and/or hypometabolism to the frontal and temporal cortices [
65‐
67]. DTI studies have also reported WM deficits in the frontotemporal regions [
68,
69]. Longitudinal and combined structural and functional MRI studies are needed to validate our hypothesis of disease progression along the functional and structural connections of the frontotemporal network.
We also found that the single-subject networks can predict the disease progression rate with an accuracy of 85%. Earlier studies also found that MRI abnormalities can be used to predict outcome in ALS. More severe abnormalities of the corticospinal tract and the spinal cord predict a poorer long-term clinical outcome in patients with ALS [
70,
71]. FA has been proven to be a sensitive DTI metric for both diagnosis [
72] and progression modeling [
73]. An earlier WM network study has also found a relationship between the FA-based connectivity degree in the frontal area and disease progression rate of patients with ALS [
19]. Similarly, using deep learning, van der Burgh and colleagues also predicted outcomes of ALS patients with high accuracy by combining the WM network, morphology and clinical information [
74]. As an important alternative approach to studying ALS pathology progression, earlier studies [
75‐
78] also found that alterations of network and other imaging features provide useful information associated with disease progression [
79] in the spatial domain. All this evidence indicates that the brain network information has significant predictive potential to predict disease progression in patients with ALS as a supplement to other clinical measures.
The study has several limitations. First, although the method for echo plane imaging-distortion correction (i.e. non-linear registration) used in the current study is common in the field, state-of-the-art distortion correction methods like file-mapping [
80],
topup-based approach [
81‐
83], or machine learning approaches [
84,
85] are encouraged to be used in future studies. Second, currently there is no widely accepted optimal approach to defining nodes and edges. We used the widely used AAL 90 template regions as nodes and mean FA values of fibers as the weighting factor in the construction of graphs. Other measures such as Harvard–Oxford atlas can also be considered for calculating network metrics [
38]. Third, this study was cross-sectional; how the WM network architecture associated with ALS evolves dynamically and how the WM network evolves in the progress of disease remain to be clarified in longitudinal studies. Fourth, in this study the progression rate was based on a retrospective interview. As ALS progression is dynamic, a prospective study design would be more suitable for the prediction analysis. Fifth, we did not collect genetic information of the ALS patients. Although genetic factors have less impact on sporadic ALS compared with familial ALS [
86], further studies should also collect the genetic information to study the interactions between MRI and genetic information. Finally, to reduce the scanning time and thus limit obstacles to participation, we chose 3-mm slice thickness DWI. This has led to non-isotropic voxels which might have a negative effect on the FA estimation and tractography. High-resolution diffusion-weighted images with isotropic voxel size would be a better choice in future studies, if scanner timing permits.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.