Accurate support vector machine identification of limb-onset amyotrophic lateral sclerosis using resting-state activity of regions within executive control network

Participants

This study investigated fourteen participants (n = 14, two female) with clinically possible, probably or definite limb-onset ALS as defined by the El Escorial diagnostic criteria (14) and eleven (n = 11, two female) gender matched control subjects who were screened for neurological disorders (Table 1). All participants recruited were over 18 years of age and able to give informed consent (Project Number: CF14/3968-2014002057).

MRI data acquisition

MR images were acquired using a 3-Tesla Skyra scanner (Siemens, Erlangen, Germany) at Monash Biomedical Imaging, Melbourne, Australia. Participants were instructed to fixate their gaze on a cross in the monitor for the duration of the scan. Functional MRI data consisted of 400 volumes of an echo planar sequence (TR = 801, TE = 21 ms, axial slices = 45, matrix= 64 × 64, voxel resolution = 3 × 3 × 3 mm, slice thickness 3 mm, interslice gap = 0mm, 50° flip angle). T1-weighted high-resolution MPRAGE images were acquired for co-registration of the functional datasets (TR = 2300, TE = 2.07 ms, voxel resolution 1 × 1 × 1 mm, field of view = 240 × 256 mm, 9° flip angle, 192 slices per volume).

MRI pre-processing

Pre-processing steps were carried using FEAT (Version 6.00) a part of the FSL toolbox (www.fmrib.ox.ac.uk/fsl). The following pre-processing steps were applied to functional data: (i) motion correction through multi-resolution rigid body co-registration using MCFLIRT (15); (ii) removal of nonbrain structures from volumes using BET (16); (iii) correction for slice timing; (iv) spatial smoothing with a 5 mm full-width half-maximum Gaussian kernel to reduce individual variability (17); (v) bias removal by normalizing each volume to its grand-mean intensity; (vi) correction for distortion with gradient-echo field map; and (vii) highpass temporal filtering with 100 second cut-off. Structural images were brain-extracted using BET and visually inspected. Registration of functional data to T1-weighted and brain extracted Montreal Neurological Institute (MNI152) 1 mm standard images was carried out using Boundary-Based Registration and 12-DOF affine transformation using FLIRT (15,18) followed by FNIRT nonlinear registration (19,20).

Independent component analysis

MELODIC (version 3.15) part of FSL was used to perform probabilistic independent component analysis (PICA) (21) for each subject where model order was estimated using the Laplace approximation (range 35-76) (22). FSL AROMA was used to automatically identify and remove motion related components from the data (23,24) with the remaining components transformed to MNI152 1mm standard space (25).

Control group data was temporally concatenated using group PICA with model order of 25 (21). For each resulting component, the component with the highest cross-correlation value to each RSN was determined to represent that network (26). These group spatial maps were then regressed into each individual dataset using dual regression to derive individual spatial maps and corresponding time courses (27,28). Each voxel contains a β-coefficient representing the estimated contribution of that voxel to the regression model (27).

Comparison of resting-state networks

Between-group comparisons of individual spatial maps were performed using non-parametric permutation tests using FSL’s Randomise with 5,000 permutations (29). Results were corrected for multiple comparisons using threshold-free cluster enhancement (TFCE) (30) with P values less than 0.05 considered to be statistically significant. Statistically significant clusters were masked for non-brain regions. To calculate the number of voxels of clusters belonging to anatomical regions, binary masks of each anatomical region within each cluster were created and subsequently used to mask the clusters to produce significant voxels parcellated by each anatomical region, with the number of voxels within these parcellated maps calculated.

Support vector machine classification of limb-onset ALS

To classify limb-onset ALS patients from controls, a mask of significant clusters from between-group analysis was created for each subject from individual spatial maps, with ß-coefficients of these regions averaged. Mean ß-coefficients of significant clusters as well as each of the ten RSNs, and class (0 = control, 1 = ALS patient) for each subject was used as input to train a SVM classifier using the scikit-learn library (https://scikit-learn.org/stable/) implemented in Python 3.8.5.

Data was randomly divided into training (70%) and validation datasets (30%) and stratified by group to ensure an equal division of class in each dataset. Each SVM was trained using a linear kernel, C-parameter value of 1, and default value of γ where γ = where is the training data ß-coefficient variance. After training, the SVM returned the probability of each subject in the validation dataset belonging to the limb-onset ALS group. Ten-fold cross-validation was performed using random splits of the data. Resulting probabilities and actual class from each of the ten splits were concatenated to generate receiver-operator-characteristic (ROC) curves and area-under-curve (AUC) values.

Regions of the executive network showed significantly decreased activity in limb-onset ALS patients

Comparisons between groups showed significantly decreased activity within a region within the executive control network (EXN) in limb-onset ALS subjects (P < 0.05 after TFCE; Figure 2). All other networks were insignificantly different, however the occipital pole and frontoparietal networks were nearing significance (Supplementary Table 1).

• Download figure
• Open in new tab

Figure 2. Regions of the EXN showing significantly decreased activity in ALS patients compared to controls.

In orange are significant P values after TFCE (α = 0.05) overlaid on (A) axial, (B) coronal, (C) sagittal, and (D) 3-d render views of the MNI152 standard brain. EXN = executive control network, TFCE = threshold-free cluster enhancement.

Overall, there were 2,671 voxels within this EXN cluster, which localized primarily to the anterior cingulate gyrus (2,134 voxels – 79.90% of total), with smaller clusters in the frontal pole (533 – 19.96%) and paracingulate gyrus (4 – 0.15%) (Table 2).

Accurate SVM classification of limb-onset ALS patients using activity within EXN cluster

Mean ß-coefficients from individual spatial maps were calculated for each of the RSNs as well as the significant EXN cluster, which were used to train a SVM classifier evaluated with ten-fold cross-validation. The SVM trained using mean EXN cluster ß-coefficients showed excellent performance on both the training (AUC = 0.93) and validation (AUC =0.93) datasets (Figure 3A & B, respectively). The optimal threshold of 0.65 which returned a sensitivity of 86% with 87% specificity (Supplementary Table 2). Excluding the entire EXN on the training data with an AUC of 0.71, all other networks displayed near random performance on both datasets (AUC range: 0.34 – 0.62).

• Download figure
• Open in new tab

Figure 3. Accurate SVM classification of limb-onset ALS patients using EXN cluster.

Shown are ROC curves for SVMs trained using β-coefficients of each RSN (dotted lines) and EXN cluster (orange solid line) on (A) training and (B) validation datasets. Solid navy line represents random performance. The EXN cluster SVM excellent performance on both the training (AUC = 0.93) and validation (AUC = 0.93) datasets. SVM = support vector machine, ROC = receiver operator characteristic, AUC = area-under-curve.

Decreased activity in executive control network of limb-onset ALS patients

Between-group analysis showed a region of the EXN significantly decreased in limb-onset ALS patients localizing to the anterior cingulate gyrus (ACG), frontal pole, and paracingulate gyrus. It is worth noting that differences in the frontoparietal networks (FTPNs) and occipital pole network were nearing significance, which may not have been detected due to sample size, which was less than half the recommended size to detect moderate to strong effects in neuroimaging studies (31,32). This trend was mentioned as several studies have shown decreased activity in temporo-occipital regions of several networks – including the FTPN and visual networks – in heavily (>80%) limb-onset ALS cohorts (13,33,34). Based on this observation, it should be explored if temporo-occipital decreases are a hallmark of this subtype.

Strengthening the involvement of these regions in this subtype are longitudinal studies in heavily limb-onset ALS cohorts which describe progressive increases over time (35–37). Given the suggestion of decreased activity compared to controls (13,33,34), it is unclear whether these progressive increases are a compensatory mechanism following these initial decreases, or reflect interneuron degeneration causing decreased inhibition, which has been demonstrated in mouse models (38,39) and post-mortem brain tissue (40). The neurodegenerative mechanisms of ALS and its implications on functional activity, particularly within temporo-occipital regions of the brain, should be a focus of future work.

Limitations

This study had several limitations, namely sample size, which may contribute to the inability to find significant effects, given studies investigating effect-sizes suggest each group should have 40 subjects to detect strong effects, and over 80 to detect moderate effects (31,32). Furthermore, an increased sample size would increase validation dataset size for the SVM. Although this was mitigated through cross-validation, increased validation cohorts would improve the reliability of our findings.

Secondly, our cohort was very skewed, with 12 of the 14 limb-onset ALS patients being male. As RSN sex-differences have been reported in ALS (41) this is both a strength, as it improves the specificity of our cohort, but also a limitation, as it decreases the translatability of findings to females. Finally, our cerebellum RSN map showed poor similarity to previously described RSNs (26) and this may affect detecting effects in this region.