Disruption of the white matter structural network and its correlation with baseline progression rate in patients with sporadic amyotrophic lateral sclerosis

Seventy-three patients with ALS and 100 age- and sex-matched HCs were included in this study. All patients with ALS fulfilled the El Escorial revised criteria of the World Federation of Neurology for definite or probable ALS and none of them had affected family members. The severity of the disease was evaluated by the ALS functional rating scale-revised (ALSFRS-R). The exclusion criteria for patients were: (1) severe dysarthria and hand weakness; (2) meeting the criteria of ALS-frontotemporal dementia; and (3) a history of other neurologic conditions that could affect the assessment. The cognitive abilities of the ALS patients were estimated using the Addenbrooke’s Cognitive Examination–revised, Chinese Version. No patient showed significant cognitive impairment according to our prior studies [33, 34] at the time of their neuropsychological assessment. The ALSFRS-R assessed three latent domains corresponding to bulbar, motor and respiratory functions [35, 36], defined as follows: bulbar score = sum of ALSFRS-R questions 1–3 (maximum score 12); motor score = sum of ALSFRS-R questions 4–9 (maximum score 24); respiratory score = sum of ALSFRS-R questions 10–12 (maximum score 12) [35]. The baseline ALS progression rate was calculated as (48 – ALSFRS-R)/time since disease onset [37]. The disease onset time was obtained based on patients’ recall and cross-checked by at least one of the close family members or against medical records if available. The anxiety and depression level of patients was measured by the Hamilton Anxiety and Depression Rating Scale. Healthy participants were recruited from local community through poster advertisements. The exclusion criteria for all participants were: (1) presence of focal brain lesions on routine MRI; (2) claustrophobia or standard MRI incompatibility; (3) history of alcohol/substance abuse; (4) comorbidity with neurological or psychiatric disorders or serious physical disease (including traumatic brain injury, cerebrovascular disease, hypertension, diabetes mellitus, ischemic heart disease, chronic liver disease, or other chronic systemic disorders); and (5) poor image quality or severe head motion via visual inspection. This study was approved by the Human Research Ethics Committee of West China Hospital and written informed consent was obtained from all participants.

All participants were scanned using the same magnetic resonance scanner (3.0 T Siemens Trio, Erlangen, Germany) with a 12-channel head coil. Head motion was minimized by foam padding. DTI images were acquired using a spin-echo echo-planar sequence with the following parameters: 64 noncollinear diffusion directions with b = 1000 s/mm² and a reference image without diffusion weighting (b value = 0), 3 mm slice thickness with no interslice gap, repetition/echo time (TR/TE) 6800/91 ms, field of view 1920 × 1920 mm², flip angle 90°, voxel 0.94 × 0.94 × 3.0 mm³ and 2 excitations. High-resolution 3D T1-weighted images were acquired using a magnetization-prepared rapid gradient-echo sequence with the following parameters: resolution 1.0 mm × 1.0 mm × 1.0 mm, TR/TE 1900/2.26 ms, inversion time 900 ms; flip angle 9°, FOV 256 × 256 mm², matrix size 256 × 256, slice thickness 1 mm, no interslice gap, voxel 1 × 1 × 1 mm³ and 176 slices.

Data preprocessing and WM network construction were conducted mainly using the PANDA software (http://​www.​nitrc.​org/​projects/​panda/​; a pipeline tool for diffusion MRI analysis) [38]. The patient and control samples did not differ in scanning head motion for rotation, transition, and frame-wise displacement (all P > 0.05, Additional file 1: Table S1).

Whole-brain anatomical networks were constructed according to the approach used previously [39]. First, to define the nodes of the network, we used the automated anatomical labeling (AAL) atlas to divide the whole brain into 90 cortical and subcortical regions [40], as discussed and used previously [39, 41]. The connections between each pair of brain anatomical regions were determined by the FA value, which resulted in a 90 × 90 matrix for each participant. More details regarding the image processing and network construction work can be found in Additional file 2.

We applied a network sparsity parameter, S, to give each network the same number of edges. According to previous studies [42, 43], we selected a range of S thresholds for the WM connectivity network such that: (1) the averaged degree over all nodes of each thresholded network was > 2 × log 90; and (2) the small‐worldness σ of all thresholded networks was > 1.1 in all participants. Based on these criteria, we defined S ranging from 0.1 to 0.34. For each network, the area under the curve (AUC), calculated over the range of S values with an interval step of 0.01, provides a summarized scalar for the topological characterization of brain networks unbiased by any single threshold.

Graph theoretical analysis was carried out on each participant’s WM network using the GRETNA software (http://​www.​nitrc.​org/​projects/​gretna/​) [44]. Both global (clustering coefficient C_p, characteristic path length L_p, normalized clustering coefficient γ, normalized characteristic path length λ, small‐worldness σ, local efficiency E_loc and global efficiency E_glob) and regional metrics (nodal degree, betweenness and efficiency) were used to characterize network topology. For global measures, high values of C_p, γ, and E_loc reflect network segregation, i.e., the ability for specialized neuronal processing carried out among densely interconnected regions; while low values of L_p, λ, and high E_glob reflect network integration, i.e., the ability for global information communication or distributed network integration; σ characterizes an optimized balance between network segregation and integration [45, 46]. For nodal measures, the three kinds of nodal centrality measurements can reflect the topological importance of nodes in the network in different ways. More detailed explanation of topological measures can be found in the Additional file 2.

To detect the altered connectivity networks in patients with ALS, we used an NBS approach (http://​www.​nitrc.​org/​projects/​nbs/​) [47] to define a set of supra-threshold links in which any of the connected components and their sizes could be determined (threshold, t = 2.62, P < 0.05 equal to Cohen’s d = 0.2). The significance of each supra-threshold link among the connected components was estimated using a nonparametric permutation method (10,000 permutations).

The demographic characteristics of the ALS and HC groups were compared using the R software (version 4.0.0). A nonparametric permutation test (repeated 10,000 times) was used to analyze between‐group differences in the AUC of global and nodal network metrics. After balancing statistical power against the risk of type I error, for all the nodal metrics, only the nodal centralities that changed in the same direction in at least two out of three different measures were reported [48]. After identification of the between-group differences in global and regional network metrics, partial Pearson’s correlation analyses were performed to assess their relationship with symptom severity including ALSFRS-R and its subscores, depression level and anxiety level, controlling for age, sex and illness duration.

We further studied whether the single-subject WM network of patients with ALS can be used to predict their baseline progression rate. To increase the robustness of the prediction, the progression rate was binarized to “fast progression” and “slow progression” using a cut-off value of 0.68 per month based on previous studies [49, 50]. To reduce the feature dimension, we converted the raw connection data into principal components (PCs) using principal component analysis (PCA). We fed the PCs which explained 80% variance of the connection data into the linear kernel support vector machine (SVM) algorithm (Additional file 1: Fig. S1). To get the unbiased classification accuracy estimation and tune the hyperparameter C for the SVM, nested cross-validation was used (details found in Additional file 2).

To assess the significance of the prediction and to make sure that the proposed results did not reflect overfitting, we re-ran the study on a randomly permuted dataset. To do this, we shuffled the progression rate tags (slow and fast progression), breaking the relationship between ALS progression and MRI data and re-ran the analysis. This process was repeated for 5000 iterations and thus quantified the ability of the model to predict noise.

Finally, the top 10 PCs with highest weights in the SVM model were then mapped back from PCA space to WM connectivity space to identify the most important brain connections for ALS progression prediction. The predictor importance score for connections was defined as the product of the absolute value of the weight of the PC in SVM model and principal component scores of the connections [51].

Demographic and clinical characteristics of the participants are summarized in Table 1. There were no significant differences between the two groups in age, sex or years of education (P > 0.05).

Both the ALS and the HC groups exhibited small-world properties of the WM structural network architecture, with γ > 1 and λ ≈ 1. The patients with ALS showed decreased C_p (P = 0.0034, t = 2.98), γ (P = 0.039, t = 2.08) and σ (P = 0.038, t = 2.10) compared with HC (Fig. 1). These differences were still significant after taking account of the outliers.

The nodal topological centralities were decreased in patients with ALS compared with the HCs in the right medial orbital frontal cortex, the left medial superior frontal cortex, the right gyrus rectus, the right paracentral lobule, the right inferior parietal cortex, the bilateral superior temporal pole, the left amygdala and the right caudate (P < 0.05, with significant change in the same direction in at least two of three centrality measures) (Fig. 2, Additional file 1: Table S2). Comparing the connections measured by FA values, we found a network less connected in ALS than in HC, with 6 nodes and 5 edges after NBS correction (Fig. 2).

Although the head motion parameters between patients and HCs were not significantly different (all P > 0.05, Additional file 1: Table S1), to further minimize the impacts of head motion on the primary results, we also re-ran the between-group comparisons taking all head motion parameters (including 3 translation, 3 rotation and 2 frame-wise displacement measures) as covariates. As expected, all between-group comparisons led to the same results as that obtained before.

Exploratory partial Pearson’s correlation analyses showed that the small-worldness indices were positively correlated with ALSFRS (r = 0.27, P = 0.017, Fig. 3a); the degree centrality of the right medial orbital frontal cortex was negatively correlated with depression symptoms (r = − 0.30, P = 0.011) (Fig. 3b) and the nodal efficiency of the right paracentral lobule was positively correlated with ALSFRS (r = 0.28, P = 0.015) (Fig. 3c). After excluding outliers, the correlation between the degree centrality of the right medial orbital frontal cortex and depression symptoms did not reach statistical significance (r = − 0.24, P = 0.051), but all other correlation results remained significant. We also found correlation relationships between network metrics and ALSFRS-R subscores (Table 2).

The demographic and head motion variables were comparable between the subgroups of fast and slow progression rate (Additional file 1: Table S3). Using WM matrices, the mean balanced classification accuracy for predicting ALS baseline progression rate was 85%, which is well above the chance expectation using the same model (P < 0.05, Additional file 1: Fig. S2). We calculated the predictive importance score for each connection in the WM network. The 50 most relevant WM connections contributing to the SVM classification are shown in Fig. 4 and Table 3.