Resting-state EEG delta and alpha power predict response to cognitive behavioral therapy in depression: a Canadian biomarker integration network for depression study

Cognitive behavioral therapy (CBT) is often recommended as a first-line treatment in depression. However, access to CBT remains limited, and up to 50% of patients do not benefit from this therapy. Identifying biomarkers that can predict which patients will respond to CBT may assist in designing optimal treatment allocation strategies. In a Canadian Biomarker Integration Network for Depression (CAN-BIND) study, forty-one adults with depression were recruited to undergo a 16-week course of CBT with thirty having resting-state electroencephalography (EEG) recorded at baseline and week 2 of therapy. Successful clinical response to CBT was defined as a 50% or greater reduction in Montgomery-Åsberg Depression Rating Scale (MADRS) score from baseline to post-treatment completion. EEG relative power spectral measures were analyzed at baseline, week 2, and as early changes from baseline to week 2. At baseline, lower relative delta (0.5–4 Hz) power was observed in responders. This difference was predictive of successful clinical response to CBT. Furthermore, responders exhibited an early increase in relative delta power and a decrease in relative alpha (8–12 Hz) power compared to non-responders. These changes were also found to be good predictors of response to the therapy. These findings showed the potential utility of resting-state EEG in predicting CBT outcomes. They also further reinforce the promise of an EEG-based clinical decision-making tool to support treatment decisions for each patient.


Results
Demographic and clinical characteristics. In total, forty-one participants were recruited for the study.
Thirty-seven completed clinical assessments at baseline and at the end of the therapy. Thirty participants completed EEG sessions at both baseline and week 2 (three participants missed the EEG session at baseline, and four participants missed the EEG session at week 2) and were included in the analysis (Table 1).
Of the thirty participants considered in this study, sixteen (53%) were responders, and fourteen (47%) were non-responders. Groups did not differ by sex (p = 0.134) or age (p = 0.400). However, responders and nonresponders differed by ethnicity (p = 0.011). No differences were observed between groups in MADRS scores at baseline and week 2 (p = 0.641 and p = 0.615, respectively).
Relative power EEG spectra. The average relative power EEG spectra for responders and non-responders at baseline and week 2 can be found in Supplementary Fig S1. Relative power measures showed excellent internal consistency with an averaged Cronbach's alpha across frequencies and electrodes of 0.911 at baseline, and 0.925 at week 2, respectively (Supplementary Fig S2). was also observed across several frequencies and regions of interest in the source space. As such, several post-hoc analyses were conducted. In what follows, the results of post-hoc t tests, correlational and predictive analyses are presented for baseline, week 2 and change in baseline to week 2. For each analysis, cluster-based permutation testing was applied to correct for multiple comparisons.
Relative power at baseline. At baseline, post-hoc analysis revealed a group difference in delta band, with responders exhibiting lower delta activity compared to non-responders (negative cluster, p = 0.016, Cohen d = 0.882, Fig. 1). In the source space, this difference was identified in several brain areas, including bilateral precentral gyrus and sulcus, bilateral central sulcus, bilateral postcentral gyrus and sulcus, bilateral precuneus, bilateral superior and inferior parietal gyri, bilateral middle occipital gyri, left midcingulate cortex (MCC), and bilateral posterior cingulate cortex (PCC) (Fig. 1).
The difference observed between groups in the delta band at baseline was found to be predictive of the treatment outcome (AUC = 0.754, p = 0.006 in the sensor space, and AUC = 0.754, p = 0.006 in the source space, Fig. 2).
Results were similar when including all participants who completed the EEG session at baseline (thirty-four participants) and can be found in Supplementary Figs S3 and S4.
Relative power at week 2. At week 2, post-hoc analysis did not reveal any differences between groups in relative power. There was also no significant correlation between relative power at week 2 and improvement in depressive symptoms.
Results did not differ when including all thirty-three participants who completed the EEG session at week 2 (thirty-three participants).
Early changes in relative power. Post-hoc analysis revealed group differences in delta and alpha bands, with responders exhibiting an increase in delta activity (positive cluster, p = 0.020, Cohen d = 1.118, Fig. 3) and a decrease in alpha activity (negative cluster, p = 0.014, Cohen d = 0.949, Fig. 3) relative to non-responders. Both differences were localized to similar brain regions, including left precentral, postcentral, and superior parietal gyri, left central sulcus, left precuneus, and left MCC and PCC. In addition, decrease in alpha was also localized to right precentral gyrus and left inferior parietal gyrus (Fig. 3).
Early changes observed in delta and alpha activity were good predictors of response to CBT. The significant prediction models had the following characteristic: early changes in delta activity model (AUC = 0.848, p = 0.001 www.nature.com/scientificreports/  Here, specificity corresponds to percentage of non-responders who were predicted to be non-responders, and sensitivity corresponds to percentage of responders who were predicted to be responders. The red circle shows the optimum operating point of the ROC curve. www.nature.com/scientificreports/ in the sensor space, and AUC = 0.799, p = 0.002 in the source space, Fig. 4), and early changes in the alpha activity model (AUC = 0.830, p < 0.001 in the sensor space, AUC = 0.768, p = 0.007 in the source space, Fig. 4).

Discussion
In this study, we showed that resting-state EEG delta and alpha power might have possibly moderate to strong predictive utility in predicting response to CBT for adults with depression. A reduced baseline delta power was observed in responders and was predictive of response to CBT. Responders also exhibited an early increase in delta power, which also proved to be a good predictor of response to the therapy. While previous authors have also reported an association between slow oscillations and response to other antidepressant treatment modalities 20,21,23-28 , results are not unidirectional and may be related to different effects and mechanisms of antidepressants and different subtypes within depression. Here, the findings in the delta activity might reflect cognitive control changes, which have been attributed to the mechanisms of CBT [53][54][55][56][57] . Previous studies have shown the functional significance of delta activity during cognitive processing [59][60][61][62][63][64][65][66][67] . For example, delta oscillations may play a role in motivational and emotional processes 61,66 and also in inhibitory control and response inhibition 59,60,63,65,67 . An increase in delta activity may also be an indicator of attention to internal processing during the performance of a mental task 62 . Regarding delta activity at rest, some previous studies reported the relevance of low-frequency intrinsic activity in the neural mechanisms of cognitive control Figure 3. Differences in relative power early changes between responders and non-responders. Cold colors show lower relative power changes in responders compared to non-responders. Warm colours show higher relative power changes in responders compared to non-responders. (A,B) The x-axis shows frequencies from 0.5 to 50 Hz. The y-axis shows all electrodes from 1 to 58. Image A shows uncorrected t-value map, image B shows significant clusters (p < 0.025, single-tailed) using cluster-based correction. (C,E) Topographies illustrate t-values at different frequencies with stars indicating electrodes that belonged to the significant cluster. (D,F) Cortical maps depict source-localized regions (p < 0.05) in the frequency band in which the cluster was found at the sensor space level. www.nature.com/scientificreports/ processes [68][69][70] . In these studies, resting-state delta power was shown to be associated with behavioral performance and neural correlates of response inhibition during a go/no-go task. In another recent study, resting-state delta oscillations were predictive of good performance on an attentional set-shifting task 71 . The results here could suggest that CBT enhances cognitive control manipulation processes in patients who display an increase in delta activity after 2 weeks of therapy. Moreover, some studies involving meditation also reported an increase in delta activity during the suppression of interference 72,73 . In these studies, stronger delta activity in the medial prefrontal cortex, during the eyes-closed resting state, was suggested to reflect a reduction of emotional and cognitive engagement. The findings here could suggest that patients with stronger emotional engagement at baseline represent a subgroup who responds well to CBT. This correlates with a study in which patients presenting excessive attention toward aversive information were more likely to respond to CBT 74 . Increased attention to aversive stimuli is suggested to be a mechanism underlying emotional disorders in patients with depression 75 . The early increase in delta observed here in responders to CBT could therefore reflect a mechanism whereby CBT reduces negative emotional engagement that interferes with decision making and situational goals. Furthermore, when investigating if differences observed in early changes were associated with improvements in specific symptoms using subdomains of MADRS, early increases in delta were found to be related to improvement in sadness, and negative thoughts subdomains (Supplementary Fig S5). In comparison, early decreases in alpha were not related to any subdomains. While further analyses are necessary to better understand the role of delta oscillations, these results support the potential link between delta activity and cognitive control when processing negative information. A significant difference was also found in the early changes in relative alpha power, with a decrease observed in responders. This change was found to have strong utility in predicting response to CBT. Reduction in alpha power has been shown to be a mechanistic marker of response to various modalities of antidepressant treatments Here, specificity corresponds to the percentage of non-responders who were predicted to be nonresponders, and sensitivity corresponds to the percentage of responders who were predicted to be responders. The red circle shows the optimum operating point of the ROC curve. www.nature.com/scientificreports/ (including medications, rTMS and ECT) across studies 20,21,28,34,47 . There is an inverse relationship between alpha power and cortical activity 76,77 . Hence, a decrease in alpha power could reflect an increase in cortical arousal toward the processing of external stimuli and cognitive engagement instead of an internally self-focused emotional processing [78][79][80][81] . Moreover, as this change was primarily over regions in left hemisphere, it could also indicate that cognitive therapies, which are highly verbal treatments, may involve cognitive processes meditated by regions in left hemisphere. This is in line with studies that found a left hemisphere advantage for verbal processing in responders to CBT 50-52 . Altogether, it is possible that CBT may be better suited for individuals with overactive cognitive presentations. In this study, responders to the therapy were found to be more likely to exhibit lower delta activity and higher alpha activity, which could indicate greater inward-focused and self-referential processing. By using CBT to alter thought patterns, it might be possible to induce changes in brain activity, leading to improved mood and emotional states.
This study has certain limitations worth noting. First, considering the small sample size, any generalization from the findings should be made with caution. It also precludes the analysis of gender-based effects or other potential moderators such as ethnicity or clinical features. Secondly, placebo effects are unknown as findings were obtained during an open-label treatment. Replication of the results in a larger sample size and with a treatment control group is required. It would also inform the quantification of the differences in EEG predictors across treatment modalities.
In summary, this is the first study to show findings possibly indicated a potential utility of resting EEG in predicting response to CBT for adults with depression. The findings suggest that baseline measures and early changes in delta and alpha frequency bands might be used to classify responders and non-responders to CBT. While baseline measures might not be enough to predict CBT treatment outcomes accurately, information that arises during the early course of therapy was found to be a strong predictor of response to CBT. Given the limited access to in-person CBT therapy and the limited number of CBT therapists, being able to make an informative decision after the two first weeks of therapy about whether to continue or switch to another treatment would have great clinical value. EEG technology is becoming more accessible with more portable and less expensive devices 82,83 , and research investigating EEG-based biomarkers can inform its strategic application in clinical contexts. Future studies are required in order to validate the results, but the insight gained here may help the development of objective clinical decision tools to guide individual patients with MDD to the optimal treatment.

Methods Participant sample. Recruitment. Participants were recruited at the Centre for Addiction and Mental
Health (CAMH) using a research registry, internal referrals, and waitlists for the CAMH Mood and Anxiety Program. All research was conducted in accordance with the Declaration of Helsinki, and all participants gave informed written consent prior to enrolment in the study. The protocol was approved by the Centre for Addiction and Mental Health Research Ethics Board and was registered with clinicaltrials.gov (https:// clini caltr ials. gov/ ct2/ show/ NCT02 883257). CBT. CBT was provided on an individual basis by a registered psychologist or graduate-level trainee under the direct supervision of a registered psychologist. Participants received 20 sessions over approximately 16 weeks, with 2 sessions per week in the first 4 weeks and 1 session per week in the remaining 12 weeks. CBT was delivered according to established protocols which comprised behavioral activation and cognitive restructuring. Optional elements such as coping and social skills training, perfectionism, and self-criticism, were also included to address individual maintaining factors. Some treatment sessions were audio-recorded and reviewed by an independent evaluator using the Cognitive Therapy Scale-Revised (CTS-R) to assess therapists' adherences. Scores met the established threshold of acceptable competence in delivering CBT (CTR-S: 50.43 ± 4.26). The primary outcome was the Montgomery-Åsberg Depression Rating Scale (MADRS) to assess depression severity. It was administered by the study coordinator and was completed at baseline and every second week during the therapy. EEG data recording. EEG data were collected at week 0, week 2, and week 16 of treatment, but only data at baseline (week 0) and week 2 were used in this study. EEG data were recorded using a 64-channel SynAmps 2 EEG system (Neuroscan, Compumedics USA, USA). The International 10-10 system was used for the placement of electrodes. The impedance level of each electrode was lowered to ≤ 5 kΩ, and an electrode positioned posterior to the Cz electrode was used as a reference. Recordings were done at a 10 kHz sampling rate, with a direct current and a low-pass filter. Five minutes of EEG activity were recorded from all participants during eyes-closed resting condition. EEG data processing. EEG data were preprocessed using EEGLAB toolbox 84 . Datasets were first standardized to those collected in previous CAN-BIND studies for potential future comparisons. The standardization involved the selection of 58 common channels in all datasets, a re-referencing to average, and a resampling to 512 Hz. A customized, fully automatic pipeline was then used to clean the data. This pipeline was adapted from the ERPEEG toolbox 85 . First, data were high-pass filtered at 0.5 Hz with a second-order Butterworth IIR filter. www.nature.com/scientificreports/ The EEGLAB plugin clean_rawdata was then used to detect bad channels and bad segments of data. In this step, bad channels were deleted, and bad segments were corrected. Power line artifacts were removed using ZapLine method 86 . Independent component analysis (ICA) was then conducted. Data were temporarily high-pass filtered at 1 Hz for better ICA decomposition 87 , and the ICLabel algorithm 88 was used to remove components associated with recurring artifacts such as eye movement, eye blinks, muscle, and cardiac artifacts from the original data filtered at 0.5 Hz. Finally, bad channels were interpolated using spherical interpolation, and the data were again re-referenced to average. EEG power spectral density. Using Welch's method, periodograms derived from 2-s non-overlapping windows were averaged to get an estimated power spectral density (absolute power) of each channel from 0.5 to 50 Hz with a frequency resolution of 0.5 Hz. The relative power was obtained by taking the ratio of the absolute power at each frequency and the total sum of absolute power. Internal consistency of relative power measures was also estimated (Supplementary Information).
EEG source localization. EEG sources were reconstructed using Brainstorm software 89  Statistical analysis. Demographic and clinical data were compared between responder and non-responder groups using independent-samples t test or Chi-squared test, wherever appropriate. Analysis of variance was performed to test the differences in relative power (0.5-50 Hz frequencies) between responders and non-responders at baseline and week 2. The main effect of CBT Response (responder, nonresponder) and Time (baseline, week 2), and the interaction effect of CBT Response x Time were evaluated across 58 channels in sensor space and 148 regions of interest in source space. For post-hoc comparisons, when applicable, independent-sample t-statistics were used across all 58 channels and across frequency bands: delta band (0.5-4 Hz), theta band (4-8 Hz), alpha band (8-12 Hz), beta band (12)(13)(14)(15)(16)(17)(18)(19)(20)(21)(22)(23)(24)(25)(26)(27)(28)(29)(30), and gamma band . Cluster-based permutation testing 93 was applied separately within each frequency band to further correct for multiple comparisons. Initial t-values exceeding an a priori threshold of p < 0.05 were clustered together based on adjacent frequency bins and neighboring electrodes. A minimum of 2 neighboring electrodes was considered for a selected sample to be included in the cluster. All t values within every cluster were summed to build clusterlevel statistics. Finally, using a Monte Carlo method with 2000 permutations, a distribution of the maximum cluster-level statistics was obtained, and the significance of each cluster in the original data was set at p < 0.025 (single-tailed). In source space, the average value of the relative power was calculated within frequency bands in which a significant cluster was found at the sensor space level. Post-hoc independent-sample t-statistics were then conducted across all 148 regions of interest and t-values exceeding a threshold of p < 0.05 were considered significant. As correction for multiple comparisons was conducted at the sensor level, and our hypotheses were only about the existence of group differences with no a priori specific brain area being involved, correction for multiple comparisons was not required at the source level 94 .
Further analyses were performed with subdomains of MADRS score 95 . More details are given in the Supplementary Information and the results can be found in Supplementary Fig S5. Predictive analysis. Logistic regression models were used to assess the predictive value of relative power to classify participants' responses to CBT. Logistic regression models were only assessed for relative power measures that were significantly different between responders and non-responders. In sensor space, an average of the relative power measures in the related significant clusters was calculated to obtain a single value per patient. In source space, a single value was obtained for each frequency band by averaging of the relative power across the regions of interest found significantly different. The level of prediction of these single values was quantified by the receiver operating characteristic (ROC) curve, plotting the sensitivity and the sensibility and across all possible thresholds. The area under the curve (AUC) was calculated to determine the significance of the prediction.
Based on previous studies showing that early improvement in symptoms predicts better treatment outcomes [96][97][98] , we also examined the predictive value of reduction in MADRS score from baseline to week 2 to classify participants' responses to CBT as an exploratory analysis. The results can be found in Supplementary Figs S6 and S7.

Ethics declaration.
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.

Data availability
Data can be made available upon reasonable request to the corresponding author. Each request will be processed in consultation with the related research ethics boards and institutional data sharing policies and guidelines. www.nature.com/scientificreports/