The systematic review was conducted according to the PRISMA guidelines (Page et al., 2021). We focused on finding representative original studies on resting-state EEG biomarker research for unipolar depression, resp. MDD diagnosis or recognition rather than treatment in order to keep the use case concise. The search was conducted in the PubMed database on October 3^rd 2024 since we expect clinically relevant studies to be published in an outlet indexed in PubMed. The search string (‘major depressive disorder’ OR MDD) AND (biomarker OR diagnosis OR detection) AND (electroencephalography OR EEG) NOT (treatment OR TMS OR ‘transcranial magnetic stimulation’ OR sleep) was used with a filter on the last 5 years to include only recent research. Papers were excluded if they met at least one of the following exclusion criteria: (1) the aim of the paper was not diagnosis or detection of MDD; (2) MEG or other image modalities were used; (3) task-EEG, event-related potentials, or sleep EEG was recorded; (4) the purpose of the study was data augmentation; (5) the study contains only bipolar patients or a specific subgroup of MDD; (6) the study does not have a HC group; (7) the paper is a review; (8) the paper is not published in English or not accessible.

We extracted information about the three categories demographic data, diagnosis and medication, and neuropsychological tests to characterize the study populations as detailed as possible. Demographic data includes the size of the study population, age, gender, and ethnicity as smallest set of overlapping information across studies. Diagnosis includes the procedures to obtain the labels MDD and HC including exclusion criteria for participants. Regarding medication, we included all information on drugs available from the methods or results of the studies. Neuropsychological tests assess the severity of psychiatric diseases, symptom dimension, or cognitive function, usually with a (self-administered) questionnaire. They can be used for screening, complement diagnosis, or provide a more detailed description of a study population. For all three categories, we also extracted information about the consideration of these factors in the analysis and their impact. Since a substantial number of studies is based on publicly available data sets, we separate the information of studies based on own data from studies based on public data.

For further demonstration of the influence of the factors investigated in the review and the heterogeneity of a study population when considering further data about the participants, we conduct exploratory analyses on two publicly available data sets that are also used in some of the studies presented in this review (for details see chapter 3.2).

The CAV data set (Cavanagh et al., 2019) provides rich diagnostic information, including two different neuropsychological tests for the assessment of depression severity, which allows for the comparison of different diagnostic scenarios. The MODMA data set (Cai et al., 2022) provides six neuropsychological tests, which is well suited for an analysis of study population heterogeneity.

Group comparisons are conducted with analyses of variance (ANOVA), t-tests, or Fisher’s exact test dependent on the number of groups and nature of the data. Significance level is assumed with α < 0.05, post-hoc tests are Bonferroni-corrected. Exploratory analyses for relationships between variables are investigated with Pearson’s correlation or general linear regression models. Exploratory patient stratification is conducted with k-means and hierarchical clustering with Euclidean distance based on min-max normalized data.

The initial database search yielded 193 papers. We excluded 91 papers during title screening and further 25 during abstract screening. From the remaining 77 papers, we excluded eleven more during full paper review. Reasons for exclusion were: Electrocardiography data was used instead of EEG (n = 1), event-related potentials were analyzed and not resting-state EEG (n = 3), no HC (n = 2), not accessible (n = 3), not in English (n = 1). Additionally, one paper reported using the MODMA data set but since the data presented did not match this data at all, the study was excluded as well. This left 66 papers to include in the review, 34 of which collected own data, and 32 worked with publicly available data sets.

The 32 studies working exclusively with public data utilized seven different data sets. The most frequently used data set was MODMA (Cai et al., 2022). Nine studies are based solely on this data (Deng et al., 2022; Liu W. et al., 2022; Wang B. et al., 2023; Wu W. et al., 2022; Zhang B. et al., 2023; Zhang et al., 2021; Zhang J. et al., 2023; Zhao et al., 2022a; Zhao et al., 2022b), five more studies additionally included other public data sets (Chu et al., 2024; Kabbara et al., 2022; Movahed et al., 2022; Sun et al., 2024; Wang et al., 2024c), and another study replicated their findings based on own data with this data set (Soni et al., 2022). MODMA is a multi-modal data set for MDD research containing two EEG experiments with eyes-closed resting-state recordings from three or 128 electrodes. Descriptions in the subsequent chapter are only for the 128-electrode data (n = 24/29 MDD/HC), because this set was used by all but one of the studies in this review.

Nearly as frequently used is the data set collected from Mumtaz and colleagues (MUM) (Mumtaz et al., 2017). MUM was used in eleven studies alone (Ataei and Wang, 2022; Ellis et al., 2024; Ellis et al., 2023; Kang et al., 2020; Khadidos et al., 2023; Mahato and Paul, 2019; Movahed et al., 2021; Saeedi et al., 2020; Tang et al., 2024; Zhou et al., 2024) and in three studies in combination with other data sets (Li et al., 2024; Movahed et al., 2022; Sun et al., 2024). The public data set contains eyes-closed and eyes-open resting-state EEG data as well as task-EEG data with 19 electrodes from 34 MDD patients and 30 HC.

The data set provided by Cavanagh and colleagues (CAV) (Cavanagh et al., 2019) was used in four studies alone (Thoduparambil et al., 2020; Trambaiolli and Biazoli, 2020; Yun and Jeong, 2021; Zandbagleh et al., 2024) and together with other data sets in five studies (Chu et al., 2024; Kabbara et al., 2022; Li et al., 2024; Sun et al., 2024; Wang et al., 2024c). The data set includes task as well as resting-state EEG data with 64 electrodes from 120 participants altogether.

The DRYAD data (Kołodziej et al., 2021) was used in one study in combination with other data sets (Chu et al., 2024) and consists of three small data sets: Nowowiejska (NOW; n = 55), DiamSar (DIA; n = 95), and Wronski (WRO; n = 82). All three data sets contain 64-channel eyes-closed resting-state EEG data.

The data sets EMBARC (Webb et al., 2016), TDBRAIN (Van Dijk et al., 2022), and B-SNIP (Tamminga et al., 2017) were used by one study each (Ciarleglio et al., 2022; Gour et al., 2023; Lechner and Northoff, 2024). The EEG data of these three data sets are part of larger data collections. EMBARC was a multi-site drug study including resting-state EEG as well as fMRI data. TDBRAIN was collected over 20 years, containing EEG data of patients with different diagnoses, most frequent were MDD, attention deficit disorder, subjective memory complaints, and obsessive-compulsive disorder. B-SNIP primarily contains data from schizophrenia, schizoaffective disorder, or psychotic bipolar I disorder patients, and their direct relatives. The data set includes resting-state and task EEG, fMRI, and blood samples.

More than half of the studies with own data collection were conducted in Asia, with more than half of the investigated subjects participating at an Asian location (Table 1). This information serves as proxy for ethnicity since most studies do not provide explicitly the ethnicity of their participants. Since two data sets were used twice each and one study did not provide the number of participants, the number of participants is based on 31 studies. The publicly available data sets were collected in China (MODMA), Malaysia (MUM), Poland (NOW, DIA, WRO), Netherlands (TDBRAIN), and USA (CAV, EMBARC, B-SNIP).

Half of the studies are based on overall sample sizes of 80 participants or fewer (Figure 1A), with the majority having about an equal distribution between the clinical groups (Figure 1B, top box). Studies tend to include more female than male participants in both clinical populations (Figure 1B, bottom boxes) with two studies investigating female participants only (Shim et al., 2023; Umemoto et al., 2021).

Age distributions of studies vary widely (Figure 2). The older the study population average the higher its standard deviation, except for one study with a rather old but confined age distribution (Wu Z. et al., 2022). One study specifically recruited adolescents up to 18 years (Umemoto et al., 2021), all other studies targeted adults.

Seven studies based on own data considered gender or age in their analysis. Two studies included gender: Lord and Allen (2023) conducted additional analyses for gender separately and found differences in EEG complexity metrics between those groups. Three studies (Benschop et al., 2022; Jang et al., 2020; Mitiureva et al., 2024) used age and gender as covariates in statistical comparison of clinical groups or regressed them out in correlation analyses. Two more studies treat only age as covariate (Périard et al., 2024; Umemoto et al., 2021); note that the latter one included only female participants. One study specifically investigated pathological aging processes and found that aging affects the diagnostic capability of EEG biomarkers (Sarisik et al., 2024).

The studies using only MODMA, CAV, or MUM data did not consider gender, age, or any other demographic information in their analysis. One study using both MOD and CAV did not find an age but a gender influence in their regression models (Kabbara et al., 2022). The two studies using EMBARC (Ciarleglio et al., 2022) and B-SNIP (Lechner and Northoff, 2024) used age and gender as covariates in statistical comparison of clinical groups.

To summarize, the majority of studies is based on small sample sizes (n < 100 per diagnostic group) with a disproportional high ratio of Asian participants. Studies are rather variable in their gender ratio and age distributions, both for studies based on own data as well as public data sets. Only few studies consider gender or age in the analysis. Those explicitly investigating the influence of these factors, however, tended to find one.

All studies with own data separated their participants into the groups MDD and HC. This diagnostic label was used for classification in 12 studies, for group comparisons in 17 studies, and for both analysis approaches in five studies. Twenty-four of the studies based on public data used the diagnostic groups MDD and HC for classification, four for group comparisons, and two for both. One study based on the CAV data set (Trambaiolli and Biazoli, 2020) did not use that dichotomous diagnostic label at all but utilized neuropsychological scores as only target variable instead.

Neuropsychological tests assess the severity of psychiatric diseases, symptom dimension, or cognitive function, usually with a (self-administered) questionnaire. They can be used for screening, complement diagnosis, and provide a more detailed description of a study population.

Five neuropsychological tests that operationalize depression severity are administered most frequently (Table 2): Hamilton Depression Scale with 17 items (HAMD-17) (Hamilton, 1960), Beck Depression Inventory (BDI) (Beck et al., 1961), currently most frequently used in version 2 (BDI-II) (Beck et al., 1996), 9-item subscale for depression from the Patient Health Questionnaire (PHQ-9) (Kroenke et al., 2001), Self-Rating Depression Scale (SDS) (Zung, 1965), and Montgomery-Åsberg Depression Rating Scale (MADRS) (Montgomery and Åsberg, 1979). Note that a clinician administers the HAMD-17 and MADRS while the other three tests are usually self-administered. Tests that operationalize anxiety sometimes complement the depression score, most frequently the Hamilton Anxiety Rating Scale (HAM-A) (Hamilton, 1969), and the State–Trait Anxiety Inventory (STAI) (Spielberger et al., 1971) used in four and one study, respectively, the latter also in the CAV data set. Further neuropsychological tests that occurred more than once were the Young Mania Rating Scale (YMRS) (Young et al., 2000) and the Mini Mental State Examination (MMSE) (Folstein et al., 1975) with four and two studies administering them, respectively.

Distributions of the two most frequently used tests demonstrate a high variability across study populations (Figure 3) with distinctive differences between diagnostic groups. Noteworthy, sometimes the tests are only administered to the MDD but not the HC group, as apparent for the HAMD-17 (Figure 3A).

Half of the studies using own data included the test scores from at least one neuropsychological test in their analysis (Table 3). From the studies pooling data across both groups, five used the BDI-II score to operationalize the severity of depression, the PHQ-9, SRS, and MADRS were used once each. From the studies conducting the analysis on the MDD group only, six used the HAMD-17 and two the BDI-II. Six studies that use public data included the test score of at least one neuropsychological test (Table 3, marked with *). One of these studies, however, shows PHQ-9 values in their graphs that are actually not contained in the public data set (Zhang et al., 2021). The spectral markers seem to be rather contradictory in correlation with depression severity or other neuropsychological test scores. Connectivity markers, however, seem to be better correlated.

In summary, the studies employed a variety of neuropsychological questionnaires, the most commonly used being the BDI and HAMD-17. These questionnaires categorize MDD into different levels of severity. However, the distribution of these scores varies greatly across studies. About half of the studies with own data also use the scores of neuropsychological questionnaires in their analysis, studies based on public data employ this practice less frequently. The results of the studies are partially inconsistent, especially regarding spectral markers.

The participants of the CAV data set can be divided into five groups based on their diagnostic procedure (Figure 4A): (1) MDD_current: diagnosis & BDI ≥ 13 (n = 11; 54.5% female); (2) MDD_remitted: diagnosis & BDI ≥ 13 (n = 12; 75.0% female); (3) highBDI_noMDD: diagnosis & BDI ≥ 13 (n = 9; 66.7% female); (4) highBDI_undiag: no diagnosis & BDI ≥ 13 (n = 14; 92.9% female); (5) lowBDI: BDI ≤ 7 (n = 75; 53.3% female).

The first four groups have on average a moderate BDI-score (Figure 4A; 22.2 ± 4.9) without sign. differences between groups (F_3,42 = 0.398; p = 0.775). The HAMD-17 score, however, differs sign. between the first three groups (Figure 4B; F_2,29 = 5.169; p = 0.012), with the MDD_current group (13.1 ± 5.3) exceeding sign. both the MDD_remitted (7.3 ± 5.3; t₂₁ = 2.637; p = 0.015; corr. p = 0.045) and the highBDI_noMDD group (6.8 ± 4.4; t₁₈ = 2.852; p = 0.011; corr. p = 0.033). Note that only diagnosed participants were assessed with the HAMD-17 test.

Several dichotomous diagnosis groups (MDD/HC) can be formed from these five groups, depending on the research question. This practice has been used in the studies based on this data set. (1) The strictest separation includes only MDD_current for the MDD group (n = 11) and lowBDI for the HC group (n = 75). (2) The second approach expands the MDD group by the MDD_remitted group (n = 23) and keeps the HC group the same. (3) The diagnosis-agnostic approach groups participants solely based on BDI, e.g., BDI ≥ 13 for the MDD group (n = 46) and BDI ≤ 7 for the HC group (n = 75). The cutoff value, however, is of minor importance in this data set since there is a gap in the scores between eight and twelve. For a data set with continuous BDI scores, however, different cutoff values might be used and are used, e.g., group splits at BDI scores of nine or 13. (4) Only diagnosed MDD (current & remitted; n = 23) are included in the MDD group while the HC group comprises all other participants (n = 98). The last approach seems counterintuitive for the given data set. However, since a substantial amount of studies comprises ill-defined HC groups, this approach mirrors this practice.

Obviously, the distribution of BDI scores differs slightly between groups of these different “diagnosis” approaches. Furthermore, we find a sign. difference for gender between diagnostic groups in approach three only (odds ratio = 2.479; p = 0.034). This means that here either gender needs to be considered as a potential confound in the analysis, or, e.g., the HC group is subsampled to match the MDD group – an approach that is possible here since the HC group is substantial larger than the MDD group. For none of the approaches we find a sign. age difference between groups (all p ≥ 0.3).

The participants of the MODMA data set are usually grouped based on the diagnostic label given, which leads to a MDD group with 24 participants (45.8% female) and a HC group with 29 participants (31.0% female). Neither age nor gender differs sign. between these groups (both p > 0.3). The education, however, is a potential confound with a sign. difference between groups (t₅₁ = −3.209; p = 0.002), confirmed by a moderate but highly sign. correlation between education and PHQ-9 score (r = −0.45; p < 0.001).

The exploratory stratification analysis on the MODMA data set is restricted to three scores, since more than three dimensions are not as intuitively to visualize. We chose the depression severity (PHQ-9) along with the anxiety (7-item Generalized Anxiety Disorder scale: GAD-7) (Spitzer et al., 2006) and sleep score (Pittsburgh Sleep Quality Index: PSQI) (Buysse et al., 1989) because these two correspond to possible symptoms of depression (American Psychiatric Association, 2013; Malgaroli et al., 2021) as well as to disorders on their own right that can be co-morbid with MDD (Liu et al., 2007; Meng and Wang, 2023; Sevillano-Garcia et al., 2007; Staner, 2010; Thaipisuttikul et al., 2014; Zimmerman et al., 2002).

Visual exploration of the clustering hierarchy revealed four clusters that adequately separate the participants in one HC group and three MDD groups (Figures 5A,B). For the latter, a small group of patients with low anxiety but sleep problems (magenta markers) stands out. The separation of the remaining MDD patients (red and orange markers) seems to be based on all three scores in a similar fashion. Based on the four clusters found in the hierarchical approach, the k-means algorithm was parametrized with four clusters as well. This algorithm separates the HC and MDD groups as well but also sub-divides the healthy group (Figures 5C,D). Especially for the HC, it seems to be mainly driven by the PSQI/GAD-7 combination (blueish markers in Figure 5D) rather than by the PHQ-9 score (Figure 5C).

To summarize, grouping MDD and HC according to different criteria demonstrates that apart from a change in the distribution of MDD severity rating in the diagnostic groups, also the distribution of other variables like demographic information may shift, which might lead to potential confounders dependent on grouping. We also demonstrate that different neuropsychological tests like HAMD-17 and BDI can have different capabilities of separating between diagnostic subgroups. Moreover, adding further neuropsychological tests related to symptoms or co-morbidities such as sleep problems or anxiety, we show that already rather small and confined samples might contain clinically relevant subgroups.