Major Depressive Disorder (MDD) is a severe, widespread and often recurring psychiatric illness that is primarily characterized by a loss of experiencing pleasure, persistent sad mood, sleep disturbances, changes in appetite and impaired concentration.1 The lifetime prevalence of MDD is estimated around 20% and 30% for men and women respectively.2 In addition, the societal and economic burden of MDD is considerable due to an increase in absenteeism, alcohol- and drug related issues, suicide attempts and general illness comorbidity, making MDD a leading cause of disability.3 It is therefore imperative to focus research efforts on the development of reliable and practical biomarkers of MDD that can be applied in clinical settings.
In the past decade, research investigating neural correlates of MDD has shifted focus from identifying specific dysfunctional brain regions to examining intrinsic neural networks implicated in depression.4 One of these networks is the cognitive control network (CCN) which comprises the dorsolateral prefrontal cortex (DLPFC), the anterior cingulate cortex (ACC) and the parietal cortex.5–7 The CCN’s main function is to regulate cognition and behavior in the pursuit of internal goals by way of directing attention towards task relevant stimuli while simultaneously inhibiting task irrelevant stimuli.8,9 In patients with depression the CCN’s connectivity is diminished, resulting in difficulties with both sustained attention and the downregulating of negative emotions.1,10,11 There is mounting evidence showing the involvement of the CCN in regulating both positive and negative emotions through indirect top-down connections with limbic regions in which the subgenual ACC (sgACC) serves as a gatekeeper between the cognitive- and emotional network.12–17 In short, the sgACC has dense connections with the amygdala and other regions of the limbic system,18–20 constituting a brain network involved in emotional processing.4,21,22 Moreover, the sgACC seems to project information from this affective network to the CCN’s frontal cortical regions,1,12,13 resulting in top-down emotion regulation.
Abnormalities in the DLPFC and sgACC have been consistently found in MDD patients. Specifically, the sgACC seems hyperactive in depression and successful treatment leads to a normalization of this hyperactivity.23–25 In contrast, the DLPFC has been found to be hypoactive in MDD patients26–29 and restoring DLPFC activity seems to elicit an antidepressant response.29,30 As a result, both the left- and right DLPFC have become popular targets for noninvasive neurostimulation techniques such as Transcranial Magnetic Stimulation (TMS) in the treatment of severe depression.31–36 Fox et al. posit that the left DLPFC and the sgACC are intrinsically anticorrelated during rest and that this anticorrelation is exacerbated in MDD.24 Consistent with this notion, a neurostimulation study from Baeken and colleagues found stronger sgACC–DLPFC anticorrelations in responders before high frequency TMS treatment while observing a normalization in sgACC–DLPFC connectivity after remission.23 Interestingly, applying low frequency TMS to the right DLPFC also seems to normalize sgACC hyperactivity34 and a meta-analysis that compared the clinical efficacy of both treatments found them equally effective.31 Consistent with these neurostimulation findings, functional Magnetic Resonance Imaging (fMRI) studies reported reduced functional connectivity between the DLPFC and the dorsal ACC in late-life depression patients during an executive-control task37 and when at rest.38 Taken together, aberrant sgACC–DLPFC connectivity seems to be a robust component of MDD, making it a promising candidate as a biomarker for depression.
A reliable biomarker should be an objective measurement of the physiologic or pathologic processes underlying a biological trait or illness.39 Moreover, a biomarker that is being applied in diagnostics or treatment outcome should be relatively accessible for it to have any clinical relevance. Indeed, a biomarker has little clinical value if the operational costs are too high or the equipment necessary for measuring it is too scarce. For example, even though MDD biomarkers that are based on fMRI studies produce valuable information regarding the neurobiological underpinnings of the disease, they are seldom used in clinical practice. Since acquiring and operating MRI’s is expensive, hospitals will simply prioritize more urgent clinical matters than an MDD diagnosis. In contrast, the electroencephalogram (EEG) is a relatively affordable, time-efficient, and commonly available neuroimaging tool that is already routinely used in psychiatric and neurologic departments during a patient’s hospital admission.40 Furthermore, EEG’s temporal resolution is superior when compared to most other neuroimaging techniques. For example, fMRI’s Blood Oxygenation level-dependent (BOLD) functional connectivity is based on the relatively slow hemodynamic response, restricting its temporal resolution to around 1Hz. EEG can measure the brains electrical activity with a precision of milliseconds, resulting in a temporal resolution of more than 1000Hz depending on the sampling rate capabilities of the amplifier. This allows researchers to estimate the functional connectivity of neural oscillations in real time over an extensive frequency range which can reveal unique electrophysiologic frequency signatures of neural processes.
One notable disadvantage of EEG is the low spatial precision which arises from the diffusion of the electrical signals caused by volume conduction of the skull.41 In brief, electrical sources in the brain are projected on the scalp which are then measured by the EEG’s electrodes. However, volume conduction and the mixing of electrical signals introduces spatial artifacts, distorting scalp projections. For instance, a relatively small source localized in the occipital cortex could present as a large frontal projection.42 These spatial artifacts are especially problematic when estimating functional connectivity since, diffused electrical signals measured by different electrodes could originate from the same neural source which would result in spurious connectivity values.43 Fortunately, great strides have been made to minimize spatial artifacts stemming from volume conduction. Advancements in source estimation techniques such as the development of realistic head models,44,45 high-density EEG (whole-head electrode coverage)46 and more reliable linear inverse solutions47 substantially improve the spatial accuracy of electrophysiological source models.48 Consequently, numerous methods to estimate EEG resting state functional connectivity have been developed.49 A study from Colclough and colleagues reports that out of 12 electromagnetic connectivity measures, amplitude envelope correlation (AEC) and partial correlation measures have the best intra-subject and between group consistency;50 while another study found AEC to best mirror connectivity results obtained using fMRI,51 making AEC an excellent measure of connectivity that can be compared across modalities.
While some EEG studies have looked at general spectral signatures of aberrant network functional connectivity in depression,52,53 none have investigated whether disturbances in resting state sgACC–DLPFC functional connectivity could be utilized as a potential biomarker for MDD. Therefore, the aim of the current EEG study was twofold. Firstly, to replicate fMRI and neurostimulation studies that observed disturbances in connectivity between the DLPFC and the sgACC in depression and secondly, to evaluate whether this EEG sgACC–DLPFC functional connectivity has any reliable biomarker capabilities. The latter goal was attempted by training a support vector machine (SVM) on the estimated sgACC–DLPFC connectivity values and subsequently test the SVM performance to reliably distinguish individual MDD patients from healthy controls based on these connectivity values. Compared to other machine learning methods, SVM has some unique properties that are advantageous in the context of identifying psychiatric biomarkers such as the ability to analyze high-dimensional datasets with small sample sizes.54 We chose to include a supervised machine learning approach, since accurately identifying MDD patients should be an essential attribute of any clinically relevant biomarker of depression.