Magnetoencephalography-based interpretable automated differential diagnosis in neurodegenerative diseases

38 39 Automating the diagnostic process steps has been of interest for research grounds and to help 40 manage the healthcare systems. Improved classification accuracies, provided by ever more 41 sophisticated algorithms, were mirrored by the loss of interpretability on the criteria for 42 achieving accuracy. In other words, the mechanisms responsible for generating the 43 distinguishing features are typically not investigated. Furthermore, the vast majority of the 44 classification studies focus on the classification of one disease as opposed to matched 45 controls. While this scenario has internal validity, concerning the appropriateness toward 46 answering scientific questions, it does not have external validity. In other words, differentiating 47 multiple diseases at once is a classification problem closer to many real -world scenarios. In 48 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint NOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice. this work, we test the hypothesis that specific data features hold most of the discriminative 49 power across multiple neurodegenerative diseases. Furthermore, we perform an explorative 50 analysis to compare metrics based on different assumptions (concerning the underlying 51 mechanisms). To test this hypothesis, we leverage a large Magnetoencephalography dataset 52 (N=109) merging four cohorts, recorded in the same clinical setting, of patients affected by 53 multiple sclerosis, amyotrophic lateral sclerosis, Parkinson’s disease, and mild cognitive 54 impairment. Our results show that it is possible to reach a balanced accuracy of 67,1% (chance 55 level = 35%), based on a small set of (non -disease specific) features. We show that edge 56 metrics (defined as statistical dependencies between pairs of brain signals) perform better 57 than nodal metrics (considering region while disregarding the interactions. Moreover, phase -58 based metrics slightly outperform amplitude-based metrics. In conclusion, our work shows that 59 a small set of phase -based connectivity metrics applied to MEG data successfully 60 distinguishes across multiple neurological diseases. 61 62 63 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint

64 65 In the last twenty-five years, the widespread availability of large-scale brain functional data in 66 health and disease has brought great hope toward discovering the mechanisms underpinning 67 brain diseases and the appearance of neurological symptoms. Cognitive functions emerge 68 from the coordinated interactions among brain regions, manifesting as statistical 69 dependencies among the corresponding brain signals. The overall statistical dependencies 70 between all pairs of signals are often referred to as “functional connectivity” (Friston, 1994). 71 Functional connectivity (FC) is subject -specific and allows subject identification (Finn & 72 Rosenberg, 2021), is altered during the execution of tasks (Corsi et al., 2020) , in different 73 environmental conditions (Shine et al., 2016), as well as in neurological diseases (Sorrentino 74 et al., n.d., 2018, 2019) . The commonest and most straightforward approach to assessing 75 statistical dependencies has been using descriptive metrics (e.g., Pearson’s correlation). This 76 approach has no underlying assumptions concerning the mechanism underlying the observed 77 statistical dependencies. Other techniques take a more mechanisms-driven approach. As an 78 example, the hypothesis of communication through coherence posits that the occurrence of 79 communication between regions might occur via more or less synchronization (Fries, 2015). 80 Then, metrics such as the Phase Locking value (PLV) were developed to quantify 81 communication via the synchronization between brain signals (such as 82 electroencephalography-EEG and magnetoencephalography -MEG) (Bastos & Schoffelen, 83 2016). These metrics have been classically used to characterize multiple neurodegenerative 84 diseases (Stam, 2010). More recently, it was shown that large -scale brain activity is far from 85 stationary, and instead, it is characterized by aperiodic, scale-free bursts of activity (Haldeman 86 & Beggs, 2005; Shriki et al., 2013; Tagliazucchi et al., 2012). Then, borrowing from statistical 87 mechanics, the dependencies among brain regions were understood as the presence of scale-88 free bursts of activities, named “neuronal avalanches”, that describe the presence of aperiodic, 89 non-linear bursts of activities spreading brain regions. Intriguingly, in several neurological 90 diseases (such as Parkinson’s disease, Amyotrophic lateral sclerosis, and Mild Cognitive 91 Impairment), brain dynamics spread differently with respect to healthy controls (Polverino et 92 al., 2024; Romano et al., 2023; Sorrentino et al., 2019), and, more importantly, changes in the 93 way aperiodic waves spread proved to be strongly predictive of individual clinical disability 94 (Polverino et al., 2024; Romano et al., 2023). 95 Despite extensive efforts, there has been a lack of replicability of the studies, regardless of 96 the particular technique adopted to estimate functional connectivity (Kelly & Hoptman, 2022). 97 In other words, the measurements and metrics devised to this day might fail to optimally 98 capture disease -relevant mechanisms comprehensively. As a consequence, automatic 99 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint classification among multiple neurological diseases cannot be achieved with high accuracy 100 based on functional data alone. 101 In this paper, we take a different approach and start from the assumption that the way 102 pathophysiological processes spread across the brain has some aspects to it that are specific 103 to a given disease and can be best measured in a set of features that are (spatially) shared 104 among multiple diseases. As a direct consequence, functional connectivity should show 105 specific elements that distinguish various diseases. Hence, the first hypothesis of our study is 106 that it is possible to identify a (small) set of features that can classify multiple neurological 107 diseases. 108 To test our hypothesis, we leveraged a vast cohort of source -reconstructed MEG data from 109 patients affected by mild cognitive impairment (MCI), amyotrophic lateral sclerosis (ALS), 110 Parkinson’s disease (PD), and Multiple Sclerosis (MS). 111 First, we compared the classification performance of four FC metrics that capture different 112 properties of the signals (AEC, PLV, Pearson’s correlation coefficient, and ATM) associated 113 with a four-class problem (i.e. MCI, PD, MS, and ALS). We considered the features that can 114 differentiate the considered neurological diseases for each FC metric taken separately. 115 Furthermore, we compared nodal and edge metrics, under the hypothesis that edges, which 116 more directly represented the interactions among brain regions, would outperform local (i.e.) 117 metrics. We compared the classification performance using three different machine learning 118 algorithms (i.e., XGBoost, Support Vector Machine (SVM), and Linear Discriminant Analysis 119 (LDA), to demonstrate that the performance of a given feature-set is algorithm-independent. 120 Finally, for each FC metric, we identified the most informative features used by the classifier, 121 under the hypothesis that such relevant features were linked to the neurophysiology of the 122 considered neurological diseases. Such a study would make the classification results more 123 interpretable and would enable us to identify clusters of brain interactions sensitive to the 124 neurophysiological mechanisms associated with the considered diseases. 125 The purpose of this work is to explore a diverse set of connectivity metrics to propose an 126 interpretable automated pipeline for differentiated diagnosis of neurodegenerative diseases. 127 128 1. Materials and Methods 129 2.1 Participants 130 131 One hundred nine patients with different neurological diseases (ALS, MCI, PD, MS) were 132 recruited from Hermitage Capodimonte Clinic in Naples (Polverino et al., 2022; Romano et al., 133 2023; Sorrentino et al., 2019, 2022) . Specifically, Thirty-two MCI patients (18 males and 14 134 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint females; mean age 71.31; SD ± 6.83; mean education 10.54; SD ± 4.33) were recruited from 135 the Center of Cognitive and Memory Disorders of the Hermitage Capodimonte Clinic in 136 Naples, Italy. The MCI diagnosis was done according to the National Institute on Ageing -137 Alzheimer’s Association criteria (Albert et al., 2011). Thirty-nine ALS patients (29 males and 138 10 females; mean age 59.63; SD ± 12.87; mean education 10.38 years SD ± 4.3) were 139 selected in collaboration with the ALS Center of the First Division of Neurology of the University 140 of Campania “Luigi Vanvitelli” (Naples, Italy). The ALS diagnosis was performed according to 141 the El-Escorial criteria (Brooks, 1994). Twenty patients (14 males and 6 females; mean age 142 64.5; SD ± 12.18; mean education 11 years SD ± 3.9) with a confirmed diagnosis of 143 Parkinson’s disease according to the United Kingdom Parkinson’s Disease Brain Bank criteria 144 (Gibb & Lees, 1988) were recruited in collaboration with the Movement Disorder Unit of 145 Cardarelli hospital in Naples. Finally, eighteen patients (6 males and 12 females; mean age 146 45.05; SD ± 9.92; mean education 14 -11 years SD ± 4.89) with Multiple Sclerosis were 147 recruited in collaboration with University of Campania Luigi Vanvitelli. The diagnosis was 148 performed following the 2017 revision of the McDonald criteria (Thompson et al., 2018). Each 149 participant underwent a specific motor and/or neuropsychological evaluation according to the 150 clinical characteristics of each disease. A complete summary of the cohort description is 151 available in Table 1. The study protocol was approved by the ‘‘Comitato Etico Campania 152 Centro’’ (Prot.n.93C.E./Reg. n.14 -17OSS) and all participants provided written informed 153 consent in accordance with the Declaration of Helsinki. 154 155 156 Type of disease Number of participants (109) Age (mean ± SD) Years of education (mean ± SD) Gender (ratio) Mild Cognitive Impairment (MCI) 32 71.31 (SD ± 6.83) 10.54 (SD ± 4.33) 18 m / 14 f Multiple Sclerosis (MS) 18 45.05 (SD ± 9.92) 14.11 (SD ± 4.89) 6m /12 f Parkinson’ s Disease (PD) 20 64.5 (SD ± 12.18) 11 (SD ± 3.9) 14 m / 6 f Amyotrophic Lateral Sclerosis (ALS) 39 59.63 (SD ± 12.87) 10.38 (SD ± 4.3) 29 m/10 f Table 1: Demographic features of the cohort: m: males; f: females; SD: Standard 157 Deviation 158 159 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint 2.2 MEG and MRI acquisition, pre-processing, and source reconstruction 160 161 MEG and MRI acquisition, preprocessing, and source reconstruction were performed similarly 162 to previous studies (Cipriano et al., 2024, p. 20; Romano et al., 2022) . Briefly, all patients 163 underwent an MRI scan using a 3T Biograph mMR tomograph (Siemens HealthcareErlangen, 164 Germany) equipped with a 12 channels head coil. Specifically, 3 dimensional T1 -weighted 165 images (gradient-echo sequence inversion recovery prepared fast spoiled gradient recalled -166 echo, time repetition = 6,988 ms, inversion time = 1,100 ms, echo time = 3.9 ms, flip angle = 167 10, voxel size = 1 × 1 × 1.2 mm3) were acquired. The MEG acquisition was performed using 168 a 163-magnetometer system placed in a magnetically shielded room (AtB Biomag UG, Ulm, 169 Germany). Fastrack (Polhemus®) was used to define the position of the head under the 170 helmet and to digitalize the position of four anatomical landmarks (nasion, right, and left 171 preauricular and apex) and four reference coils. Each patient performed two recordings of 3.5 172 minutes each, with a one -minute break, during a resting state, with eyes closed. 173 Electrocardiographic and electrooculographic signals were recorded to remove physiological 174 artifacts. Data were acquired with a sampling frequency of 1024 Hz. A Principal component 175 analysis (PCA) was used to reduce the environmental noise, and an independent component 176 analysis (ICA) was used to remove physiological artifacts (namely ocular and cardiac 177 artifacts). Finally, to obtain the source-reconstructed time series of the patients, according to 178 the Automated Anatomical Labeling (AAL) atlas, we used a beamformer algorithm and the 179 volume conduction model proposed by Nolte (Nolte, 2003) . The time series were filtered 180 between 0.5 and 48 Hz. 181 2.3 Connectivity Metrics 182 183 Phase Locking Value (PLV) 184 185 The PLV measures the phase synchronization between two narrowband signals, and it is 186 computed as : (Lachaux et al., 1999). 187 𝑃𝐿𝑉 = |𝐸 [ 𝑒⬚𝑗𝛥𝛷𝑥𝑦(𝑡)]|, 188 189 where ∆ Φxy(t) represents the difference between Φx(t) - ∆Φy(t), [E ] is the statistical 190 expectation, and ∆Φx, y(t) are the instantaneous phases of the analytical signals. 191 192 Correlation Coefficient (CC) 193 194 We computed Pearson's correlation coefficient to estimate the pairwise synchronization 195 between signals of different brain regions. 196 197 Avalanche Transition Matrix (ATM) 198 199 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint The ATM describes the probability that after the activation of region i at the time t , the region 200 j will be active at the time t +δ (Sorrentino et al., 2021). The ATMs are computed starting from 201 neuronal avalanches, which are defined as events that start when at least one region is above 202 the threshold and end when all the regions return to their baseline activity. Hence, there is one 203 ATM for each avalanche. More specifically, the ATM contains, in the ijth position, the 204 probability that region j is active at time t+1 given that region t is active at time t.ATMs were 205 then averaged element-wise over all the avalanches for a subject, and finally symmetrized. 206 207 208 Amplitude Envelope Correlation (AEC) 209 210 The amplitude envelope is used to estimate the statistical interdependencies between brain 211 regions. It is computed as the correlation coefficient between the analytical amplitude of two 212 signals. High values of amplitude correlation between the envelopes indicate that two brain 213 regions display a coordinated behavior (Brookes et al., 2011, 2012). 214 215 Nodal analysis 216 217 Each of the connectivity metrics yields an adjacency matrix. We have compared directly a 218 subset of the entries of the matrices (see section 3.2), that is “edge-metrics” or nodal metrics. 219 Three different edge -specific metrics were used: betweenness centrality, eigenvector 220 centrality, and the degree. Betweenness centrality is a centrality measure that is equal to the 221 number of the shortest paths passing through a given node. Another centrality measure is 222 eigenvector centrality, which determines a node's relative importance within a network. Lastly, 223 the degree of a node is the sum of the weights of the edges incident upon the node. 224 225 2.4 Classification Algorithms 226 227 To evaluate the discriminative ability of different feature sets (PLV, CC, ATM, AEC) and 228 compare them with each other, we applied three different Machine Learning (ML) algorithms. 229 Balanced accuracy was used as an evaluation metric, since we have imbalanced classes. ML 230 algorithms include Linear Discriminant Analysis (LDA), Support Vector Machines (SVM), and 231 Extreme Gradient Boosting (XGBoost). The general modelling workflow is summarized in Fig. 232 1 233 234 Linear Discriminant Analysis 235 236 LDA is a widely used approach for solving multi -class classification problems. The algorithm 237 separates multiple classes (in our study - 4 classes) with multiple features through a data 238 dimensionality reduction approach. LDA aims to find a hyperplane that best separates the 239 classes while minimizing the overlap within each class. Related work has revealed that LDA 240 performs well with multiclass diagnosis problems (Lin et al., 2021). 241 242 243 244 245 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint 246 Figure 1 The general workflow of modeling 247 248 Support Vector Machines 249 250 SVM is another widely used technique for solving supervised tasks with multiple classes. 251 Several studies identified SVM as an outstanding algorithm for solving tasks with multiple 252 classes (Maqsood et al., 2022) . SVM performs complex data transformations (according to 253 the selected kernel function) and maximizes the separation boundaries between the data 254 points depending on the classes. 255 256 Extreme Gradient Boosting 257 258 Recent studies showed that XGBoost is a state-of-the-art tree-based machine learning model 259 that outperforms many other algorithms, including deep learning models (Grinsztajn et al., 260 2022). Moreover, the XGBoost algorithm provides an assessment of the relative importance 261 of individual predictors, which allows us to interpret our findings (Manju et al., n.d.) 262 XGBoost is an ensemble method that builds a predictive model by combining predictions of 263 multiple individual decision trees. It uses weak learner trees, these are decision trees with a 264 single split, called decision stumps. The algorithm works by sequentially adding weak learners 265 to the ensemble, with each new learner focusing on correcting the errors made by the previous 266 one. 267 XGBoost is known for its high accuracy and has been shown to outperform other machine 268 learning algorithms in many predictive modeling tasks. In addition, it is highly scalable and can 269 handle large datasets. 270 271 2.5 Statistical Analysis 272 273 Kruskal-Wallis test 274 275 For each connectivity metric taken separately, to identify the most statistically significant 276 different features among the four groups (PD, MCI, SLA, MS) to be considered for the 277 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint classification, we used the Kruskal-Wallis test. Since the brain is a non-linear dynamic system, 278 we relied on a non -parametric statistical test, checking the null hypothesis that two or more 279 independent groups were drawn from the same underlying distribution. The same approach 280 was used for both edge and nodal metrics. 281 282 Multiple comparison correction 283 284 Since we have numerous features to be considered for a given FC metric, we used the false 285 discovery rate to correct for inflated significance. The False Discovery Rate (FDR) is used to 286 control the expected proportion of false positives. The FDR is the expected ratio of the number 287 of false positive classifications, or false “discoveries”, to the total number of positive 288 classifications (rejections of the null hypothesis). The p-values of the Kruskal-Wallis test were 289 corrected accordingly. Finally, we sorted the features according to the corrected p -values in 290 ascending order. 291 292 Spearman Correlation 293 We used Spearman correlation to evaluate the correlation between the features’ ranks. 294 Spearman’s rank correlation coefficient is a non-parametric measure of statistical dependence 295 between two variables. This way, we evaluated the relation between the ranks of the nodal or 296 edge features across different FC metrics (PLV, AEC, ATM). 297 298 Repeated Stratified K-fold splits 299 300 To get valid results and avoid overfitting, we applied Repeated Stratified K -fold cross -301 validation, which repeats k-folds n times with different randomization for each repetition (J.-H. 302 Kim, 2009). Then, for each fold, we have pooled our results across multiple randomization. 303 First, our whole dataset was split into two parts. For the first part, we use Stratified K -folds 304 cross-validation to tune hyperparameters and find an optimal set that gives the best result. 305 After tuning the hyperparameters on the first part of the dataset, then we used Stratified K-fold 306 cross-validation 10 times for the second part. Then, accuracies obtained by each set are 307 averaged. This way, we prevent data leakage, and it helps to get a more robust estimation of 308 the accuracy by averaging over all repetitions and all folds. We used 10 repetitions of 10-fold 309 cross-validations, and therefore we ensure that our evaluation is not affected by the specific 310 choice of the validation set. 311 2.7 Evaluation metrics 312 313 Balanced accuracy 314 315 We used balanced accuracy as an evaluation metric for the classification algorithms since our 316 dataset is imbalanced (NALS=39, NMCI=32, NPD=20, NMS=18) . The balanced accuracy is 317 calculated by taking the average of the recalls obtained in each class (Thölke et al., 2023). 318 319 320 Recall 321 322 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint Recall is an evaluation metric that measures how often a classification algorithm correctly 323 identifies positive instances among all the actual positive samples in the dataset. 324 325 𝑅𝑒𝑐𝑎𝑙𝑙 = 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 + 𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 326 327 Receiver Operating Characteristic (ROC) curve 328 329 ROC curve is a graph that displays the performance of a binary classification algorithm of 330 predicting a positive class at all possible thresholds. The lower the classification threshold, the 331 more observations are successfully classified. ROC curve uses False Positive Rate on the x-332 axis and True Positive Rate on its y-axis. 333 The area under the ROC curve (AUC) is an evaluation measure that measures the area 334 underneath the ROC curve, and its maximum possible value equals one. In this manuscript, 335 we compute the ROC curve for each class separately. 336 337 Confusion Matrix 338 339 A confusion matrix is an N x N matrix, where N is the number of classes. It has the true labels 340 on the rows and the predicted labels on the second axis. This way, a confusion matrix shows 341 how many times each class was classified correctly and also how often it was misclassified 342 (and how). 343 We used a confusion matrix for 4 classes, therefore we have a 4 x 4 matrix, where 𝑇𝑃𝑖 344 represents the observations that were correctly classified for class 𝑖 , and 𝐸𝑖𝑗 represents where 345 true class 𝑗 was misclassified with predicted class 𝑖. After that, we took the relative 346 percentages across columns to see the whole picture in percentages, therefore each column’s 347 values will sum up to 100%. This is done by dividing each element of each column by the sum 348 of all elements of that column and multiplying by 100. For example, for column 4 and its third 349 element, it is done as follows: 350 𝐸43 (𝐸41 + 𝐸42 + 𝐸43 + 𝑇𝑃4) × 100 351 352 353 354 355 Code availability 356 357 The code used to perform the analysis of this study is publicly available at 358 https://github.com/dklpp/multiclass_meg_features_analysis 359 360 361 2. Results 362 363 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint 3.1 Kruskal-Wallis Test 364 365 Each adjacency matrix obtained from a given FC metric (namely PLV, AEC, ATM, or CC) is 366 a square matrix with the dimension of 𝑛𝑟𝑒𝑔𝑖𝑜𝑛𝑠 × 𝑛𝑟𝑒𝑔𝑖𝑜𝑛𝑠, where 𝑛𝑟𝑒𝑔𝑖𝑜𝑛𝑠is equal to 116 367 regions of interest. All matrices are symmetric and contain ones on the main diagonal., Hence, 368 we take the triangular matrix, excluding the main diagonal elements, leading to 6670 edge -369 wise features. Given the high dimensionality of the feature space, we identified the most 370 statistically significant different features among the four groups to be considered for the 371 classification. 372 A non-parametric statistical Kruskal-Wallis test was performed for each feature to compare 373 the four independent groups (PD, SLA, MS, MCI). After applying Kruskal-Wallis Test and False 374 Discovery Rate correction, we found that there were more than 120 statistically significant 375 edge features (pFDR < 0.002) for each of the 4 edge-specific FC metrics. The lowest corrected 376 with FDR p -value p<0.0001 (pFDR = 2,46 × 10−7 ) was obtained with the edge -wise PLV 377 metric between the right frontal superior gyrus and the right postcentral gyrus (see Fig 2) 378 379 380 Figure 2 The most significant features’ (PLV values between the right frontal superior gyrus and the right postcentral 381 gyrus) boxplots with the observations for 4 classes with FDR p-value (pFDR = 2,46 x 10-7) 382 383 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint 3.2 Classification algorithms 384 385 Based on the significant edges, we have then classified the participants. We used a 386 consecutive iterative search technique, starting with the 15 best features (according to their 387 corrected p-values), and sequentially added features and compared the accuracies, and this 388 procedure was repeated until 39 features were fed to the classifier (Table 2). Stability had 389 been reached at this point, and further increasing the number of features led to a slight 390 worsening of the performance (not shown). Furthermore, given the relatively small size of our 391 sample, we kept a lower number of features to reduce overfitting. 392 393 394 395 Num pred n=15 n=16 n=18 n=22 n=25 n=28 n=30 n=31 n=34 n=36 n=38 n=39 Metrics Edge metrics PLV 0.643 0.629 0.639 0.611 0.647 0.653 0.639 0.631 0.671 0.651 0.627 0.635 AEC 0.568 0.588 0.584 0.605 0.611 0.614 0.613 0.610 0.581 0.604 0.630 0.608 ATM 0.567 0.554 0.520 0.514 0.470 0.515 0.529 0.500 0.504 0.512 0.514 0.517 CC 0.466 0.477 0.521 0.511 0.495 0.539 0.565 0.560 0.556 0.545 0.529 0.525 Nodal metrics AEC (eign. centr.) 0.481 0.478 0.491 0.468 0.468 0.459 0.434 0.492 0.454 0.446 0.418 0.399 AEC (betw. centr.) 0.358 0.345 0.350 0.381 0.362 0.312 0.329 0.334 0.314 0.333 0.341 0.348 AEC (degree) 0.293 0.286 0.296 0.311 0.366 0.313 0.321 0.315 0.337 0.330 0.356 0.348 PLV (degree) 0.407 0.389 0.376 0.345 0.377 0.347 0.321 0.312 0.311 0.302 0.323 0.317 PLV (betw. centr.) 0.436 0.454 0.443 0.392 0.401 0.384 0.371 0.384 0.373 0.372 0.345 0.363 PLV (eign. centr.) 0.350 0.354 0.346 0.384 0.414 0.408 0.387 0.378 0.436 0.443 0.435 0.432 ATM (eign. centr.) 0.479 0.475 0.489 0.475 0.460 0.454 0.438 0.480 0.431 0.427 0.409 0.403 ATM (betw. centr.) 0.358 0.350 0.350 0.381 0.362 0.312 0.329 0.334 0.314 0.333 0.341 0.349 ATM (degree) 0.251 0.251 0.263 0.289 0.302 0.276 0.279 0.269 0.289 0.274 0.308 0.317 CC (betw. centr.) 0.506 0.532 0.512 0.530 0.523 0.486 0.468 0.447 0.452 0.468 0.452 0.453 CC (degree) 0.396 0.389 0.396 0.447 0.446 0.478 0.455 0.443 0.421 0.399 0.421 0.417 396 Table 2:Balanced accuracy for the number of features which contain the best accuracy across different 397 metrics (PLV, AEC, ATM, CC). For visual purposes, it is demonstrated only with the LDA algorithm. The 398 accuracies obtained with the XGBoost and the SVM algorithms are in the supplementary material (see 399 S4 and S5). 400 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint 401 We decided to display the balanced accuracies for each FC metric taken separately (both 402 edge-based and node-based metrics) to see a clearer picture of different sets’ performances 403 (Fig 3). 404 405 406 Figure 3 The balanced accuracies for all feature sets with LDA classifier. Each bar plot displays the averaged 407 accuracy with its standard errors. The boxplots for other algorithms are available in the supplementary material 408 (Fig S1-S2). 409 410 As shown in Fig3, we observed that all edge metrics consistently outperformed nodal metrics. 411 Consistently, we observed that the standard deviations (over different repetitions of the K -412 folds) were higher for nodal metrics. Note that the results refer to the best-performing feature 413 selection (i.e. the number of features is not fixed across different metrics). 414 Finally, we identified the optimal number of features (i.e. the features that showed the lowest 415 corrected p-values and which led to the highest balanced accuracy) for each of the 3 different 416 Machine Learning classification algorithms considered (namely XGBoost, SVM, LDA)(Fig. 3). 417 The exhaustive search algorithm yielded the feature sets (which nodes/ edges) with the best-418 balanced accuracies for each algorithm across different FC metrics. 419 For the sake of simplicity, we discuss here the two best -performing FC metrics per 420 classification algorithm (Fig. 4). In the case of the SVM algorithm, the AEC showed a balanced 421 accuracy of 67.8% with a total of 31 top features, the PLV presented a balanced accuracy of 422 66.5 % with 36 top features. With the XBoost classifier, the CC showed a balanced accuracy 423 of 63.8% with 35 top features), and the PLV presented a balanced accuracy of 62.8%, with 424 15 top features. Finally, in the case of the LDA classifier, the PLV showed a balanced accuracy 425 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint of 67.1%, with 34 top features and the AEC presented a performance of 63.0% with 38 top 426 features. 427 Given the lower performance obtained with the nodal metrics, we shall proceed with the 428 analyses exclusively on the edges. 429 430 Figure 4 Balanced accuracies for different metrics across 3 Machine Learning algorithms with its standard errors. 431 We observe consistent higher performance of PLV and AEC, in comparison to ATM and CC. The chance level for 432 our dataset equals 35%. 433 434 The chance level for the balanced problem with 4 classes is equal to 25%. However, since 435 we have a dataset with unbalanced classes, in such tasks the chance level is usually assumed 436 to be the probability of predicting the most frequent class label in the target.In our case, SLA, 437 which contains 38 patients out of 108, is the most numerous class. Therefore, the chance level 438 is calculated as follows: 439 440 𝑝 = 38 / 108 ≃ 0,35 𝑜𝑟 35% (1) 441 442 In this task, a more objective evaluation metric is the balanced accuracy. Nevertheless, it is 443 also useful to compare and evaluate overall accuracies. The trend remains the same – i.e. the 444 same edge -specific metrics stay as the top features sets. Still, the accuracies are slightly 445 higher: AEC (73.3%, 28 features), PLV (72.7%, 36 features), ATM (67.5%, 32 features), CC 446 (69.1%, 36 features) for SVM classifier; AEC (68.2%, 38 features), PLV (69.5%, 34 features), 447 ATM (63.2%, 15 features), CC (60.4%, 30 features) for LDA classifier; AEC (68.5%, 26 448 features), PLV (68.7%, 15 features), ATM (64.2%, 39 features), CC (70.5%, 35 features) for 449 XGBoost classifier. 450 451 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint Since the PLV is the most performant metric, we now focus on the PLV for sensitivity analyses. 452 453 3.3 ROC curves 454 455 After repetitive stratified K-folds, we can estimate probabilities for each class to be correctly 456 predicted (i.e. the probabilities sum up to 1) with different classification algorithms (LDA, SVM, 457 XGB). We applied the One -vs-All technique, where we fix one desired class and all other 458 classes are treated as one class. This way, we can replace our multi classification task to a 459 binary class, and it enables us to build the ROC curves. 460 We calculated the average ROC curve for each repetition of the 10 folds, and the red curve 461 displays the overall average ROC curve across 10 repetitions. The ROC curves were built for 462 each class separately (Fig 5) which display the trade -off between False Positive Rate on x 463 axis, and True Positive Rate on y axis. 464 ALS patients display the best results in terms of classification accuracy, and PD patients the 465 worst results. Accordingly, it is worth mentioning that deviations of the ROC curves for PD 466 patients are also much higher in comparison to other classes. 467 468 469 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint 470 471 Figure 5 ROC curves for 4 classes (MS, SLA, PD, MCI) for LDA machine learning classifier with overall mean 472 curve and mean curves for each repetition of Stratified K-folds. ROC curves are built with PLV edge-based features 473 with 10 repetitions over 10 k-folds. 474 475 3.4 Confusion Matrix 476 477 The confusion matrix was built to depict the whole picture of the classifier’s performance: it 478 allows seeing what percentage of each class was classified correctly, and where mistakes 479 were made (Fig 6). 480 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint 481 Figure 6 Confusion Matrix with relative percentage representation for 4 classes (SLA, PD, MS, MCI) with true 482 values on y axis and predicted values on x axis for PLV features with LDA machine learning classifier. 483 484 Again, one can observe that the accuracy for PD patients is the worst, while the results for 485 ALS subjects are the best. While these results might be affected by imbalanced classes, the 486

for all classes are well above chance level (which is equal to 35%). 487 488 3.5 Feature importance 489 490 The XGBoost classification algorithm allows us to quantify and compare the relative 491 importance of the features during the classification process. To get valid results, we have 492 defined 10 cross-validations with stratified KFolds. That is, the balance of the classes in the 493 train and test splits are preserved. We run 40 times these stratified cross-validation iterations 494 to reach convergence and then validate our results. After these steps, we obtained the feature 495 importance for each of the features obtained from each FC metric taken separately (we focus 496 here on the PLV, AEC, and ATM). Then, we evaluated and compared the features that 497 showed the highest importance values for each of the sets considered (Fig. 7A) 498 499 500 501 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint 502 503 504 Figure 7. A. Feature importance for XGBoost across FC metrics along with these features on the brain plots. 505 B. Connectome of overlapping features and its brain plot. The list of the edges is reported in the supplementary 506

(see S6-S7-S8). 507 508 There are 9 overlapping features (i.e. edges) across 3 FC metrics (PLV, ATM, AEC), where 509 we focused on the first 20 features with the largest feature importance according to XGBoost 510 evaluation. 511 As an example of the consistency of our results across different metrics, the edge between 512 the left supplementary motor area and the left paracentral lobule is the most important in all 3 513 feature sets. At the same time, the edge between the right frontal superior gyrus and right post 514 central gyrus is also among the top 3 features across three feature sets. 515 Another interesting observation is that the left cuneus appears three times among the top pairs 516 of edges (pairs of edges left cuneus and right lingual, left cuneus and Occipital middle gyrus, 517 the left cuneus and the right calcarine cortex), meanwhile the left supplementary motor area, 518 the right frontal superior gyrus and the right postcentral gyrus appear twice (for the exhaustive 519 list of the region of interest see the supplementary material S3) . 520 521 522 We selected 20 of the most significant features, according to their corrected pFDR values, and 523 ranked them across feature sets (where rank 1 means the most significant, and rank 20 524 indicates the least significant feature). This way, we identified that edge features for the next 525 pairs of ROIs: left supplementary motor area and left paracentral lobule, right frontal superior 526 gyrus and right postcentral gyrus, right paracentral lobule and right middle cingulum are 527 ranked as the top features (ranks 1 and 2) for three edge -specific FC metrics (AEC, PLV, 528 ATM), therefore we see a strong overlap of features across metrics. 529 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint We noticed that 9 edges out of 20 selected were the same for 3 different metrics – AEC, PLV 530 and ATM (Fig 7B) 531 To systematically test these findings, we applied pairwise Spearman correlation for the ranks 532 of three metrics (Fig 8). This revealed that, indeed, PLV and AEC features’ ranks are highly 533 positively correlated with a correlation coefficient equal to 0.89. 534 We observed that AEC, PLV and ATM have 9 overlapping features out of the 20 best features 535 according to the p -values, including the Cross Correlations leaves only one overlapping 536 feature. 537 538 539 540 Figure 8. Feature importance for XGBoost across metrics 541 542 543 544 545 546 547 548 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint 3. Discussion 549 550 In this study, we set out to identify a set of functional biomarkers to perform automated 551 differential diagnosis (among MS, MCI, PD, and ALS) from MEG data. Our work focuses on 552 the interpretability of the biomarkers, which is why we compared multiple connectivity metrics 553 (AEC, PLV, Pearson’s correlation coefficient, and ATM) that are different in terms of 554 interpretation. We tested the robustness of our analyses by feeding the data features to 555 multiple classification algorithms (i.e., XGBoost, SVM, LDA). In particular, we used a vast 556 amount of MEG data from a total number of 109 subjects affected by four different neurological 557 diseases: ALS, PD, MCI, and MS. Firstly, from each cohort, we extracted different feature sets 558 from four different FC metrics (PLV, ATM, AEC, and CC), each of which was obtained starting 559 from a symmetric matrix, leading to a total number of 6670 edge -wise features. Due to the 560 high dimensionality of our sample, we performed a Kruskal -Wallis test to reduce the 561 dimensionality and to consider only the most discriminative features. 562 563 Firstly, our results showed more than 120 significant edge features among the four different 564 feature sets (PLV, AEC, ATM, CC) for all the diseases. In particular, the Kruskal -Wallis test 565 showed statistical significance (pFDR<0.0001) of edge -wise PLV values between the right 566 frontal superior gyrus and the right postcentral gyrus. This finding might be related to the fact 567 that the frontal lobe is involved in physiological processes related to motor function (which are 568 notably impaired in SLA, PD, and MS) as well as to cognitive function, such as long -term 569 memory (which is often impaired in MCI). 570 571 We then move on to estimate the accuracy of the selected features by adding the features in 572 an iterative manner according to their p values. Such an approach enabled us to determine 573 the optimal number of top features for each FC metric taken separately. The AEC reached a 574 balanced accuracy of 63.07% with 38 features added, while the edge-wise PLV displayed the 575 best-balanced accuracy (67,14%) with a total number of 34 added features. To our knowledge, 576 no previous research has combined MEG data from patients with MCI, MS, PD, and ALS. 577 However, some research has focused on these conditions individually. As an example, López 578 ME et al. (López et al., 2014a) examined 105 subjects (36 controls and 69 MCI cases). They 579 identified spectral bio-marked changes in the theta, alpha, and beta frequency bands in MEG 580 data in MCI. Kim MJ et al. (M.-J. Kim et al., 2023) utilized a large EEG dataset including 417 581 MCI cases and applied a neural network that detected dementia with 81.1% accuracy. In the 582 closely related work, Giovannetti A. et al. (Giovannetti et al., 2021) presented the Deep-MEG 583 neural network, which was tested on 54 Alzheimer’s patients, each undergoing a five -minute 584 resting state task. Similarly to our study, they used functional connectivity indices, phase 585 locking values for classification. They reported an 87.4% AUC -ROC in identifying early MCI 586 symptoms. For multiple sclerosis disease, using EEG, Kiiski H. et al. (Kiiski et al., 2018) 587 assessed the responses of 35 subjects with multiple sclerosis during event -related potential 588 cognitive tasks over three years. They found significant correlations between ERP visual 589 components and cognitive function, identified using machine learning techniques. In related 590 research, Karaca et al. employed a continuous wavelet transform to differentiate nine multiple 591 sclerosis patients from 11 controls, achieving accuracy rates between 80%-88% in their best-592 performing models (Karaca et al., 2021) . Furthermore, Ahmadi A. et al. analyzed five MS 593 patients, developing a detection model using phase locking values and an online sequential 594 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint extreme learning classifier, with performance scores ranging from 82% to 96% across different 595 tasks (Ahmadi et al., 2019). 596 597 Our results are consistent across the three different ML algorithms (see Fig 3) with PLV and 598 AEC showing the best accuracy with respect to ATM and CC. These results are in agreement 599 with Chaturvedi et al., who showed that features extracted from the Phase lagIndex (PLI) were 600 able to better discriminate between PD patients with and without MCI as compared to spectral 601 features (Chaturvedi et al., 2019) . These results might suggest that phase -based metrics 602 might be more suitable in classification performance analysis as compared to amplitude-based 603 metrics (i.e., power spectra). Phase-based metrics specifically capture synchronization among 604 brain signals (defined as a bounded average phase difference). Synchronization is typically 605 measured in the framework of the communication -through-coherence hypothesis (Fries, 606 2015), whereby communication among brain regions might be captured by the coherent 607 activities of the corresponding brain signals. On the other hand, non-periodic activities, as well 608 as simple correlation coefficients, seem to perform less well in this context. Regardless of the 609 chosen FC metric, we see a consistent trend where the estimates at the edge level outperform 610 those at the nodal level in terms of disease classification. On the one hand, nodal metrics 611 capture the local activities and are predominantly sensitive to the dynamics of the local 612 activations. On the other hand, edge metrics focus primarily on how brain regions interact 613 among themselves. Therefore, our results might be interpreted as evidence that 614 neurodegenerative diseases primarily alter how regions interact with each other at the large -615 scale level. In particular, since the PLV is the best-performing metric, this might be interpreted 616 as the neurodegenerative diseases altering the ability of brain regions that are far apart to 617 synchronize their activities. 618 While the PLV is sensitive to volume conduction, volume conduction does not offer a 619 reasonable explanation for the ability to classify different subjects according to diagnosis. 620 Furthermore, the edges of the ATM (that are more robust to volume conduction artifacts) also 621 confirm the ability to correctly diagnose patients well above chance level. 622 Furthermore, the set of edges that contributed more to the classification were FC metric -623 independent. In general, it is interesting to note that the edges that are relevant to classification 624 irrespective of the metrics are typically longer range connections, either in the antero-posterior 625 direction of cross -hemispheric. Again, these results are representative of significant 626 involvement and, as a consequence, impairment, of the frontal lobe in PD, ALS, MS and MCI 627 respectively (Foong et al., 1997; Kendi et al., 2008; Trojsi et al., 2012; Wang et al., 2012). 628 629 Not many studies explain the rationale behind the feature extraction and selection method 630 choice. It usually consists in a trade-off between enriching the information of interest and the 631 risk of adding irrelevant inputs that could reduce the classification performance. Two types of 632 approaches have been proposed. The first one consists in considering that fusing features will 633

in an improvement of the classification performance. For instance, Geraedts et al fused 634 features obtained from the estimation of the power spectra in seven frequency bands (resulting 635 in 16 674 features per EEG) before selecting them to discriminate cognitive functions in 636 patients with Parkinson’s Disease during Deep Brain Stimulation (Geraedts et al., 2021) . 637 Similarly, López et al extracted spectral and non -linear metrics before fusing them and 638 selecting them via their fast correlation-based filter to discriminate early Alzheimer’s disease 639 and its prodromal form from healthy subjects (López et al., 2014b). 640 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint Another approach consists in fusing the classifiers’ output rather than the different types of 641 features. Fusing the classifiers’ outputs confers a higher reliability and robustness through 642 redundancy and facilitates the integration of heterogeneous data without 643 normalizing them (Roli, 2009; Roli & Fumera, 2002; Ruta & Gabrys, 2000) . In a recent work, 644 we proposed a framework that was based on Riemannian geometry extended to functional 645 connectivity measures through an ensemble learning method. We validated it on numerous 646 publicly available datasets (Corsi et al., 2022) . Such an approach notably ranked 1st in a 647 clinical challenge that consisted in discriminating mental states from data obtained from stroke 648 patients (Corsi et al., 2021 ). Future work will consist in considering this type of approach to 649 enrich the information of interest used to discriminate diseases. 650 In conclusion, our is the first study investigating automated differential diagnosis in several 651 neurological diseases, based on different connectivity metrics as well as on different 652 classification algorithms. Our results demonstrate the existence of a common set of edges 653 that drive the classification performance, irrespective of the particular metric chosen or the 654 algorithm. These results demonstrate the existence of a robust set of long-range connections 655 that are altered in neurodegeneration, across multiple diseases, and valid in terms of 656 distinguishing specific diseases. Future studies will have to confirm the external validity of our 657

to different datasets and extend our analyses to more neurodegenerative diseases. 658 Data availability statement 659 The magnetoencephalography data and the reconstructed avalanches are available upon 660 request to the corresponding author, conditional on appropriate ethics approval at the local 661 site. The availability of the data was not previously included in the ethical approval, and 662 therefore data cannot be shared directly. In case data are requested, the corresponding author 663 will request an amendment to the local ethical committee. 664 Competing interests 665 The authors report no competing interests 666 Funding 667 668 This work was financially supported by Ministero Sviluppo Economico (Contratto di sviluppo 669 industriale "Farmaceutica e Diagnostica" [CDS 000606]); European Union 670 “NextGenerationEU,” (Investimento 3.1. M4. C2), project IR0000011, EBRAINS-Italy of PNRR 671 and Contratto di sviluppo industriale - “Progetto CDS000904 - agevolazioni ex DM del 672 09/12/2014” 673 674 675 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint

676 677 Ahmadi, A., Davoudi, S., & Daliri, M. R. (2019). Computer Aided Diagnosis System for 678 multiple sclerosis disease based on phase to amplitude coupling in covert visual 679 attention. Computer Methods and Programs in Biomedicine, 169, 9–18. 680 https://doi.org/10.1016/j.cmpb.2018.11.006 681 Albert, M. S., DeKosky, S. T., Dickson, D., Dubois, B., Feldman, H. H., Fox, N. C., Gamst, 682 A., Holtzman, D. M., Jagust, W. J., Petersen, R. C., Snyder, P. J., Carrillo, M. C., 683 Thies, B., & Phelps, C. H. (2011). The diagnosis of mild cognitive impairment due to 684 Alzheimer’s disease: Recommendations from the National Institute on Aging-685 Alzheimer’s Association workgroups on diagnostic guidelines for Alzheimer’s 686 disease. Alzheimer’s & Dementia: The Journal of the Alzheimer’s Association, 7(3), 687 270–279. https://doi.org/10.1016/j.jalz.2011.03.008 688 Bastos, A. M., & Schoffelen, J.-M. (2016). A Tutorial Review of Functional Connectivity 689 Analysis Methods and Their Interpretational Pitfalls. Frontiers in Systems 690 Neuroscience, 9. https://doi.org/10.3389/fnsys.2015.00175 691 Brookes, M. J., Woolrich, M., Luckhoo, H., Price, D., Hale, J. R., Stephenson, M. C., Barnes, 692 G. R., Smith, S. M., & Morris, P. G. (2011). Investigating the electrophysiological 693 basis of resting state networks using magnetoencephalography. Proceedings of the 694 National Academy of Sciences of the United States of America, 108(40), 16783–695 16788. https://doi.org/10.1073/pnas.1112685108 696 Brookes, M. J., Woolrich, M. W., & Barnes, G. R. (2012). Measuring functional connectivity 697 in MEG: A multivariate approach insensitive to linear source leakage. 698 https://doi.org/10.1016/j.neuroimage.2012.03.048 699 Brooks, B. R. (1994). El Escorial World Federation of Neurology criteria for the diagnosis of 700 amyotrophic lateral sclerosis. Subcommittee on Motor Neuron Diseases/Amyotrophic 701 Lateral Sclerosis of the World Federation of Neurology Research Group on 702 Neuromuscular Diseases and th. Journal of the Neurological Sciences, 124, 96–107. 703 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint Chaturvedi, M., Bogaarts, J. G., Kozak (Cozac), V. V., Hatz, F., Gschwandtner, U., Meyer, 704 A., Fuhr, P., & Roth, V. (2019). Phase lag index and spectral power as QEEG 705 features for identification of patients with mild cognitive impairment in Parkinson’s 706 disease. Clinical Neurophysiology, 130(10), 1937–1944. 707 https://doi.org/10.1016/j.clinph.2019.07.017 708 Cipriano, L., Minino, R., Liparoti, M., Polverino, A., Romano, A., Bonavita, S., Pirozzi, M. A., 709 Quarantelli, M., Jirsa, V., Sorrentino, G., Sorrentino, P., & Troisi Lopez, E. (2024). 710 Flexibility of brain dynamics is increased and predicts clinical impairment in 711 relapsing–remitting but not in secondary progressive multiple sclerosis. Brain 712 Communications, 6(2), fcae112. https://doi.org/10.1093/braincomms/fcae112 713 Corsi, M.-C., Chavez, M., Schwartz, D., George, N., Hugueville, L., Kahn, A. E., Dupont, S., 714 Bassett, D. S., & De Vico Fallani, F. (2020). Functional disconnection of associative 715 cortical areas predicts performance during BCI training. NeuroImage, 209, 116500. 716 https://doi.org/10.1016/j.neuroimage.2019.116500 717 Corsi, M.-C., Chevallier, S., de Vico Fallani, F., & Yger, F. (2022). Functional connectivity 718 ensemble method to enhance BCI performance (FUCONE). IEEE Transactions on 719 Biomedical Engineering, 1–1. https://doi.org/10.1109/TBME.2022.3154885 720 Corsi, M.-C., Yger, F., Chevallier, S., & Noûs, C. (2021). Riemannian Geometry on 721 Connectivity for Clinical BCI. ICASSP 2021 - 2021 IEEE International Conference on 722 Acoustics, Speech and Signal Processing (ICASSP), 980–984. 723 https://doi.org/10.1109/ICASSP39728.2021.9414790 724 Finn, E. S., & Rosenberg, M. D. (2021). Beyond fingerprinting: Choosing predictive 725 connectomes over reliable connectomes. NeuroImage, 239, 118254. 726 https://doi.org/10.1016/j.neuroimage.2021.118254 727 Foong, J., Rozewicz, L., Quaghebeur, G., Davie, C. A., Kartsounis, L. D., Thompson, A. J., 728 Miller, D. H., & Ron, M. A. (1997). Executive function in multiple sclerosis. The role of 729 frontal lobe pathology. Brain, 120(1), 15–26. https://doi.org/10.1093/brain/120.1.15 730 Fries, P. (2015). Rhythms For Cognition: Communication Through Coherence HHS Public 731 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint Access. Neuron, 88(1), 220–235. https://doi.org/10.1016/j.neuron.2015.09.034 732 Friston, K. J. (1994). Functional and effective connectivity in neuroimaging: A synthesis. 733 Human Brain Mapping, 2(1–2), 56–78. https://doi.org/10.1002/hbm.460020107 734 Geraedts, V. J., Koch, M., Contarino, M. F., Middelkoop, H. A. M., Wang, H., van Hilten, J. 735 J., Bäck, T. H. W., & Tannemaat, M. R. (2021). Machine learning for automated 736 EEG-based biomarkers of cognitive impairment during Deep Brain Stimulation 737 screening in patients with Parkinson’s Disease. Clinical Neurophysiology, 132(5), 738 1041–1048. https://doi.org/10.1016/j.clinph.2021.01.021 739 Gibb, W. R. G., & Lees, A. J. (1988). A comparison of clinical and pathological features of 740 young‐ and old‐onset Parkinson’s disease. Neurology, 38(9), 1402–1402. 741 https://doi.org/10.1212/WNL.38.9.1402 742 Giovannetti, A., Susi, G., Casti, P., Mencattini, A., Pusil, S., López, M. E., Di Natale, C., & 743 Martinelli, E. (2021). Deep-MEG: Spatiotemporal CNN features and multiband 744 ensemble classification for predicting the early signs of Alzheimer’s disease with 745 magnetoencephalography. Neural Computing and Applications, 33(21), 14651–746 14667. https://doi.org/10.1007/s00521-021-06105-4 747 Grinsztajn, L., Oyallon, E., & Varoquaux, G. (2022). Why do tree-based models still 748 outperform deep learning on tabular data? (arXiv:2207.08815). arXiv. 749 https://doi.org/10.48550/arXiv.2207.08815 750 Haldeman, C., & Beggs, J. M. (2005). Critical Branching Captures Activity in Living Neural 751 Networks and Maximizes the Number of Metastable States. Physical Review Letters, 752 94(5), 058101. https://doi.org/10.1103/PhysRevLett.94.058101 753 Karaca, B. K., Akşahin, M. F., & Öcal, R. (2021). Detection of multiple sclerosis from photic 754 stimulation EEG signals. Biomedical Signal Processing and Control, 67, 102571. 755 https://doi.org/10.1016/j.bspc.2021.102571 756 Kelly, R. E., & Hoptman, M. J. (2022). Replicability in Brain Imaging. Brain Sciences, 12(3), 757 397. https://doi.org/10.3390/brainsci12030397 758 Kendi, A. T. K., Lehericy, S., Luciana, M., Ugurbil, K., & Tuite, P. (2008). Altered Diffusion in 759 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint the Frontal Lobe in Parkinson Disease. American Journal of Neuroradiology, 29(3), 760 501–505. https://doi.org/10.3174/ajnr.A0850 761 Kiiski, H., Jollans, L., Donnchadha, S. Ó., Nolan, H., Lonergan, R., Kelly, S., O’Brien, M. C., 762 Kinsella, K., Bramham, J., Burke, T., Hutchinson, M., Tubridy, N., Reilly, R. B., & 763 Whelan, R. (2018). Machine Learning EEG to Predict Cognitive Functioning and 764 Processing Speed Over a 2-Year Period in Multiple Sclerosis Patients and Controls. 765 Brain Topography, 31(3), 346–363. https://doi.org/10.1007/s10548-018-0620-4 766 Kim, J.-H. (2009). Estimating classification error rate: Repeated cross-validation, repeated 767 hold-out and bootstrap. Computational Statistics & Data Analysis, 53(11), 3735–768 3745. https://doi.org/10.1016/j.csda.2009.04.009 769 Kim, M.-J., Youn, Y. C., & Paik, J. (2023). Deep learning-based EEG analysis to classify 770 normal, mild cognitive impairment, and dementia: Algorithms and dataset. 771 NeuroImage, 272, 120054. https://doi.org/10.1016/j.neuroimage.2023.120054 772 Lachaux, J. P., Rodriguez, E., Martinerie, J., & Varela, F. J. (1999). Measuring phase 773 synchrony in brain signals. Human Brain Mapping, 8(4), 194–208. 774 https://doi.org/10.1002/(sici)1097-0193(1999)8:43.0.co;2-c 775 Lin, W., Gao, Q., Du, M., Chen, W., & Tong, T. (2021). Multiclass diagnosis of stages of 776 Alzheimer’s disease using linear discriminant analysis scoring for multimodal data. 777 Computers in Biology and Medicine, 134, 104478. 778 https://doi.org/10.1016/j.compbiomed.2021.104478 779 López, M. E., Cuesta, P., Garcés, P., Castellanos, P. N., Aurtenetxe, S., Bajo, R., Marcos, 780 A., Delgado, M. L., Montejo, P., López-Pantoja, J. L., Maestú, F., & Fernandez, A. 781 (2014a). MEG spectral analysis in subtypes of mild cognitive impairment. Age 782 (Dordrecht, Netherlands), 36(3), 9624. https://doi.org/10.1007/s11357-014-9624-5 783 López, M. E., Cuesta, P., Garcés, P., Castellanos, P. N., Aurtenetxe, S., Bajo, R., Marcos, 784 A., Delgado, M. L., Montejo, P., López-Pantoja, J. L., Maestú, F., & Fernandez, A. 785 (2014b). MEG spectral analysis in subtypes of mild cognitive impairment. Age 786 (Dordrecht, Netherlands), 36(3), 9624. https://doi.org/10.1007/s11357-014-9624-5 787 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint Manju, N., Harish, B. S., & Prajwal, V. (n.d.). Ensemble Feature Selection and Classification 788 of Internet Traffic using XGBoost Classifier. International Journal of Computer 789 Network and Information Security, 11(7), 37. 790 Maqsood, S., Damaševičius, R., & Maskeliūnas, R. (2022). Multi-Modal Brain Tumor 791 Detection Using Deep Neural Network and Multiclass SVM. Medicina (Kaunas, 792 Lithuania), 58(8), 1090. https://doi.org/10.3390/medicina58081090 793 Nolte, G. (2003). The magnetic lead field theorem in the quasi-static approximation and its 794 use for magnetoencephalography forward calculation in realistic volume conductors. 795 Physics in Medicine and Biology, 48(22), 3637–3652. https://doi.org/10.1088/0031-796 9155/48/22/002 797 Polverino, A., Lopez, E. T., Minino, R., Liparoti, M., Romano, A., Trojsi, F., Lucidi, F., Gollo, 798 L., Jirsa, V., Sorrentino, G., & Sorrentino, P. (2022). Flexibility of Fast Brain 799 Dynamics and Disease Severity in Amyotrophic Lateral Sclerosis. Neurology. 800 https://doi.org/10.1212/WNL.0000000000201200 801 Polverino, A., Troisi Lopez, E., Liparoti, M., Minino, R., Romano, A., Cipriano, L., Trojsi, F., 802 Jirsa, V., Sorrentino, G., & Sorrentino, P. (2024). Altered spreading of fast aperiodic 803 brain waves relates to disease duration in Amyotrophic Lateral Sclerosis. Clinical 804 Neurophysiology, 163, 14–21. https://doi.org/10.1016/j.clinph.2024.04.003 805 Roli, F. (2009). Multiple Classifier Systems. In S. Z. Li & A. Jain (Eds.), Encyclopedia of 806 Biometrics (pp. 981–986). Springer US. 807 http://link.springer.com/referenceworkentry/10.1007/978-0-387-73003-5_148 808 Roli, F., & Fumera, G. (2002). Analysis of Linear and Order Statistics Combiners for Fusion 809 of Imbalanced Classifiers. Multiple Classifier Systems, 252–261. 810 https://doi.org/10.1007/3-540-45428-4_25 811 Romano, A., Troisi Lopez, E., Cipriano, L., Liparoti, M., Minino, R., Polverino, A., Cavaliere, 812 C., Aiello, M., Granata, C., Sorrentino, G., & Sorrentino, P. (2023). Topological 813 changes of fast large-scale brain dynamics in mild cognitive impairment predict early 814 memory impairment: A resting-state, source reconstructed, 815 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint magnetoencephalography study. Neurobiology of Aging, 132, 36–46. 816 https://doi.org/10.1016/j.neurobiolaging.2023.08.003 817 Romano, A., Trosi Lopez, E., Liparoti, M., Polverino, A., Minino, R., Trojsi, F., Bonavita, S., 818 Mandolesi, L., Granata, C., Amico, E., Sorrentino, G., & Sorrentino, P. (2022). The 819 progressive loss of brain network fingerprints in Amyotrophic Lateral Sclerosis 820 predicts clinical impairment. NeuroImage: Clinical, 35, 103095. 821 https://doi.org/10.1016/j.nicl.2022.103095 822 Ruta, D., & Gabrys, B. (2000). An Overview of Classifier Fusion Methods. Computing and 823 Information Systems, 7(1), 1–10. 824 Shine, J. M., Bissett, P. G., Bell, P. T., Koyejo, O., Balsters, J. H., Gorgolewski, K. J., 825 Moodie, C. A., & Poldrack, R. A. (2016). The Dynamics of Functional Brain Networks: 826 Integrated Network States during Cognitive Task Performance. Neuron, 92(2), 544–827 554. https://doi.org/10.1016/j.neuron.2016.09.018 828 Shriki, O., Alstott, J., Carver, F., Holroyd, T., Henson, R. N. A., Smith, M. L., Coppola, R., 829 Bullmore, E., & Plenz, D. (2013). Neuronal avalanches in the resting MEG of the 830 human brain. Journal of Neuroscience, 33(16), 7079–7090. 831 https://doi.org/10.1523/JNEUROSCI.4286-12.2013 832 Sorrentino, P., Petkoski, S., Sparaco, M., Troisi Lopez, E., Signoriello, E., Baselice, F., 833 Bonavita, S., Pirozzi, M. A., Quarantelli, M., Sorrentino, G., & Jirsa, V. (2022). Whole-834 brain propagation delays in multiple sclerosis, a combined tractography—835 Magnetoencephalography study. The Journal of Neuroscience, JN-RM-0938-22. 836 https://doi.org/10.1523/JNEUROSCI.0938-22.2022 837 Sorrentino, P., Rucco, R., Baselice, F., Micco, R. D., Tessitore, A., Hillebrand, A., Mandolesi, 838 L., Breakspear, M., Gollo, L. L., & Sorrentino, G. (2019). Extensive functional 839 repertoire underpins complex behaviours: Insights from Parkinson’s disease. In 840 bioRxiv (p. 823849). Cold Spring Harbor Laboratory. https://doi.org/10.1101/823849 841 Sorrentino, P., Rucco, R., Jacini, F., Trojsi, F., Lardone, A., Baselice, F., Femiano, C., 842 Santangelo, G., Granata, C., Vettoliere, A., Monsurrò, M. R., Tedeschi, G., & 843 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint Sorrentino, G. (2018). Brain functional networks become more connected as 844 amyotrophic lateral sclerosis progresses: A source level magnetoencephalographic 845 study. NeuroImage: Clinical, 20, 564–571. https://doi.org/10.1016/j.nicl.2018.08.001 846 Sorrentino, P., Rucco, R., Lardone, A., Liparoti, M., Lopez, E. T., Cavaliere, C., Soricelli, A., 847 Jirsa, V., Sorrentino, G., & Amico, E. (n.d.). Clinical connectome fingerprints of 848 cognitive decline. https://doi.org/10.1101/2020.10.09.332635 849 Sorrentino, P., Seguin, C., Rucco, R., Liparoti, M., Troisi Lopez, E., Bonavita, S., Quarantelli, 850 M., Sorrentino, G., Jirsa, V., & Zalesky, A. (2021). The structural connectome 851 constrains fast brain dynamics. eLife, 10, e67400. https://doi.org/10.7554/eLife.67400 852 Stam, C. J. (2010). Use of magnetoencephalography (MEG) to study functional brain 853 networks in neurodegenerative disorders. Journal of the Neurological Sciences, 854 289(1–2), 128–134. https://doi.org/10.1016/j.jns.2009.08.028 855 Tagliazucchi, E., von Wegner, F., Morzelewski, A., Brodbeck, V., & Laufs, H. (2012). 856 Dynamic BOLD functional connectivity in humans and its electrophysiological 857 correlates. Frontiers in Human Neuroscience, 6(DEC). 858 https://doi.org/10.3389/fnhum.2012.00339 859 Thölke, P., Mantilla-Ramos, Y.-J., Abdelhedi, H., Maschke, C., Dehgan, A., Harel, Y., 860 Kemtur, A., Mekki Berrada, L., Sahraoui, M., Young, T., Bellemare Pépin, A., El 861 Khantour, C., Landry, M., Pascarella, A., Hadid, V., Combrisson, E., O’Byrne, J., & 862 Jerbi, K. (2023). Class imbalance should not throw you off balance: Choosing the 863 right classifiers and performance metrics for brain decoding with imbalanced data. 864 NeuroImage, 277, 120253. https://doi.org/10.1016/j.neuroimage.2023.120253 865 Thompson, A. J., Banwell, B. L., Barkhof, F., Carroll, W. M., Coetzee, T., Comi, G., Correale, 866 J., Fazekas, F., Filippi, M., Freedman, M. S., Fujihara, K., Galetta, S. L., Hartung, H. 867 P., Kappos, L., Lublin, F. D., Marrie, R. A., Miller, A. E., Miller, D. H., Montalban, X., 868 … Cohen, J. A. (2018). Diagnosis of multiple sclerosis: 2017 revisions of the 869 McDonald criteria. The Lancet Neurology, 17(2), 162–173. 870 https://doi.org/10.1016/S1474-4422(17)30470-2 871 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint Trojsi, F., Monsurro, M. R., Esposito, F., & Tedeschi, G. (2012). Widespread structural and 872 functional connectivity changes in amyotrophic lateral sclerosis: Insights from 873 advanced neuroimaging research. Neural Plasticity, 2012. 874 Wang, Z., Jia, X., Liang, P., Qi, Z., Yang, Y., Zhou, W., & Li, K. (2012). Changes in thalamus 875 connectivity in mild cognitive impairment: Evidence from resting state fMRI. 876 European Journal of Radiology, 81(2), 277–285. 877 https://doi.org/10.1016/j.ejrad.2010.12.044 878 879 880 881 882 883 884 885 886 887 888 889 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint Supplementary materials Supplementary materials 1: The balanced accuracies for all feature sets with SVM classifier. Each bar plot displays the averaged accuracy with its standard errors. Supplementary materials 2: The balanced accuracies for all feature sets with XGBoost classifier. Each bar plot displays the averaged accuracy with its standard errors. . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint Supplementary materials 3: Exhaustive list of regions of interest N° ROIS Anatomical correspondance N° ROIS Anatomical correspondance N° ROI S Anatomical correspondance 1 Rectus L 40 Rectus R 79 Hippocampus L 2 Olfactory L 41 Olfactory R 80 Hippocampus R 3 Frontal Superior Orbital L 42 Frontal Superior Orbital R 81 Amygdala L 4 Frontal Medial Orbital L 43 Frontal Medial Orbital R 82 Amygdala R 5 Frontal Medial Orbital L 44 Frontal Medial Orbital R 83 Caudate L 6 Frontal Inferior Orbital L 45 Frontal Inferior Orbital R 84 Caudate R 7 Frontal Superior L 46 Frontal Superior R 85 Putamen L 8 Frontal Medial L 47 Frontal Medial R 86 Putamen R 9 Frontal Inferior Opercolum L 48 Frontal Inferior Opercolum R 87 Pallidum L 10 Frontal Inferior Triangular L 49 Frontal Inferior Triangular R 88 Pallidum R 11 Frontal Superior Medial L 50 Frontal Superior Medial R 89 Thalamus L 12 Supplementary Motor area L 51 Supplementary Motor area R 90 Thalamus R 13 Paracentral Lobule L 52 Paracentral Lobule R 91 Cerebellum Crus1 L 14 Precentral L 53 Precentral R 92 Cerebelum Crus1 R 15 Rolandic Opercolum L 54 Rolandic Opercolum R 93 Cerebelum Crus2 L 16 Postcentral L 55 Postcentral R 94 Cerebelum Crus2 R . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint 17 Parietal Superior L 56 Parietal Superior R 95 Cerebellum 3 L 18 Parietal Inferior L 57 Parietal Inferior R 96 Cerebellum 3 R 19 Supra Marginal L 58 Supra Marginal R 97 Cerebellum 4 5 L 20 Angular L 59 Angular R 98 Cerebellum 4 5 R 21 Precuneus L 60 Precuneus R 99 Cerebellum 6 L 22 Occipital Superior L 61 Occipital Superior R 100 Cerebellum 6 R 23 Occipital Medial L 62 Occipital Medial R 101 Cerebellum 7b L 24 Occipital Inferior L 63 Occipital Inferior R 102 Cerebellum 7b R 25 Calcarine L 64 Calcarine R 103 Cerebellum 8 L 26 Cuneus L 65 Cuneus R 104 Cerebellum 8 R 27 Lingual gyrus L 66 Lingual gyrus R 105 Cerebellum 9 L 28 Fusiform gyrus L 67 Fusiform gyrus R 106 Cerebellum 9 R 29 Heschl L 68 Heschl R 107 Cerebellum 10 L 30 Temporal Superior L 69 Temporal Superior R 108 Cerebellum 10 R 31 Temporal Medial L 70 Temporal Medial R 109 Vermis 1 2 32 Temporal Inferior L 71 Temporal Inferior R 110 Vermis 3 33 Temporal Pole Superior L 72 Temporal Pole Superior R 111 Vermis 4 5 34 Temporal Pole Medial L 73 Temporal Pole Medial R 112 Vermis 6 35 ParaHippocampal L 74 ParaHippocampal R 113 Vermis 7 36 Cingulum Anterior L 75 Cingulum Anterior R 114 Vermis 8 37 Cingulum Medial L 76 Cingulum Medial R 115 Vermis 9 38 Cingulum Posterior L 77 Cingulum Posterior R 116 Vermis 10 39 Insula L 78 Insula R Supplementary materials 4: Balanced accuracy for the number of features which contain the best accuracy across different metrics (PLV, AEC, ATM, CC) with XGBoost. n=15 n=17 n=18 n=19 n=24 n=25 n=26 n=27 n=28 n=35 n=38 n=39 Edge metrics PLV 0.628 0.606 0.597 0.618 0.575 0.577 0.573 0.563 0.572 0.544 0.559 0.557 AEC 0.554 0.606 0.597 0.601 0.594 0.611 0.626 0.622 0.617 0.602 0.602 0.622 ATM 0.513 0.533 0.533 0.551 0.55 0.535 0.524 0.519 0.531 0.561 0.554 0.575 CC 0.529 0.565 0.559 0.574 0.591 0.586 0.595 0.594 0.591 0.638 0.612 0.608 Nodal metrics CC (betw.) 0.496 0.489 0.504 0.483 0.468 0.482 0.48 0.467 0.476 0.47 0.459 0.448 PLV (betw.) 0.379 0.385 0.388 0.384 0.376 0.364 0.357 0.356 0.374 0.354 0.341 0.341 PLV (eign.) 0.378 0.372 0.372 0.373 0.378 0.376 0.401 0.405 0.402 0.383 0.399 0.4 AEC (eign.) 0.414 0.421 0.435 0.404 0.401 0.407 0.401 0.403 0.413 0.405 0.399 0.403 AEC (betw.) 0.374 0.359 0.358 0.356 0.334 0.323 0.327 0.321 0.332 0.342 0.35 0.352 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint ATM (betw.) 0.374 0.359 0.358 0.356 0.334 0.323 0.327 0.321 0.332 0.342 0.35 0.352 ATM (eign.) 0.422 0.427 0.423 0.415 0.424 0.414 0.418 0.419 0.434 0.41 0.408 0.412 PLV (degree) 0.333 0.343 0.328 0.327 0.341 0.371 0.363 0.364 0.348 0.361 0.335 0.339 AEC (degree) 0.298 0.308 0.289 0.289 0.297 0.299 0.31 0.298 0.32 0.284 0.27 0.262 ATM (degree) 0.373 0.377 0.372 0.378 0.397 0.376 0.368 0.375 0.376 0.346 0.373 0.368 CC (degree) 0.537 0.557 0.548 0.539 0.522 0.529 0.535 0.533 0.534 0.531 0.528 0.544 Supplementary materials 5: Balanced accuracy for the number of features which contain the best accuracy across different metrics (PLV, AEC, ATM, CC) with SVM. n=16 n=19 n=20 n=21 n=22 n=25 n=28 n=30 n=31 n=32 n=34 n=35 n=36 n=37 n=39 Edge metrics PLV 0.61 4 0.64 0.63 5 0.63 6 0.62 8 0.63 4 0.63 3 0.60 1 0.63 1 0.64 0.65 1 0.60 7 0.66 5 0.64 8 0.61 AEC 0.54 1 0.57 4 0.61 1 0.60 1 0.61 3 0.64 8 0.65 2 0.66 1 0.67 9 0.66 5 0.66 5 0.64 4 0.63 2 0.65 4 0.64 7 ATM 0.53 6 0.54 4 0.54 1 0.53 5 0.53 2 0.45 4 0.54 4 0.58 6 0.59 7 0.60 5 0.60 2 0.59 6 0.59 7 0.59 9 0.58 CC 0.50 5 0.48 3 0.54 3 0.53 7 0.50 1 0.55 8 0.57 7 0.57 5 0.56 8 0.58 4 0.6 0.62 4 0.62 1 0.62 5 0.55 9 Nodal metrics AEC (eign. centr.) 0.48 6 0.49 3 0.49 5 0.48 9 0.49 0.47 2 0.45 9 0.45 9 0.45 9 0.46 1 0.46 1 0.45 5 0.45 4 0.45 4 0.43 9 AEC (betw. centr.) 0.29 3 0.28 8 0.29 3 0.29 9 0.30 4 0.31 7 0.25 0.34 6 0.34 2 0.34 2 0.35 0.34 1 0.34 1 0.33 4 0.34 CC (betw. centr.) 0.48 7 0.48 0.49 7 0.49 0.51 2 0.49 5 0.53 0.45 3 0.49 2 0.49 6 0.48 6 0.47 9 0.49 5 0.47 4 0.45 7 PLV (betw. centr.) 0.41 2 0.42 7 0.40 1 0.39 4 0.37 8 0.36 1 0.33 0.35 8 0.34 4 0.34 4 0.36 3 0.35 4 0.34 9 0.32 8 0.36 9 PLV (eign. centr.) 0.39 7 0.42 6 0.43 2 0.42 2 0.45 8 0.46 7 0.45 8 0.46 2 0.48 7 0.48 8 0.45 4 0.43 6 0.40 2 0.42 3 0.49 2 ATM (betw. centr.) 0.29 3 0.28 8 0.29 3 0.29 9 0.30 4 0.31 7 0.33 9 0.34 6 0.34 2 0.34 2 0.35 0.34 1 0.34 1 0.33 4 0.34 ATM (Eign. centr.) 0.47 5 0.47 5 0.48 3 0.48 4 0.48 3 0.45 4 0.46 3 0.45 4 0.45 6 0.45 9 0.44 8 0.44 9 0.44 2 0.43 5 0.42 3 PLV (degree) 0.39 1 0.36 2 0.38 0.4 0.40 7 0.38 9 0.38 7 0.37 9 0.37 7 0.37 6 0.40 6 0.39 8 0.34 3 0.39 4 0.35 3 AEC (degree) 0.36 5 0.34 3 0.30 4 0.35 0.33 7 0.36 9 0.32 7 0.32 6 0.32 4 0.32 7 0.32 3 0.32 5 0.32 2 0.33 4 0.32 2 ATM (degree) 0.31 0.31 5 0.35 8 0.34 1 0.32 7 0.31 0.29 2 0.29 3 0.31 5 0.29 2 0.33 1 0.32 1 0.30 2 0.32 7 0.31 4 CC (degree) 0.46 1 0.41 4 0.35 9 0.40 4 0.39 1 0.37 9 0.41 9 0.38 8 0.4 0.39 4 0.40 8 0.38 2 0.38 9 0.37 9 0.37 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint Supplementary materials 6: Features importance. List of associated edges for PLV. 12-13 Supplementary Motor area L and Paracentral Lobule L 46-55 Frontal Superior R and Postcentral R 52-76 Paracentral Lobule R and Cingulum Medial R 12-48 Supplementary Motor area L and Frontal Inferior Opercolum R 26-66 Cuneus L and Lingual gyrus R 37-54 Cingulum Medial L and Rolandic Opercolum R 38-77 Cingulum Posterior L and Cingulum Posterior R 26-62 Cuneus L and Occipital Medial R 46-57 Frontal Superior R and Parietal Inferior R 46-53 Frontal Superior R and Precentral R 56-76 Parietal Superior R and Cingulum Medial R 21-37 Precuneus L and Cingulum Medial L 54-58 Rolandic Opercolum R and Supra Marginal R 37-90 Cingulum Medial and L Thalamus R 26-64 Cuneus L and Calcarine R 7-46 Frontal Superior L and Frontal Superior R 53-55 Precentral R and Postcentral R 5-85 Frontal Medial Orbital L and Putamen L 12-83 Supplementary Motor area L and Caudate L 58-86 Supra Marginal R and Putamen R . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint Supplementary materials 7: Features importance. List of associated edges for AEC. 12-13 Supplementary Motor area L and Paracentral Lobule L 51-52 Paracentral Lobule R and Paracentral Lobule R 46-55 Frontal Superior R and Postcentral R 26-66 Cuneus L and Lingual gyrus R 52-76 Paracentral Lobule R and Cingulum Medial R 5-85 Frontal Medial Orbital L and Putamen L 46-53 Frontal Superior R and Precentral R 53-55 Precentral R and Postcentral R 38-77 Cingulum Posterior L and Cingulum Posterior R 54-58 Rolandic Opercolum R and Supra Marginal R 37-54 Cingulum Medial L and Rolandic Opercolum R 58-90 Supra Marginal R and Thalamus R 26-62 Cuneus L and Occipital Medial R 53-84 Precentral R and Caudate R 26-64 Cuneus L and Calcarine R 7-46 Frontal Superior L and Frontal Superior R 63-111 Occipital Inferior R and Vermis 4 5 54-88 Rolandic Opercolum R Pallidum R 13-83 Paracentral Lobule L and Caudate L 53-57 Precentral R and Parietal Inferior R . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint Supplementary materials 8: Features importance. List of associated edges for ATM. 12-13 Supplementary Motor area L and Paracentral Lobule L 46-55 Frontal Superior R and Postcentral R 67-99 Fusiform gyrus R and Cerebelum 6 L 52-76 Paracentral Lobule R and Cingulum Medial R 97-108 Cerebelum 4 5 L and Cerebelum 10 R 53-55 Precentral R and Postcentral R 51-52 Paracentral Lobule R and Paracentral Lobule R 46-53 Frontal Superior R and Precentral R 55-90 Postcentral R and Thalamus R 9-16 Frontal Inferior Opercolum L and Postcentral L 26-63 Cuneus L and Occipital Inferior R 15-16 Rolandic Opercolum L and Postcentral L 21-37 Precuneus L and Cingulum Medial L 26-66 Cuneus L and Lingual gyrus R 58-90 Supra Marginal R and Thalamus R 7-46 Frontal Superior L and Frontal Superior R . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint 12-83 Supplementary Motor area L and Caudate L 26-62 Cuneus L and Occipital Medial R 54-88 Rolandic Opercolum R Pallidum R 26-64 Cuneus L and Calcarine R . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprintthis version posted June 17, 2024. ; https://doi.org/10.1101/2024.06.17.24309023doi: medRxiv preprint