Electroencephalographic Network Topologies Predict Antidepressant Responses in Patients With Major Depressive Disorder

Medication therapy seems to be an effective treatment for major depressive disorder (MDD). However, although the efficacies of various medicines are equal or similar on average, they vary widely among individuals. Therefore, an understanding of methods for the timely evaluation of short-term therapeutic response and prediction of symptom improvement after a specific course of medication at the individual level at the initial stage of treatment is very important. In our present study, we sought to identify a neurobiological signature of the response to short-term antidepressant treatment. Related brain network analysis was applied in resting-state electroencephalogram (EEG) datasets from patients with MDD. The corresponding EEG networks were constructed accordingly and then quantitatively measured to predict the efficacy after eight weeks of medication, as well as to distinguish the therapeutic responders from non-responders. The results of our present study revealed that the corresponding resting-state EEG networks became significantly weaker after one week of treatment, and the eventual medication efficacy was reliably predicted using the changes in those network properties within the one-week medication regimen. Moreover, the corresponding resting-state networks at baseline were also proven to precisely distinguish those responders from other individuals with an accuracy of 96.67% when using the spatial network topologies as the discriminative features. These findings consistently provide a deeper neurobiological understanding of antidepressant treatment and a reliable and quantitative approach for personalized treatment of MDD.

Moreover, the corresponding resting-state networks at 23 baseline were also proven to precisely distinguish those disorder characterized by sustained negative mood [1], 35 a persistent lack of motivation, and difficulty experiencing 36 pleasure that substantially affects patients' quality of daily 37 life [2]. Overall, 78% of patients with severe depression were 38 diagnosed with at least one comorbid psychiatric disorder, 39 such as psychotic disorder, past panic disorder, anxiety, and 40 even suicide risk [3], [4]. MDD is not a homogeneous 41 disorder but a complex disease with a variety of etiolo- 42 gies. Many studies have also shown dysfunctions in the 43 areas of the brain modulated by corresponding systems, 44 including the frontal cortex, amygdala, hippocampus, and 45 basal ganglia, in depression patients. These specific brain 46 regions are highly vulnerable to the effects of stress, prob-47 ably accounting for the adverse effects of life events on 48 MDD [5]. The severity of MDD is indexed by a com-49 posite of several behavioral measures or aspects of depres-50 sive perception. Therefore, clinicians score the degree of 51 depressive symptoms using the 17-item clinician-administered 52 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ Furthermore, investigators have recently utilized related 111 brain networks to distinguish the differences between patients 112 with MDD and healthy controls, assisting in providing a 113 better understanding of the mechanism underlying depression 114 [28], [29]. For example, in a previous study conducted by 115 Mohammadi and Moradi, the potential relationship between 116 the regional activity in patients with MDD and their depression 117 severity was not only identified but also provided a quantitative 118 depression severity prediction [30]. Gamma wave coherence 119 has also been found to help discriminate patients with mild 120 depression from healthy controls, as they manifested lower 121 gamma coherence than healthy controls [31]. However, studies 122 of the treatment response, especially short-term treatment 123 response, thus far, are not yet sufficient and still await 124 further investigation by performing EEG network analyses. 125 Consequently, the identification of robust predictors provides 126 significant benefits in terms of understanding and predicting 127 that variation [15]. 128 In the present study, the data we utilized were 129 downloaded from a public data archive, the National 130 Institute of Mental Health Data Archive (NDA). The 131 data are publicly available through the official website 132 (https://nda.nih.gov/edit_collection.html?id=2199). Recently, 133 multiple studies have been conducted using these datasets. For 134 example, Zhang and colleagues identified clinically relevant 135 MDD subtypes using (un)supervised machine learning based 136 on distinct network patterns [1]. Yu and colleagues investi-137 gated network differences within and between resting-state 138 networks in patients with MDD and healthy controls and 139 found that traumatic childhood experiences and dimensional 140 symptoms are linked to abnormal network architecture in these 141 patients with MDD [32]. Although these studies have been 142 implemented, the treatment response, especially short-term 143 treatment response, has not yet been studied extensively and is 144 still awaiting further investigation by performing EEG network 145 analyses. Therefore, in our present study, we analyzed the 146 resting-state EEG datasets from patients with MDD collected 147 before and after their one-week antidepressant medication 148 therapy. Related brain networks of these patients with MDD 149 were constructed and then statistically compared to explore 150 the brain fluctuations after one-week medication, as well as 151 predict its eight-week efficacy. The EEG data were selected from the study "Establishing 155 Moderators and Biosignatures of Antidepressant Response in 156 Clinic Care (EMBARC)" in the National Institute of Mental 157 Health Data Archive (NDA). In this study, patients were 158 recruited from different centers. However, in addition to 159 Columbia University, data from fewer qualified subjects were 160 obtained from the other centers due to the poor quality of 161 EEG data and failure of follow-up visits in the later stage of 162 medication therapy. Therefore, only the EEG data from 30 163 patients treated at Columbia University were included in our 164 current study. The 30 participants with MDD participated in a 165 randomized trial and received sertraline, and their clinical and 166 biological markers of outcomes were evaluated [33]. In detail, 167 the Clinical Global Improvement scale (CGI), and subjects 183 who received a CGI score worse than "much improved" 184 (i.e., having a CGI score greater than 2) were deemed non-185 responders, while the remaining patients who scored "much 186 improved" or "very much improved" (i.e., having a CGI score 187 of 1 or 2) were considered responders.  1) EEG Preprocessing: In the present study, we mainly 195 focused on investigating the potential capacity of the 196 resting-state brain network to quantitatively evaluate the 197 brain fluctuations after short-term medication and distinguish 198 responders from non-responders. Therefore, concerning these 199 resting-state EEG datasets, before preprocessing, the first and 200 last ten seconds of EEG signals were first excluded, and mul-201 tiple preprocessing procedures were then applied to complete 202 data preprocessing, which included averaging-referencing, [1], 203 [42] Hz bandpass filtering, and 5-s-length data segmentation. 204 Thereafter, a threshold of ±100 μV was subsequently used 205 to automatically exclude segments with absolute magnitudes 206 exceeding 100 μV from any recorded electrode. Additionally, 207 as denser electrodes might provoke more severe volume con-208 duction effects on connectivity, sparse electrodes were used 209 to reduce the effect of volume conduction on EEG networks. 210 Concretely, 21 of 72 channels, according to the international 211 10-20 system, were selected in our present study to perform 212 the analyses described below.

213
2) Functional Brain Networks: As proven in previous studies, 214 brain network analysis is remarkably helpful in explaining 215 the neurophysiological pathogenicity of MDD [2], and a 216 related analysis was thus implemented in our study. Actually, 217 an EEG network has been typically modeled as a collection 218 of nodes (i.e., EEG electrodes) and edges that are evaluated 219 between paired electrodes [34], [35]. When constructing the corresponding EEG network for all patients with MDD, the 221 same definition of the EEG network is considered. Specifically, 222 we set 21 electrodes as the network nodes, and after extracting 223 the EEG signal for each electrode, the estimated paired-224 electrode interactions were then regarded as the network edges.

225
Here, when constructing resting-state networks for patients 226 with MDD, the synchronization likelihood between pairwise 227 electrodes was considered [36], [37]. As suggested in previous 228 studies, the phase locking value (PLV) [38], [39] that experts in x(t) and y(t) as follows: where HT x (t) and HT y (t) are the HTs of both time series, 241 x(t) and y(t), which are defined as follows: where the P.V. denotes the Cauchy principal value. Afterward, 244 the corresponding analytical signal phases, i.e., ϕ x (t) and 245 ϕ y (t), were computed as follows: Finally, the PLV was formulated as follows: where w plv is the connection weight estimated using the PLV,  (1 W) or between responders and non-responders, respectively.

261
Furthermore, two weighted network properties were cal-262 culated using these constructed EEG networks, the clus-263 tering coefficient (C) and characteristic path length (L), 264 to quantitatively measure the network efficiency in processing 265 information. Here, these properties were calculated from the 266 weighted EEG networks without any thresholding processing. 267 Concretely, d i j represents the shortest weighted path length 268 between nodes i and j , n represents the number of network 269 nodes, and represents the total set of network nodes. The 270 two parameters were formulated as follows: Afterward, we statistically analyzed the potential differences 274 in these weighted network properties between responders and 275 non-responders using an independent t-test and between 0 W 276 and 1 W using a paired t-test, which was then corrected 277 for multiple tests using the Bonferroni correction to further 278 validate treatment response in patients with MDD.

) The Prediction of Medication Efficacy Using a Multiple 280
Linear Regression Model: In the present study, the changes 281 in the two resting-state network properties (i.e., C and L) 282 were selected as the variables in the multiple linear regression 283 model for building a model to predict medication efficacy. 284 Based on both C and L, the corresponding prediction 285 model was formulated as follows: where Y denotes the predicted eight-week medication efficacy, 288 β 0...2 denotes the regression coefficients of the network prop-289 erty changes, and ε denotes the error term.

290
Here, the leave-one-out cross-validation (LOOCV) strategy 291 was used to predict the eight-week medication efficacy in all 292 patients with MDD [40]. Specifically, for N samples (N = 293 30 in this study), N-1 samples were used for training, and 294 the remaining 1 sample was used for testing in each LOOCV 295 run. The regression coefficient for each variable was estimated 296 to build a prediction model for the current N-1 samples, 297 which was then used to predict the treatment outcome of an 298 individual in the test set. This procedure was repeated N times 299 until all samples served as testing sets one time.

4) Discrimination of Responders From Non-Responders 301
Based on Resting-State Networks: Eventually, these 302 resting-state networks were further tested to prove whether 303 they also promote the clinical selection of optimal therapeutic 304 strategies for MDD, which were validated using two different 305 types of network features. First, these resting-state network 306 properties were adopted. In detail, all of these patients with 307 MDD were divided into training and testing subgroups. 308 During the training process, the corresponding training 309 weighted network properties (i.e., C and L) were calculated 310 using Eqs. (5) and (6). Second, the linear discriminant 311 analysis (LDA) classifier was trained on these training 312 features. Thereafter, during the testing process, those 313 testing network properties were also calculated accordingly 314 and then input into the trained LDA classifier, which 315 finally discriminated the testing data into responders or 316 non-responders, and output the classification accuracy.

317
As proven previously, these network properties are direct 318 statistical measurements of brain networks. Although they 319 might also quantitatively capture the overall network dif-320 ferences under our conditions of interest, the corresponding 321 spatial network information is still unmined [41]. Thus, the 322 network property can be used to describe the overall brain 323 network but does not encompass all features contained in the 324 network topology. In our previous studies, we developed the 325 spatial pattern of the network (SPN) [42] to extract the dis-326 criminative spatial pattern contained in a given brain network 327 [42]. As the SPN was described in detail in our previous 328 studies [42], a brief introduction to this method was primarily 329 provided here.

330
M 1 and M 2 correspond to the adjacency matrices for respon-331 ders and non-responders estimated using PLV, respectively.

332
The SPN filters are the projections derived by maximizing 333 the following function: where z denotes the SPN filter (an objective projection) and A 1 336 and A 2 are the covariance matrices of the adjacency matrices 337 M 1 and M 2 , respectively.

338
Because the scaling of projection z will have no effect on  For multiple i SPN filters, Eq. (11) is solved as follows: where Z is composed of the eigenvectors of A where M denotes the adjacency matrix of patients with MDD, 362 z denotes the SPN filter (an objective projection), and V S P N 363 is a 21 × i matrix composed of SPN filters.

364
As clarified in previous studies, the increasing pairs of 365 SPN filters might facilitate the classification of situations of 366 interest [41], [42]; for example, different pairs of SPN features 367 (i.e., 1 pair, 2 pairs, and 3 pairs) have been used to achieve 368 the classification of psychogenic nonepileptic seizures from 369 epilepsy, and 3 pairs of SPN features achieved the highest 370 classification accuracy [42]. Consistent with the protocols used 371 in previous studies [43], [44], in our present study, three 372 pairs of SPN filters were accordingly adopted to achieve the 373 classification of responders and non-responders. In particular, 374 for a 21 × 21 adjacency matrix, M, each SPN filter was 375 a 21-length vector, and therefore, three pairs of SPN filters 376 comprised a 21 × 6 matrix. Afterward, the corresponding 377 SPN features were acquired as a vector with a length of 6 by 378 applying these SPN filters to the constructed resting-state EEG 379 networks and then calculating the variance of each row of 380 weighted nodes.

381
When using the SPN features to classify the responders 382 and non-responders, protocols similar to those used to ana-383 lyze network properties were performed. Specifically, after 384 acquiring the trained SPN filters during the training process, 385 we subsequently calculated the corresponding training SPN 386 features (Eq. (13)) that would be used to train the LDA 387 classifier. Additionally, the trained filters were further applied 388 to the testing sets to acquire the testing SPN features. Even-389 tually, the testing SPN features were input into the trained 390 LDA classifier, and the related classification performance was 391 evaluated accordingly.

392
As the present dataset was relatively small, the LOOCV test 393 was also used to recognize responders and non-responders, 394 as described in previous studies [45], [46]. Based on the 395 LOOCV, the corresponding indices, including accuracy (ACC), 396 sensitivity (SEN), and specificity (SPE), were then calculated 397 to evaluate the performance. Let N Res and N Non denote the 398 total number of responders and non-responders, respectively, 399 and let n Res and n Non denote the correctly discriminated 400 number of responders and non-responders, respectively. The 401 detailed equations used to calculate these indices were as 402 follows: Medication treatment visits occurred at baseline and 411 weeks 1, 2, 3, 4, 6, and 8 to ensure that the delivery was 412 appropriate and to record the HAMD 17 score. Although these 413 scales have some defects, they are still currently crucial for the 414  Fig. 2(b).  (i.e., 1 W), in which the red solid lines denote the reduced 458 interelectrode connectivity (0 W > 1 W). Specifically, rela-459 tively stronger connectivity (i.e., red long-range edges) among 460 the frontal and occipital lobes was observed at the baseline 461 stage ( p < 0.05, Bonferroni-corrected) than at 1 W. Thereafter, 462 the network properties, including C and L, were calculated and 463 compared to quantitatively measure the network fluctuation 464 after one week of treatment, as shown in Fig. 3(b), in which 465 the red and blue bars denote the network properties corre-466 sponding to 0 W and 1 W, respectively. Herein, to statistically 467 explore the potential differences, the linear mixed model was 468 adopted. In detail, the model included the network properties 469 (e.g., clustering coefficients) as the dependent variable, Time 470 (0 W and 1 W) and Group (responder and non-responder) 471 were then treated as independent variables. In addition, the 472 model included participants as the random factor. Models 473 were then compared using log-likelihood ratio tests to deter-474 mine the best model, and backward algorithms were used 475 for model comparisons. The results showed that the best-fit 476 model included the main effect of Group and Group-x-Time 477 interactions. Herein, when taking the network properties of 478 Week 0 and the non-responders as the baselines, the statistics 479 demonstrated that for responders, C showed a decreasing trend 480 at 1 W compared to 0 W, while L showed an increasing 481 trend. Although the non-responders received one week of 482 antidepressant medication treatment, significant differences 483 ( p > 0.05) were not observed either in network topology or 484 in properties between the two stages.

2) Prediction of the Medication Efficacy Based on 486
Resting-State Network Properties: Because these resting-state 487 network characteristics (both topologies and properties) 488 were identified to help evaluate the brain fluctuations after 489 one-week medication therapy, we then intended to explore 490 whether any potential relationship between the HAMD 17 491 score fluctuations and the corresponding network property 492 changes (C and L) would be identified. As shown 493 in Fig. 4, within the alpha band, C (r = 0.426, p = 494 0.019) showed a significant positive correlation with the 495 change in the HAMD 17 score at one week, while L was 496 negatively correlated with the change in the HAMD 17 score 497 (r = −0.428, p = 0.018). In fact, these patterns were 498 also observed within multiple bands, e.g., delta, theta, and 499 beta bands. As similar findings across multiple bands were 500 acquired and the activity of the alpha band in the brain at rest 501    clarified the smaller C and longer L for responders. At the one-544 week medication stage, the potential differences ( p < 0.05) in 545 both network topology and properties between the responders 546 and non-responders were further increased.

2) Categorization Into Responders and Non-Responders 548
Based on Resting-State Networks: Based on the statisti-549 cally significant difference in network properties between 550 responders and non-responders at the baseline stage shown 551 in Fig. 6(a), we next examined whether the corresponding 552 network properties also discriminated non-responders from 553 responders before the actual treatment, which might facilitate 554 the precise treatment of MDD in the clinic if this medication 555 therapy was effective in these patients with MDD. These 556 results might guide the design of a more effective treatment 557 protocol. In fact, although significant differences in network 558 properties were identified between both groups, the 76.67% 559 classification accuracy acquired here was not as satisfying as 560   And based on these merged participants, the prediction 600 was also achieved. Herein, by further using related network 601 Fig. 8. Classification of responders and non-responders based on related network topologies. The scatterplot corresponding to the SPN features extracted from the first pair of the most discriminating filters between the Responders and Non-responders. metrics to predict the efficacy after eight-week medication, 602 as displayed in Fig. 9, the predicted HAMD 17 changes were 603 indeed correlated with diagnosed ones (r = 0.50, p = 0.001, 604 RMSE = 6.31).

IV. DISCUSSION 606
The HAMD 17 is the most widely used clinician-607 administered depression assessment scale, containing 17 items 608 assessing depression experienced over previous weeks, which 609 provides an indication of depression and serves as a guide to 610 evaluate recovery. Considering the reliability and validity of 611 HAMD 17 [54], we analyzed HAMD 17 scores recorded in the 612 first two stages to illustrate the treatment efficacy after a rela-613 tively short-term medication period, as well as to explore if the 614 HAMD 17 was highly sensitive in identifying responders and 615 non-responders. Concretely, we sought to use the efficacy of 616 one week of medication to classify these patients with MDD. 617 As illustrated in our present study, although these patients 618 were divided into responders and non-responders based on 619 their eventual HAMD 17 scores recorded after eight weeks of 620 medication, no significant difference was observed when we 621 first explored the potential fluctuations in HAMD 17 scores 622 between 0 W and 1 W of medication. Therefore, the model 623 failed to describe the one-week treatment efficacy and further 624 discriminate responders and non-responders. As changes in 625 the HAMD 17 score did not show an obvious difference in 626 distinguishing responders from non-responders, the HAMD 17 627 scores were then directly compared between responders and 628 non-responders at the two time points mentioned above.

629
Unfortunately, the responders did not show a significant dif-630 ference from the non-responders at either stage, as shown in 631 Fig. 2(b). These results might be attributed to the fact that 632 the short-term treatment was not long enough to significantly 633 affect the brain of patients with MDD and failed to alter occipital lobes (Fig. 3(a)), as well as smaller properties 654 ( Fig. 3(b)) than their 0 W measures. As reported in previous and executive demands and focus attention [63], [64], [65].

671
The increased beta activity was related to a deterioration in 672 cognitive flexibility and control [66]. Therefore, the decreased  Moreover, resting-state brain activity has been proven to 678 comprise the basis of the related cognitive process [67], [68], 679 and many studies have performed related resting-state analyses 680 when investigating MDD dysfunction [69], [70]. In fact, 681 the increased resting-state multiregional synchronization in 682 patients with MDD has been proven to be accompanied 683 by increased self-rumination, which is considered a princi-684 pal cause of the psychophysiology of depression [71], [72]. 685 Antidepressant medicines are commonly utilized to enhance 686 monoaminergic neurotransmission and reverse some of these 687 stress-induced neurophysiological changes, further inhibiting 688 the abnormal activity of the amygdala, and are thus proposed 689 to be helpful for MDD therapy [73], [74]. Here, the decreased 690 network topologies and properties consistently clarified that 691 one week of medication worked for responders by significantly 692 alleviating their overall connectivity. Unfortunately, although 693 receiving the same therapeutic intervention, the networks 694 of those non-responding patients with MDD remained in 695 their initial state, as no significant differences were observed 696 between the 0 W and 1 W sessions. Here, the statistics relying 697 on the linear mixed model did report both main effects of 698 Group and Group-x-Time interactions for the explorations 699 into the brain networks. The decreased network parameters 700 consistently clarified that one week of medication worked for 701 responders by significantly alleviating their overall connec-702 tivity. Unfortunately, although receiving the same therapeutic 703 intervention, the networks of those non-responding patients 704 with MDD remained in their initial state, as no significant 705 differences were observed between the 0 W and 1 W sessions. 706 And considering these findings, we thus assumed that the 707 network investigation performed in our present study might 708 be sensitive to capture the corresponding brain fluctuations 709 occurring after one week of medication.

710
As resting-state networks were validated to quantitatively 711 measure brain fluctuations after short-term treatment in 712 patients with MDD, even within one week, the corresponding 713 network metrics (e.g., network properties) were thus postulated 714 to be robust biomarkers to predict therapeutic efficacy in these 715 patients. Changes in both C and L (Fig. 4) were significantly 716 correlated with fluctuations in HAMD 17 scores between base-717 line and one week of medication therapy, which primarily 718 illustrated the possibility of the subsequent prediction analy-719 sis. Accordingly, the network properties were then utilized 720 to predict the long-term treatment outcome. Concretely, the 721 changes in network properties calculated by subtracting the 722 network properties at 1 W (after one week of medication) 723 from those at 0 W (baseline) were selected as the predictors 724 of the eight-week antidepressant treatment response. Fig. 5 725 shows the scatterplots of the actual and predicted changes 726 in the HAMD 17 scores for all of these patients with MDD, 727 where the dashed diagonal line indicates the ideal prediction, 728 and the blue filled circles distributed along the dashed line 729 denote that the regression model estimated from the train-730 ing set was capable of accurately predicting an individual's 731 eight-week antidepressant outcome. In addition, considering 732 the close correlation between network property changes and 733 HAMD 17 score fluctuations identified in the previous analysis, 734 this robust prediction of antidepressant outcome further vali-735 dated that the network properties (C and L) indeed served as 736 influential features to predict individual long-term medication 737 efficacy and verified the reliability of resting-state networks in 738 promoting personalized medication strategies.

739
Due to the significant difference between 0 W and 1 W 740 observed in responders, whereas no difference was observed 741 in non-responders between these two stages, we further com-742 pared and analyzed the significant differences between respon-743 ders and non-responders at baseline and after one week of 744 medication. Further analysis of the data from our current 745 study revealed that brain networks illustrated the difference 746 between both groups, even before antidepressant treatment.

747
Additionally, this difference between the two groups increased 748 significantly after one week of the therapeutic intervention.

749
Concerning the pretreatment comparison, as displayed in 750 the left panel of Fig. 6(a), non-responders showed enhanced 751 network connectivity compared to responders, as mainly 752 manifested as long-range connectivity between the tempo-753 ral lobe and occipital lobe. Corresponding network proper-754 ties further quantitatively revealed stronger brain activity in 755 non-responders than in responders. As validated in previous 756 studies [75], [76], [77], [78], non-responders experienced network nodes by emphasizing those important nodes with 796 larger coefficients but suppressing others with much smaller 797 coefficients (close to zero). As displayed in Fig. 7, those 798 occipital electrodes (e.g., O1, O2, and Oz) are shown in a 799 deep red or blue color and implied that great differences 800 occurred at these electrodes, which indeed coincided with the 801 topological differences shown in Fig. 6(a). In fact, based on 802 these varying strategies, the SPN succeeds in extracting the 803 network spatial information exactly and guarantees its capacity 804 in classifying patients with MDD. Here, satisfactory perfor-805 mance was achieved, as the accuracy was further improved 806 to 96.67% when using these SPN features as discriminative 807 features. Notably, this classification is based on the EEGs 808 recorded before the actual treatment, and it is very helpful to 809 instruct the clinician in designing a more efficient therapeutic 810 protocol for patients with MDD. Further, to evaluate the 811 generalizability of our method, a total of 80 participants from 812 four different sites were picked and then included in the 813 categorization analysis. The recognition of responders from 814 non-responders also obtained acceptable performance, as the 815 accuracy of 77.50% was achieved when using SPN features 816 as the discriminative features. Additionally, based on these 817 merged participants, we further used related network metrics to 818 predict the efficacy after eight-week medication; the predicted 819 HAMD 17 changes were indeed correlated with diagnosed 820 ones (r = 0.50, p = 0.001, RMSE = 6.31). These results 821 further validated the generalizability of our proposed method 822 in treatment selection of MDD.

823
Considering these findings described above, the network 824 changes were indeed observed for the responders but not 825 for the non-responders after short-term antidepressant treat-826 ment; and the differences between the two groups became 827 even greater after one-week medication. Replicated evidence 828 indicated that according to the HAMD 17 scale, antidepressant 829 efficacy was not observed obviously until at least 4 weeks of 830 medication [80]. However, compared with the clinical scale, 831 the brain network seems to be a more sensitive biomarker and 832 has a great capacity for evaluating brain changes after short-833 term medication, as well as distinguishing responders from 834 non-responders. Although the exact relationship between EEG 835 metrics and clinical scales was still unveiled, related EEG met-836 rics Li et al.; Zhang et al. have been proved to reliably index 837 the deficits occurring in brain diseases, as well as evaluate 838 the treatment response from patients. Following the similar 839 protocol used previously [81], [82], in this study, we thus 840 utilized brain networks to quantify the brain fluctuations after 841 one-week medication for MDD patients, and further explored 842 the relationship between EEG metrics and clinical scale, which 843 aimed to provide potential early biomarkers for subsequent 844 prediction analysis and treatment selection. 845 However, although we have verified our method on the 846 EMBARC study and achieved good performance, the con-847 clusions should be further verified, and the potential capacity 848 for big data analysis should also be confirmed. Meanwhile, 849 we will then develop related algorithms to mitigate the site 850 effect induced by different amplifiers and/or electrode mon-851 tages to improve the generalizability and robustness of our 852 method.