Resting-state EEG delta and alpha power predict response to cognitive behavioral therapy in depression: a Canadian biomarker integration network for depression study · Schwartzmann B, Quilty LC, Dhami P et al.

Methods

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://clinicaltrials.gov/ct2/show/NCT02883257).

Eligibility for CBT

Participants had to be 18 years or older to be enrolled in the study. Inclusion criteria included: (1) clinical diagnosis of MDD or Persistent Depressive Disorder according to the DSM-5; (2) fluency in English; (3) capacity to give informed consent. The exclusion criteria can be found in the Supplementary Information.

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 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm$$\end{document}± 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 toolbox84. 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 toolbox85. First, data were high-pass filtered at 0.5 Hz with a second-order Butterworth IIR filter. 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 method86. Independent component analysis (ICA) was then conducted. Data were temporarily high-pass filtered at 1 Hz for better ICA decomposition87, and the ICLabel algorithm88 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 software89. For each participant, eyes-closed resting condition EEG data were imported into Brainstorm software89. First, the locations of the 58 EEG channels were co-registered to the ICBM152 MRI template. The OpenMEEG BEM head model90 was then used to calculate the forward model. An identity matrix was used as noise covariance. The inverse problem was derived from the minimum norm estimate model sLORETA91, and the solution space was constrained to the cortex surface. In the MNI coordinate space, the Destrieux Atlas92 was used, providing 148 cortical regions that served as regions of interest. Finally, relative power measures were computed at these 148 reconstructed sources for all subjects.

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, non-responder) 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 (12–30 Hz), and gamma band (30–50 Hz). Cluster-based permutation testing93 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 cluster-level 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 level94.

Further analyses were performed with subdomains of MADRS score95. 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 outcomes96–98, 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.

All analyses were performed using MATLAB (Version R2020b; MathWorks, Natick, MA, USA).

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.