A systematic review of EEG-based biomarkers for depression, anxiety, and bipolar disorder: trends in explainable artificial intelligence (XAI) Authors: Lijiang Zhai, Min Zhao, Jia Zhang, Muhammad Jamil, Rida Naz, Caina Li Published: 2025 ABSTRACT BackgroundMood and anxiety disorders, including major depressive disorder (MDD), bipolar disorder (BD), and anxiety disorders, affect millions worldwide and pose significant diagnostic challenges. Electroencephalography (EEG) has emerged as a non-invasive neurophysiological tool for identifying biomarkers of these disorders. However, the adoption of deep learning and machine learning in EEG analysis introduces interpretability concerns that limit clinical trust and applicability. Explainable Artificial Intelligence (XAI) offers transparent insights into model decisions, aiming to improve trust and clinical adoption of AI tools.ObjectiveThis systematic review aims to critically evaluate the landscape of EEG-based biomarkers for depression, anxiety, and bipolar disorder as informed by XAI techniques. Specifically, it explores the nature of EEG features selected through explainable models, their physiological plausibility, diagnostic performance, methodological quality, and clinical translatability.MethodsFollowing PRISMA guidelines, a comprehensive literature search was conducted across PubMed, Scopus, IEEE Xplore, Web of Science, and Embase databases for studies published between 2015 and 2025. Inclusion criteria were studies employing EEG-derived features for mood disorder classification using machine learning or deep learning, coupled with at least one post-hoc XAI method (e.g. SHAP, LIME, attention maps). A total of 145 studies were included. Thematic synthesis was organized into key domains: EEG protocol and feature extraction, model performance, XAI methods, interpretability outcomes, risk of bias, translational barriers, and publication trends.ResultsThe review revealed that spectral features (e.g. alpha and beta power), functional connectivity metrics (e.g. PLI, PCC), and time-domain measures (e.g. Hjorth parameters) were the most commonly exploited EEG features. XAI-enhanced models consistently identified physiologically valid biomarkers such as frontal alpha asymmetry in MDD and frontal–parietal beta connectivity in BD. Machine learning models—including SVM, Random Forest, and deep neural networks—achieved high diagnostic accuracies (75–99%) across studies. SHAP and LIME were the most widely used interpretability methods, offering feature importance explanations that aligned well with known pathophysiological patterns. However, approximately 65% of studies exhibited high risk of bias due to small sample sizes, lack of external validation, and methodological inconsistencies. Ethical and translational concerns included limited clinician trust, lack of regulatory frameworks, and potential algorithmic bias. Publication trends showed exponential growth in EEG-XAI research post-2020, particularly in studies targeting depression.ConclusionEEG-XAI frameworks represent a transformative approach to neuropsychiatric diagnostics by coupling physiological insights with algorithmic transparency. While explainable models enhance diagnostic confidence and interpretability, their clinical utility is contingent upon methodological rigor, ethical safeguards, and cross-disciplinary validation. Future research should prioritize multi-site data harmonization, larger cohorts, and longitudinal designs to facilitate the safe and effective deployment of EEG-XAI tools in clinical psychiatry.Clinical trial numberNA.Supplementary informationThe online version contains supplementary material available at 10.1186/s12888-025-07589-3. Introduction Electroencephalography (EEG) is a widely used non-invasive technique for investigating neural dynamics in psychiatric disorders [1]. Although anxiety disorders are diagnostically distinct from mood disorders under DSM-5 and ICD-11, they are included in this review because of their high prevalence, shared neurophysiological features with mood disorders, and frequent co-occurrence in EEG-based AI studies. The application of EEG in the study of depression, anxiety, and bipolar disorder has gained considerable momentum, driven by its ability to reflect dynamic changes in brain function associated with these conditions [2]. Traditional methods of EEG analysis often rely on visual inspection or basic quantitative measures, which may be limited in their capacity to detect the intricate patterns and subtle nuances in the data [3]. In response to these limitations, researchers are increasingly turning to advanced signal processing techniques and machine learning (ML) algorithms to extract more informative features from EEG recordings. Deep learning methodologies, in particular, have demonstrated significant promise in decoding complex brain states and classifying psychiatric disorders using EEG signals [4]. The success of deep learning in EEG analysis stems from its capacity to automatically learn hierarchical data representations, capturing intricate relationships that conventional approaches may overlook [5]. However, the inherent “black box” nature of deep learning models poses a barrier to clinical implementation, as the rationale behind their predictions is often opaque [6]. To overcome this challenge, explainable artificial intelligence (XAI) methods have been developed to improve transparency and interpretability of AI models, supporting clinical trust [7]. XAI techniques help identify which EEG features or patterns are most relevant for diagnosis or prediction, offering insights into the neural mechanisms associated with psychiatric conditions [8]. The convergence of EEG-based biomarkers and XAI holds substantial promise for improving our understanding and treatment of depression, anxiety and bipolar disorder Figure 1 illustrates the end-to-end EEG-XAI pipeline, showing how raw EEG data are transformed through signal processing and feature extraction into machine learning models, and subsequently interpreted via XAI for clinical decision support. This horizontal flowchart (Fig. 1) provides a clear visual guide to the workflow stages that are critical for developing transparent and clinically actionable EEG-based biomarkers.Fig. 1Horizontal flowchart of the EEG-XAI workflow Horizontal flowchart of the EEG-XAI workflow The application of deep learning to EEG signal decoding has seen increasing research interest [9]. Deep neural networks such as convolutional neural networks (CNNs) and long short-term memory (LSTM) networks are being widely applied in psychiatric classification tasks [10]. These models offer powerful tools for large-scale data analysis, enabling both prediction and classification in biomedical domains [11]. The use of deep belief networks and other architectures for EEG signal analysis has also contributed to improved physiological understanding [12]. Key to the effectiveness of ML algorithms in EEG analysis is the selection of relevant features and preprocessing strategies [13]. In the context of psychiatric disorders, these models can assist in identifying individuals at risk, monitoring disease progression, and predicting treatment outcomes. Importantly, the ability to interpret model decisions is vital for clinical trust and integration. Many ML models, especially deep learning models, lack interpretability, which hinders their adoption in real-world settings [8, 14]. Figure 2 contrasts the traditional “black-box” deep learning paradigm, in which EEG signals are input into an opaque model, with an explainable AI approach that uses SHAP and LIME values to reveal feature importance. The right side of Fig. 2 demonstrates how XAI techniques provide transparent insights into the EEG patterns driving predictions, thereby supporting clinical trust and decision-making.Fig. 2Comparison of traditional black-box deep learning and explainable AI (XAI) for EEG signal Comparison of traditional black-box deep learning and explainable AI (XAI) for EEG signal While XAI is still a nascent field, especially in psychiatric EEG research, its growth is rapid and promising. Techniques such as SHAP values, LIME, and attention mechanisms are enabling researchers to pinpoint EEG features that distinguish between mental health conditions [7]. By improving interpretability, XAI enhances clinical decision-making, supports targeted intervention strategies, and promotes personalized care. Nevertheless, applying XAI to deep learning models comes with computational costs, especially for large EEG datasets. There is a pressing need to develop computationally efficient XAI approaches that remain clinically meaningful [15]. Moreover, evaluation frameworks to ensure the clinical utility of these explanations are currently underdeveloped [16]. Clinically validated studies are essential to demonstrate whether XAI-enhanced EEG analysis truly improves diagnostic accuracy and treatment outcomes [17]. XAI is especially critical in healthcare due to the ethical and regulatory need for transparency and fairness [18, 19]. Without explainability, clinicians may hesitate to adopt even highly accurate models [20]. Transparent models are more likely to be accepted by clinicians, regulators, and patients alike [21]. Most current XAI work focuses on interpreting black-box models, but the shift toward inherently interpretable models is gaining traction in domains such as medical imaging and psychiatric diagnostics [22, 23]. Finally, while global feature importance values help summarize model behavior, they often fail to represent subgroup or individual-level predictions [15]. Tools like FeatureExplorer may enhance prediction performance but often neglect explainability. The rationale for this review stems from this gap: many EEG-based ML models for psychiatric disorders lack interpretability, hindering clinical applicability. This review aims to address that gap by systematically evaluating the integration of XAI techniques in EEG biomarker research across depression, anxiety, and bipolar disorder [24]. As shown in Fig. 3, the global burden of mental health disorders varies substantially by region, with prevalence rates exceeding 16% in several countries. These estimates reflect the combined prevalence of multiple mental health conditions, including depression, anxiety, bipolar disorder, eating disorders, and schizophrenia.Fig. 3Global prevalence of mental health disorders (2021), including depression, anxiety, bipolar disorder, eating disorders, and schizophrenia. Data source: IHME, global burden of disease (2024). Source: https://ourworldindata.org/grapher/share-with-mental-and-substance-disorders?time=latest Global prevalence of mental health disorders (2021), including depression, anxiety, bipolar disorder, eating disorders, and schizophrenia. Data source: IHME, global burden of disease (2024). Source: https://ourworldindata.org/grapher/share-with-mental-and-substance-disorders?time=latest Figure 4 shows a color-coded world map of age-standardized depressive disorder prevalence by country for the latest available year, illustrating global hotspots and regional disparities. This visualization underscores the varying burden of depression worldwide and provides context for prioritizing EEG-XAI biomarker research in high-prevalence areas.Fig. 4World map showing the age-standardized prevalence (%) of depressive disorders by country for the most recent year, based on IHME GBD data. Source: https://ourworldindata.org/grapher/depressive-disorders-prevalence-ihme?time=latest World map showing the age-standardized prevalence (%) of depressive disorders by country for the most recent year, based on IHME GBD data. Source: https://ourworldindata.org/grapher/depressive-disorders-prevalence-ihme?time=latest Objectives This systematic review aims to provide a comprehensive overview of the current state of research on explainable AI in EEG-based studies of depression, anxiety, and bipolar disorder. The goal is to systematically investigate how explainable AI techniques are employed to enhance the interpretability of machine learning models utilizing EEG data for the identification and classification of depression, anxiety, and bipolar disorder. This review seeks to achieve a comprehensive understanding of the methodologies, applications, and outcomes associated with explainable AI in the context of EEG-based analysis for mental health disorders. By examining the existing literature, we aim to identify key trends, challenges, and opportunities in this emerging field and the review seeks to outline the prevalent XAI methodologies employed in conjunction with EEG data, elucidating their strengths, limitations, and applicability to the study of depression, anxiety, and bipolar disorder. This includes examining diverse XAI techniques such as feature importance analysis, rule extraction, and visualization methods, and evaluating their effectiveness in enhancing the interpretability of machine learning models. Table 1 outlines the seven thematic components of our review, detailing their specific aims, scopes, and in-depth XAI-focused research questions. By systematically addressing each question—from XAI-enhanced diagnostic value to publication trends—this framework ensures a comprehensive evaluation of EEG-XAI biomarkers in mood and anxiety disorders (Table 1).Table 1Overview of the aims, scope, and critical-thinking research questions for each major thematic component of the systematic review on AI innovations in EEG-Based biomarkers for depression, anxiety, and bipolar disorder (2015–2025)Thematic ComponentAimScopeReview QuestionXAI-Enhanced Clinical Impact & Diagnostic ValueDemonstrate how explainability adds clinical value over raw EEG featuresComparative analyses of XAI-guided vs. traditional EEG-based predictions in clinical cohortsIn what ways do XAI-driven explanations of EEG features (e.g., SHAP-ranked spectral bands or connectivity metrics) improve clinician decision-making, diagnostic confidence, and patient stratification compared to models without interpretability?XAI-Informed Feature Extraction & ValidationIdentify EEG features that XAI methods consistently highlight and validate against known neurobiological markersEvaluation of XAI-selected features (e.g., via SHAP/LIME) across tasks, preprocessing pipelinesWhich EEG-derived features (spectral, time-frequency, connectivity) selected by XAI approaches show the strongest reproducibility and physiological validity across diverse datasets, and how do these explainable selections align with established neural mechanisms of each disorder?Machine Learning ModelsEvaluate ML algorithms appliedSVM, Random Forest, CNN, ensemble methodsWhat ML algorithms have been used to classify mood and anxiety disorders based on EEG data, and what are their performance metrics?Explainable AI TechniquesExamine interpretability methodsSHAP, LIME, attention/saliency mapsWhich XAI methods have been applied to EEG-based models, and how do they enhance model transparency?Methodological Quality & BiasAssess risk of bias and evidence qualityQUADAS-2, PROBAST, methodological heterogeneityWhat is the methodological robustness of EEG-XAI studies, and what biases may affect their findings?Clinical Translation & EthicsExplore pathways to clinical deployment and ethical considerationsRegulatory approval, clinician acceptance, data privacy, algorithmic fairnessWhat are the translational barriers and ethical implications for deploying EEG-XAI biomarkers in clinical practice?Publication & Trend AnalysisMap publication trends over timeYearly publication counts, disorder-specific focus from 2015 to 2025How have publication trends and research focus Table 1Overview of the aims, scope, and critical-thinking research questions for each major thematic component of the systematic review on AI innovations in EEG-Based biomarkers for depression, anxiety, and bipolar disorder (2015–2025)Thematic ComponentAimScopeReview QuestionXAI-Enhanced Clinical Impact & Diagnostic ValueDemonstrate how explainability adds clinical value over raw EEG featuresComparative analyses of XAI-guided vs. traditional EEG-based predictions in clinical cohortsIn what ways do XAI-driven explanations of EEG features (e.g., SHAP-ranked spectral bands or connectivity metrics) improve clinician decision-making, diagnostic confidence, and patient stratification compared to models without interpretability?XAI-Informed Feature Extraction & ValidationIdentify EEG features that XAI methods consistently highlight and validate against known neurobiological markersEvaluation of XAI-selected features (e.g., via SHAP/LIME) across tasks, preprocessing pipelinesWhich EEG-derived features (spectral, time-frequency, connectivity) selected by XAI approaches show the strongest reproducibility and physiological validity across diverse datasets, and how do these explainable selections align with established neural mechanisms of each disorder?Machine Learning ModelsEvaluate ML algorithms appliedSVM, Random Forest, CNN, ensemble methodsWhat ML algorithms have been used to classify mood and anxiety disorders based on EEG data, and what are their performance metrics?Explainable AI TechniquesExamine interpretability methodsSHAP, LIME, attention/saliency mapsWhich XAI methods have been applied to EEG-based models, and how do they enhance model transparency?Methodological Quality & BiasAssess risk of bias and evidence qualityQUADAS-2, PROBAST, methodological heterogeneityWhat is the methodological robustness of EEG-XAI studies, and what biases may affect their findings?Clinical Translation & EthicsExplore pathways to clinical deployment and ethical considerationsRegulatory approval, clinician acceptance, data privacy, algorithmic fairnessWhat are the translational barriers and ethical implications for deploying EEG-XAI biomarkers in clinical practice?Publication & Trend AnalysisMap publication trends over timeYearly publication counts, disorder-specific focus from 2015 to 2025How have publication trends and research focus Overview of the aims, scope, and critical-thinking research questions for each major thematic component of the systematic review on AI innovations in EEG-Based biomarkers for depression, anxiety, and bipolar disorder (2015–2025) Methodology We performed a systematic literature search in major bibliographic databases—PubMed/MEDLINE, IEEE Xplore, Scopus, Web of Science, and Embase—limited to English-language, peer-reviewed articles published between 2015 and 2025. The inclusion of Embase ensured broader coverage of biomedical and clinical research, particularly studies relevant to psychiatric and neurological disorders. A combination of controlled-vocabulary terms (MeSH in PubMed) and free-text keywords was used. The core search string (adapted for each database) is shown in the following Table 2.Table 2Sample search strings used in databasesComponentSearch TermsEEG Terms“EEG” OR “electroencephalography”Disorder Terms“depression” OR “major depressive disorder” OR “anxiety” OR “generalized anxiety disorder” OR “bipolar disorder”Biomarker Terms“biomarker” OR “feature” OR “marker” OR “neural signature”Machine Learning Terms“machine learning” OR “deep learning” OR “classification” OR “support vector machine” OR “random forest” OR “convolutional neural network” OR “CNN” OR “ensemble”Explainable AI Terms“explainable AI” OR “XAI” OR “interpretability” OR “explainability” OR “SHAP” OR “LIME” OR “attention map” OR “saliency map”Combined String(“EEG” OR “electroencephalography”)AND (“depression” OR “major depressive disorder” OR “anxiety” OR “generalized anxiety disorder” OR “bipolar disorder”)AND (“biomarker” OR “feature” OR “marker” OR “neural signature”)AND (“machine learning” OR “deep learning” OR “classification” OR “support vector machine” OR “random forest” OR “convolutional neural network” OR “CNN” OR “ensemble”)AND (“explainable AI” OR “XAI” OR “interpretability” OR “explainability” OR “SHAP” OR “LIME” OR “attention map” OR “saliency map”) Table 2Sample search strings used in databasesComponentSearch TermsEEG Terms“EEG” OR “electroencephalography”Disorder Terms“depression” OR “major depressive disorder” OR “anxiety” OR “generalized anxiety disorder” OR “bipolar disorder”Biomarker Terms“biomarker” OR “feature” OR “marker” OR “neural signature”Machine Learning Terms“machine learning” OR “deep learning” OR “classification” OR “support vector machine” OR “random forest” OR “convolutional neural network” OR “CNN” OR “ensemble”Explainable AI Terms“explainable AI” OR “XAI” OR “interpretability” OR “explainability” OR “SHAP” OR “LIME” OR “attention map” OR “saliency map”Combined String(“EEG” OR “electroencephalography”)AND (“depression” OR “major depressive disorder” OR “anxiety” OR “generalized anxiety disorder” OR “bipolar disorder”)AND (“biomarker” OR “feature” OR “marker” OR “neural signature”)AND (“machine learning” OR “deep learning” OR “classification” OR “support vector machine” OR “random forest” OR “convolutional neural network” OR “CNN” OR “ensemble”)AND (“explainable AI” OR “XAI” OR “interpretability” OR “explainability” OR “SHAP” OR “LIME” OR “attention map” OR “saliency map”) Sample search strings used in databases (“EEG” OR “electroencephalography”) AND (“depression” OR “major depressive disorder” OR “anxiety” OR “generalized anxiety disorder” OR “bipolar disorder”) AND (“biomarker” OR “feature” OR “marker” OR “neural signature”) AND (“machine learning” OR “deep learning” OR “classification” OR “support vector machine” OR “random forest” OR “convolutional neural network” OR “CNN” OR “ensemble”) AND (“explainable AI” OR “XAI” OR “interpretability” OR “explainability” OR “SHAP” OR “LIME” OR “attention map” OR “saliency map”) Table 2 displays the core keyword components and their specific terms used to retrieve relevant studies on EEG, mood and anxiety disorders, machine learning, and explainable AI across multiple databases. By combining these components into a single search string, we ensured comprehensive coverage of EEG-XAI research published from 2015 to 2025 (Table 2). Study selection Two reviewers (XX and YY) independently screened all retrieved records by title/abstract and then full text, using predefined eligibility criteria. Disagreements were resolved through discussion and consensus, with a third reviewer (ZZ) consulted if needed. Inter-rater agreement was high (Cohen’s κ = 0.81 at title/abstract stage, κ = 0.84 at full-text stage), indicating strong reliability. Exclusion criteria included: (i) non-peer-reviewed articles, (ii) studies not using EEG, (iii) studies not applying AI/ML methods, (iv) reviews, editorials, and conference abstracts, and (v) studies without quantitative performance results. Inclusion and exclusion criteria Eligible studies were original research articles involving adult (≥18 years) human participants diagnosed with major depressive disorder, generalized anxiety disorder, or bipolar disorder that applied a machine learning or deep learning algorithm to classify or predict one of these conditions using EEG-derived features. Each included study explicitly incorporated at least one post-hoc explainability method (e.g., SHAP, LIME, attention or saliency mapping) to interpret model outputs and reported quantitative performance metrics (e.g., accuracy, AUC, sensitivity, specificity). We excluded reviews, meta-analyses, case reports, editorials, conference abstracts, and commentaries; studies that relied solely on non-EEG modalities (e.g., fMRI, MEG); animal or in vitro investigations; and any study that did not include a model interpretability or explainability component. Exclusion criteria were: (i) non-peer-reviewed articles, (ii) studies not using EEG or not applying AI/ML methods, (iii) studies without performance metrics, (iv) reviews/editorials, and (v) non-human studies Table 3 summarizes the predefined inclusion and exclusion criteria used to select EEG-XAI studies, ensuring that only human adult research applying machine learning with interpretability to depression, anxiety, or bipolar disorder is included. By clearly delineating permissible study designs, modalities, and outcomes, this framework promotes consistency and transparency in the screening process (Table 3).Table 3Inclusion and exclusion criteriaCriterionInclusionExclusionPublication Date2015–2025Before 2015PopulationHuman adults (≥18 years) diagnosed with depression, anxiety, or bipolar disorderNon-human studies; pediatric ( < 18 years); healthy-only populations without psychiatric diagnosisModalityStudies using scalp EEG recordingsNon-EEG modalities only (e.g., fMRI, MEG); EEG combined with other modalities without separate EEG analysisAI MethodMachine learning or deep learning models applied to EEG dataNo AI/ML models; purely statistical or non-automated analysesExplainabilityUse of post-hoc XAI techniques (e.g., SHAP, LIME, attention/saliency maps)No model interpretability component; black-box only without any explainable AI methodStudy DesignOriginal research articles (cohort, case–control, diagnostic accuracy, predictive modeling)Reviews, meta-analyses, editorials, case reports, conference abstracts, commentariesLanguageEnglishNon-EnglishOutcomesDiagnostic or prognostic performance metrics (e.g., acc uracy, AUC, sensitivity, specificity) and XAI findings (feature importance, interpretability)Studies reporting only qualitative findings without quantitative performance or explainability metrics Table 3Inclusion and exclusion criteriaCriterionInclusionExclusionPublication Date2015–2025Before 2015PopulationHuman adults (≥18 years) diagnosed with depression, anxiety, or bipolar disorderNon-human studies; pediatric ( < 18 years); healthy-only populations without psychiatric diagnosisModalityStudies using scalp EEG recordingsNon-EEG modalities only (e.g., fMRI, MEG); EEG combined with other modalities without separate EEG analysisAI MethodMachine learning or deep learning models applied to EEG dataNo AI/ML models; purely statistical or non-automated analysesExplainabilityUse of post-hoc XAI techniques (e.g., SHAP, LIME, attention/saliency maps)No model interpretability component; black-box only without any explainable AI methodStudy DesignOriginal research articles (cohort, case–control, diagnostic accuracy, predictive modeling)Reviews, meta-analyses, editorials, case reports, conference abstracts, commentariesLanguageEnglishNon-EnglishOutcomesDiagnostic or prognostic performance metrics (e.g., acc uracy, AUC, sensitivity, specificity) and XAI findings (feature importance, interpretability)Studies reporting only qualitative findings without quantitative performance or explainability metrics Inclusion and exclusion criteria Two reviewers independently screened titles and abstracts. Eligible records then underwent full-text review. Discrepancies were resolved by consensus or, when necessary, third-party arbitration. Reasons for exclusion were documented. Screening data are summarized in the PRISMA flow diagram (Fig. 5).Fig. 5PRISMA 2020 flow diagram PRISMA 2020 flow diagram Risk of bias and quality assessment We assessed methodological quality and risk of bias using three established tools: Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2), Prediction model Risk of Bias ASsessment Tool (PROBAST), and Methodological Index for Non-Randomized Studies (MINORS). QUADAS-2 primarily evaluates four domains—patient selection, index test, reference standard, and flow/timing. In our review, this tool frequently flagged issues related to patient selection and index test domains, particularly in studies that used retrospective EEG datasets without clear diagnostic thresholds or blinded review procedures. PROBAST focuses on risk of bias in prediction model studies across domains including participants, predictors, outcomes, and analysis. Most concerns were identified in the analysis domain, often due to limited model validation or exclusion of confounding variables. MINORS, designed to evaluate non-randomized studies, consistently highlighted weaknesses in prospective data collection, sample representativeness, and baseline comparability. Figure 6 illustrates the proportion of 145 included studies sourced from each database, with PubMed contributing the largest share (40.0%), followed by Scopus (29.7%), Web of Science (15.2%), IEEE Xplore (10.3%), and Embase (4.8%).This visualization highlights the primary data sources for EEG-XAI research in mood and anxiety disorders over the 2015–2025 period.Fig. 6Distribution of 145 included studies by database: PubMed (40.0%), Scopus (29.7%), IEEE Xplore (10.3%), web of science (15.2%), and embase (4.8%) Distribution of 145 included studies by database: PubMed (40.0%), Scopus (29.7%), IEEE Xplore (10.3%), web of science (15.2%), and embase (4.8%) Results Study characteristics This section presents a comprehensive demographic and methodological summary of EEG-based XAI techniques in diagnosing and analyzing depression, anxiety, and bipolar disorder. These studies span 2015–2025 and represent a broad range of machine learning approaches, EEG signal processing techniques, and clinical applications. Distribution by disorder Out of the 145 included studies, the majority (85 studies) focused on major depressive disorder (MDD), reflecting its global prevalence and the growing demand for non-invasive, neurophysiological biomarkers in its diagnosis and treatment. A total of 35 studies targeted various forms of anxiety disorders, including generalized anxiety disorder (GAD), panic disorder, and social anxiety disorder (SAD), demonstrating the rising use of EEG-based models in identifying psychophysiological patterns associated with anxiety. Meanwhile, 25 studies explored bipolar disorder (BD), covering both depressive and manic phases. These studies often employed comparative models to distinguish BD from other mood and anxiety disorders using machine learning and XAI-driven EEG analysis [17, 25]. Geographic distribution The geographic origin of the studies reveals a strong international interest in EEG-based XAI applications across diverse regions: Asia accounted for approximately 40% of the studies, led by contributions from China, India, South Korea, and Japan. Research from this region often employed advanced deep learning architectures and large EEG datasets for depression and anxiety detection [7, 26–28]. Europe contributed roughly 35%, particularly from countries such as Germany, the United Kingdom, Poland, and Greece. European studies emphasized interpretability and ethical deployment of XAI models in clinical neuropsychiatry [17, 29, 30]. North America comprised about 20%, primarily from the United States and Canada, focusing on translational research, biomarker validation, and early treatment prediction in psychiatric disorders [31]. The remaining 5% of studies came from Australia, South America, and Africa, showing emerging interest in the use of EEG signals and interpretable AI tools in mental health diagnostics. This global distribution highlights the interdisciplinary and international nature of this research field, with growing emphasis on population-specific EEG biomarkers and culturally informed AI models. Publication years and trends The temporal trend of publications reflects the increasing academic and clinical interest in integrating EEG analysis with machine learning and explainability frameworks: 2015–2017: A small number of foundational studies were published during this period, focusing largely on signal processing techniques and basic classification models for psychiatric diagnosis [30, 32]. 2018–2020: There was a noticeable growth in research activity as more studies began employing convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and attention-based models for EEG classification tasks [1, 30]. Importantly, this phase also marked the initial use of XAI methods like SHAP and LIME in psychiatric EEG analysis. 2021–2025: This period saw an exponential increase in publications, accounting for over 60% of the total included studies. Recent works leveraged graph neural networks (GNNs), hybrid multimodal approaches, and advanced anomaly detection frameworks with embedded XAI mechanisms for personalized mental health monitoring [25, 28, 33, 34]. The surge aligns with growing computational capabilities, open EEG datasets, and clinical interest in deploying AI for individualized diagnostics. Table 4 presents a representative subset of 20 studies selected for their methodological diversity and clinical relevance on EEG-based XAI applications for diagnosing depression, anxiety, and bipolar disorder. Each entry outlines the AI methodology, key performance metrics (e.g., AUC, sensitivity, specificity), and the clinical implications of the findings.Table 4Peer-reviewed studies on EEG-based XAI in depression, anxiety, and bipolar disorder (2015–2025)Study (Authors)YearAI MethodsResults (AUC, Sens, Spec)Clinical Insights/ImplicationsKwok et al.2025CNN, LSTM, GRU with bias mitigation strategiesDemonstrated bias in EEG datasets; mitigation strategies improved fairnessHighlighted the importance of addressing bias in EEG-based depression detection modelsAldayel & Al-Nafjan2024Random Forest with DWT featuresAccuracy: 87.5%Demonstrated the efficacy of ensemble learning in EEG-based anxiety detectionNazari et al.2024K-Nearest Neighbors combining EEG and cognitive testsBalanced Accuracy: 93%Showed improved differentiation between BD II and BPD using multimodal dataHwang et al.2025Multiscale Fuzzy Entropy and Relative Power analysesDifferentiated BD from SZ with high accuracyProvided EEG-based biomarkers for distinguishing between bipolar disorder and schizophreniaZhang et al.2025Graph Convolutional Networks with sample confidence methodsAccuracy: ~81%Addressed subjectivity in depression severity classification using EEGTeng et al.2022Graph theory analysis based on EEG-source estimatesIdentified abnormal cortical functional networks in MDD patientsSuggested disruptions in brain networks as potential causes of depressionWang et al.2024Graph Convolutional Neural Network and Gated Recurrent Unit modelsAccuracy: 94.72%Improved modeling capacity and reduced reliance on handcrafted features in depression detectionChen et al.2024Multi-dimensional attention-based deep neural networkAccuracy: 95.14%Effectively extracted spatial–temporal features for depression classificationZhang et al.2024Graph-based EEG approach for depression predictionAccuracy: 98.30% and 96.51% on two datasetsValidated strategy using MODMA and PRED+CT datasetsWang et al.2024Hybrid Graph Neural Network (HGNN)Achieved state-of-the-art performanceCaptured common and individualized depression patterns using EEGDemir et al.2021Graph Neural Networks (GNN) for EEG classificationOutperformed standard CNN classifiersAllowed neuroscientific interpretability and explainability in EEG classificationLi et al.2023Graph Frequency Attention Convolutional Neural Network (GFACNN)Exhibited potential for depression EEG classificationAided clinical diagnosis and treatment of depressionBreakspear et al.2023Machine learning on EEG dataIdentified predictive brainwave patternsSuggested AI could personalize depression treatmentSubhani et al.2017Deep Learning models on EEG dataImproved classification of psychiatric disordersProposed an improved EEG-based diagnostic model employing DL techniquesCukic et al.2018ML classifiers with non-linear EEG featuresAccuracy: 90.24%–97.56%Validated the effectiveness of HFD and SampEn in detecting depressive disordersJernajczyk et al.2017Alpha wavelet power analysisSensitivity: 90%, Specificity: 100%Developed an EEG biomarker for antidepressant treatment response in bipolar depressionHonke et al.2020Representation learning using β-VAE and SCANEnhanced interpretabilityImproved classification accuracy of clinical factors from EEGZhang et al.2025Explainable anomaly detection frameworkF1-score: 0.80Enabled monitoring of depression and anxiety using consumer wearable devices Table 4Peer-reviewed studies on EEG-based XAI in depression, anxiety, and bipolar disorder (2015–2025)Study (Authors)YearAI MethodsResults (AUC, Sens, Spec)Clinical Insights/ImplicationsKwok et al.2025CNN, LSTM, GRU with bias mitigation strategiesDemonstrated bias in EEG datasets; mitigation strategies improved fairnessHighlighted the importance of addressing bias in EEG-based depression detection modelsAldayel & Al-Nafjan2024Random Forest with DWT featuresAccuracy: 87.5%Demonstrated the efficacy of ensemble learning in EEG-based anxiety detectionNazari et al.2024K-Nearest Neighbors combining EEG and cognitive testsBalanced Accuracy: 93%Showed improved differentiation between BD II and BPD using multimodal dataHwang et al.2025Multiscale Fuzzy Entropy and Relative Power analysesDifferentiated BD from SZ with high accuracyProvided EEG-based biomarkers for distinguishing between bipolar disorder and schizophreniaZhang et al.2025Graph Convolutional Networks with sample confidence methodsAccuracy: ~81%Addressed subjectivity in depression severity classification using EEGTeng et al.2022Graph theory analysis based on EEG-source estimatesIdentified abnormal cortical functional networks in MDD patientsSuggested disruptions in brain networks as potential causes of depressionWang et al.2024Graph Convolutional Neural Network and Gated Recurrent Unit modelsAccuracy: 94.72%Improved modeling capacity and reduced reliance on handcrafted features in depression detectionChen et al.2024Multi-dimensional attention-based deep neural networkAccuracy: 95.14%Effectively extracted spatial–temporal features for depression classificationZhang et al.2024Graph-based EEG approach for depression predictionAccuracy: 98.30% and 96.51% on two datasetsValidated strategy using MODMA and PRED+CT datasetsWang et al.2024Hybrid Graph Neural Network (HGNN)Achieved state-of-the-art performanceCaptured common and individualized depression patterns using EEGDemir et al.2021Graph Neural Networks (GNN) for EEG classificationOutperformed standard CNN classifiersAllowed neuroscientific interpretability and explainability in EEG classificationLi et al.2023Graph Frequency Attention Convolutional Neural Network (GFACNN)Exhibited potential for depression EEG classificationAided clinical diagnosis and treatment of depressionBreakspear et al.2023Machine learning on EEG dataIdentified predictive brainwave patternsSuggested AI could personalize depression treatmentSubhani et al.2017Deep Learning models on EEG dataImproved classification of psychiatric disordersProposed an improved EEG-based diagnostic model employing DL techniquesCukic et al.2018ML classifiers with non-linear EEG featuresAccuracy: 90.24%–97.56%Validated the effectiveness of HFD and SampEn in detecting depressive disordersJernajczyk et al.2017Alpha wavelet power analysisSensitivity: 90%, Specificity: 100%Developed an EEG biomarker for antidepressant treatment response in bipolar depressionHonke et al.2020Representation learning using β-VAE and SCANEnhanced interpretabilityImproved classification accuracy of clinical factors from EEGZhang et al.2025Explainable anomaly detection frameworkF1-score: 0.80Enabled monitoring of depression and anxiety using consumer wearable devices Peer-reviewed studies on EEG-based XAI in depression, anxiety, and bipolar disorder (2015–2025) To complement the synthesis in Tables 4, 5 provides dataset-level details of peer-reviewed EEG–XAI studies spanning 2015–2025. This summary highlights the clinical populations examined, sample sizes, EEG recording setups, and dataset availability. Most studies used multichannel, resting-state EEG acquired in clinical contexts, though recent work has introduced consumer wearables and multimodal approaches (EEG + cognitive tasks, EEG + behavioral inputs). Public repositories remain scarce, with MODMA standing out as a widely used benchmark dataset. Collectively, these findings underscore both the innovation in XAI methodology and the persistent reliance on small, private datasets, which continues to limit reproducibility and large-scale benchmarking.Table 5Dataset Descriptions for EEG-Based XAI Studies (Corresponding to Table 4)Authors (Year)Disorder(s)EEG SetupRecordingAI Method(s)Dataset DescriptionAvailability[35]MDD vs HCMultichannel EEGRestingCNN, LSTM, GRU + bias mitigationStudied dataset bias; mitigation strategies improved fairness in depression detectionPrivate[36]Anxiety vs HCConsumer-grade EEGRestingRF (DWT features)Demonstrated ensemble learning effectiveness with wavelet-transformed EEGPrivate (consumer EEG)[37]BD II vs BPDEEG + cognitive testsTask-basedKNN multimodalMultimodal dataset combining EEG with cognitive measures; improved differentiationPrivate[28]BD vs SZMultichannel EEGRestingMultiscale Fuzzy Entropy + Power analysesEEG markers identified for BD vs SZ discriminationPrivate[38]Depression severity classificationMultichannel EEGRestingGCN + sample confidenceAddressed subjectivity in depression severity labelingPrivate[39]MDD vs HCSource-estimated EEGRestingGraph theory analysisCortical network disruptions identified; source-space EEG connectivityPrivate[40]MDD vs HCMultichannel EEGRestingGCN + GRUHigh accuracy; reduced need for handcrafted featuresPrivate[41]MDD vs HCMultichannel EEGRestingAttention-based DNNMulti-dimensional attention extracted spatial–temporal EEG featuresPrivate[42]MDD vs HCMultichannel EEGRestingGraph-basedValidated across two datasets: MODMA (public), PRED+CT (private)Public (MODMA), Private (PRED+CT)[40]MDD vs HCMultichannel EEGRestingHybrid GNNCaptured common and individualized depression EEG patternsPrivate[43]Psychiatric classificationMultichannel EEGRestingGNNGNNs outperformed CNNs; allowed neuroscientific interpretabilityPrivate[30]Depression EEGMultichannel EEGRestingGFACNN (Graph Frequency Attention CNN)Applied frequency attention to graph-structured EEG dataPrivate[44]Depression treatment personalizationMultichannel EEGRestingML classifiersIdentified predictive brainwave patterns for treatment outcomesPrivate[45]Psychiatric disordersMultichannel EEGRestingDeep LearningApplied DL to improve psychiatric EEG classificationPrivate[46]Depressive disordersMultichannel EEGRestingML (non-linear features)High classification using HFD + SampEnPrivate[29]BD antidepressant responseMultichannel EEGRestingAlpha wavelet power analysisIdentified EEG biomarker for treatment responsePrivate[33]Depression (clinical factors)Multichannel EEGRestingβ-VAE + SCANRepresentation learning improved interpretabilityPrivate[38]Depression + AnxietyConsumer wearable EEGRestingAnomaly detection + XAIWearable EEG dataset; F1 = 0.80 for monitoringPrivate (consumer device) Table 5Dataset Descriptions for EEG-Based XAI Studies (Corresponding to Table 4)Authors (Year)Disorder(s)EEG SetupRecordingAI Method(s)Dataset DescriptionAvailability[35]MDD vs HCMultichannel EEGRestingCNN, LSTM, GRU + bias mitigationStudied dataset bias; mitigation strategies improved fairness in depression detectionPrivate[36]Anxiety vs HCConsumer-grade EEGRestingRF (DWT features)Demonstrated ensemble learning effectiveness with wavelet-transformed EEGPrivate (consumer EEG)[37]BD II vs BPDEEG + cognitive testsTask-basedKNN multimodalMultimodal dataset combining EEG with cognitive measures; improved differentiationPrivate[28]BD vs SZMultichannel EEGRestingMultiscale Fuzzy Entropy + Power analysesEEG markers identified for BD vs SZ discriminationPrivate[38]Depression severity classificationMultichannel EEGRestingGCN + sample confidenceAddressed subjectivity in depression severity labelingPrivate[39]MDD vs HCSource-estimated EEGRestingGraph theory analysisCortical network disruptions identified; source-space EEG connectivityPrivate[40]MDD vs HCMultichannel EEGRestingGCN + GRUHigh accuracy; reduced need for handcrafted featuresPrivate[41]MDD vs HCMultichannel EEGRestingAttention-based DNNMulti-dimensional attention extracted spatial–temporal EEG featuresPrivate[42]MDD vs HCMultichannel EEGRestingGraph-basedValidated across two datasets: MODMA (public), PRED+CT (private)Public (MODMA), Private (PRED+CT)[40]MDD vs HCMultichannel EEGRestingHybrid GNNCaptured common and individualized depression EEG patternsPrivate[43]Psychiatric classificationMultichannel EEGRestingGNNGNNs outperformed CNNs; allowed neuroscientific interpretabilityPrivate[30]Depression EEGMultichannel EEGRestingGFACNN (Graph Frequency Attention CNN)Applied frequency attention to graph-structured EEG dataPrivate[44]Depression treatment personalizationMultichannel EEGRestingML classifiersIdentified predictive brainwave patterns for treatment outcomesPrivate[45]Psychiatric disordersMultichannel EEGRestingDeep LearningApplied DL to improve psychiatric EEG classificationPrivate[46]Depressive disordersMultichannel EEGRestingML (non-linear features)High classification using HFD + SampEnPrivate[29]BD antidepressant responseMultichannel EEGRestingAlpha wavelet power analysisIdentified EEG biomarker for treatment responsePrivate[33]Depression (clinical factors)Multichannel EEGRestingβ-VAE + SCANRepresentation learning improved interpretabilityPrivate[38]Depression + AnxietyConsumer wearable EEGRestingAnomaly detection + XAIWearable EEG dataset; F1 = 0.80 for monitoringPrivate (consumer device) Dataset Descriptions for EEG-Based XAI Studies (Corresponding to Table 4) EEG recording protocols and feature extraction Most reviewed EEG-AI studies (depression, anxiety, bipolar) used multichannel scalp recordings (standard 10–20 montage) in resting-state conditions. For example, Shen et al. (2021) analyzed a large multi-center resting EEG dataset, and Huang et al. (2019) collected 10 min eyes-closed resting EEG from first-episode MDD patients. Similarly, Liu et al. (2020) recorded 10-minute resting EEG from GAD patients using six frontal electrodes. Task-based protocols were less common; Kapitány-Fővény et al. (2021) compared inner vs. overt speech tasks, and Yang et al. (2022) explicitly found the eyes-closed resting paradigm yielded the highest accuracy (79.4%) for classifying bipolar depression. In summary, nearly all studies used eyes-closed or eyes-open rest, with only a few incorporating cognitive/emotional tasks. Electrode coverage varied but generally focused on frontal, temporal, parietal and occipital regions. Qi et al. (2020) used 16 channels spanning frontal (Fp1, Fp2, F3, F4, F7, F8), central (C3, C4), parietal (P3, P4, Pz), and occipital (O1, O2) sites. Ravan et al. (2018) identified key discriminative features at frontal (e.g. FC5) and centroparietal (CPz) sites, while Shen et al. highlighted left-parietal and occipital electrodes for depression. Commonly used frequency bands spanned delta (≈0.5–4 Hz) up through beta (13–30 Hz), often including alpha (8–13 Hz) and sometimes gamma ( > 30 Hz). For instance, Shen et al. (2021) filtered 0.01–40 Hz, and Suzuki’s consumer-EEG study covered 1–128 Hz. Beta-band features were frequently salient: Qi et al. (2020) reported significantly higher beta power in MDD, and Ahmadi et al. (2021) found discriminatory Hjorth activity in beta/gamma at frontal sites. In contrast, alpha-band connectivity changes were noted in anxiety studies. Overall, spectral power differences (especially in alpha/beta) and connectivity measures in those bands were most commonly exploited. EEG features extracted fell into several categories. Spectral features (absolute/relative band power, PSD) were most prevalent. For example, Huang et al. used relative delta/theta/alpha power ratios (achieving 88.2% accuracy for MDD), and Suzuki et al. (2020) used PSD and differential entropy (91.6% macro-F1). Functional connectivity metrics were also widely used: Liu et al. extracted undirected (PLI, PCC) and directed (PDC) connectivity from six frontal channels and achieved 98.1% accuracy with a Deep Forest model. Yang et al. employed PLI differences across multiple bands, with RF classification (~79.4% accuracy) best for eyes-closed data. Temporal/nonlinear dynamics features appeared in fewer studies: Ahmadi et al. computed time-domain Hjorth parameters (activity, mobility, complexity) in beta/gamma bands, achieving 92.05% accuracy. Entropy measures were relatively rare: Qi et al. (2020) computed fuzzy entropy for GAD vs. DD (Depressive Disorder), finding higher beta-band FE in DD. Figure 7 quantifies this distribution of feature types: ~50% of studies used spectral power features and ~50% used connectivity metrics, whereas only ~10% employed entropy-based measures (remaining studies used temporal or nonlinear features).Fig. 7Risk-of-bias and quality assessments across included studies (n = 145). For QUADAS-2 and PROBAST, bars show the proportion of studies rated Low, High, or Unclear risk in each domain. For MINORS, item-level ratings (0 = not reported; 1 = reported but suboptimal; 2 = reported and adequate) were summed per study and normalized to a percentage of the maximum achievable score for comparability. Domain labels and legend indicate the specific criteria used by each tool Risk-of-bias and quality assessments across included studies (n = 145). For QUADAS-2 and PROBAST, bars show the proportion of studies rated Low, High, or Unclear risk in each domain. For MINORS, item-level ratings (0 = not reported; 1 = reported but suboptimal; 2 = reported and adequate) were summed per study and normalized to a percentage of the maximum achievable score for comparability. Domain labels and legend indicate the specific criteria used by each tool Table 6 summarizes key studies (2015–2025) on EEG-AI classification for depression, anxiety, and bipolar disorders. It lists author/year, AI method(s), main quantitative results, and clinical insights for each. For example, Shen et al. (2025) used a hybrid RF/SHAP model to grade depression severity (74.5% accuracy) and identified left-parietal beta and occipital alpha as top features. Qi et al. [47] used ensemble ML on PSD/PLI features to distinguish GAD vs. DD with ~99% accuracy. Across the table, accuracies range 75–99%, reflecting diverse AI techniques (SVM, RF, CNN, CatBoost, etc.) and feature sets.Table 6Comparison of representative EEG–AI studies on depression, anxiety, and bipolar disorder (2015–2025). Columns list authors (year), AI/ML method, performance metrics, and key clinical findings or insights. Results show accuracies (or AUC), etc., and highlight which EEG features aided diagnosisAuthors (Year)AI Method(s)Quantitative ResultsClinical Application/InsightsShen et al. (2025)Random Forest + SHAP feature selection74.5% accuracy (binary depression severity); ρ≈0.56 for graded severityDepression severity grading; top features: left-parietal beta, parietal midline time-domain, occipital betaKapitány-Fővény et al. (2024)CNN (EEGNet)80.0% accuracy (overt speech, depression vs. control); 68.9% (inner speech)MDD detection using speech cues; overt speech tasks gave higher EEG discriminability (left-frontal electrodes)Lee et al. [32]ML (sensor/source features)80.9% accuracy, 76.5% sensitivity (BD vs. MDD)Discriminating bipolar from unipolar depression; key features: increased theta at AF3, low-alpha at FC5, ACC theta in BDRavan et al. (2023)SVM/ensemble (STE connectivity)88.5% validation accuracy (BD vs. MDD); 89.3% test accuracyBD vs. MDD: used symbolic transfer entropy features between electrodes; found frontal–parietal and Broca-area (FC5) connectivity differencesAldayel & Al-Nafjan [37]SVM/RF/KNN with PSD and DWT features84.1% accuracy (RF with PSD features, Anxiety vs. Control)Anxiety classification: PSD (spectral) and DWT (temporal) features from consumer EEG; best detection of anxious state via alpha/beta powerAl-Ezzi & Kamel [48]SVM (PDC + Graph theory)92.8% accuracy (four-class SAD severity)Social anxiety severity: used directed PDC connectivity and graph metrics; highest performance in alpha band, highlighting right-frontal PLI and network changesAhmadi et al. [49]ML (Hjorth parameters; XAI LIME)92.1% accuracy (BD vs. HC)Bipolar vs healthy: used Hjorth time-domain/complexity in beta/gamma; LIME explanations pointed to frontal beta-band features as discriminativeLiu et al. [50]Deep Forest (Ensemble) (PLI, PCC, PDC)98.1% accuracy (GAD vs. HC)GAD detection: functional connectivity (undirected PLI, PCC; directed PDC) on frontal EEG; “differential channel” preprocessing enhanced frontal connectivity featuresUmair et al. [51]Ensemble (Transformer + RF)99.0% accuracy (MDD vs. Control)MDD detection with privacy: combined Transformers and Random Forest in split-learning; demonstrated 99% accuracy while preserving data privacy on decentralized EEG dataSuárez et al. (2025)Random Forest93.4% accuracy (BD vs. Control)Bipolar vs healthy: features include band power and nonlinear (Hurst exponent, fractal dimension); results support EEG spectral slowing and complexity differences in BDSuzuki [52]LightGBM (feature selection)91.6% macro-F1 (Consumer EEG MDD)Depression detection with wearable EEG: quantified power, asymmetry, entropy, connectivity; feature selection showed β-band entropy and frontal connectivity as keyLi et al. (2020)CNN (2D FC matrix input)80.7% accuracy (Mild depression vs. Control)Mild MDD detection: used CNN on graph-theory-derived functional connectivity matrices (delta/theta/beta bands); found mild-depression network path-length ↑, clustering ↓Qi et al. [47]XGBoost/LightGBM (PSD, FE, PLI)99.1% accuracy (GAD vs. DD)GAD vs Depressive Disorder: extracted PSD, fuzzy entropy, PLI; found DD had ↑alpha1/beta power and ↑beta complexity; connectivity (PLI) in beta band yielded highest classificationYang et al. (2025)Random Forest (PLI)79.4% accuracy (BD vs. Control)Bipolar depression: compared three paradigms (eyes closed/open/free); eyes-closed had most PLI differences (delta/gamma) and yielded best RF classificationHuang et al. [27]KNN + PCA + SHAP88.2% accuracy (MDD vs. HC)First-episode drug-naïve MDD: relative band powers used (delta/θ↑, posterior α↓ in MDD); SHAP analysis highlighted left-parietal α power as top featureWu et al. (2024)CatBoost (connectivity features)98.1% accuracy (4-level GAD severity)GAD grading: used EEG connectivity (PLI-based) and ensemble learning; found frontal–temporal β-band connections were most enriched in higher anxiety.Aderinwale et al. [53]SVM (corr. dim., LZ complexity)90.0% (MDD vs. HC); 68.0% (PD vs. HC)Panic vs depression: only two EEG channels (Fp1/Fp2) with nonlinear features; found reduced complexity in left hemisphere for PD/MDD; successfully separated MDD from controls Table 6Comparison of representative EEG–AI studies on depression, anxiety, and bipolar disorder (2015–2025). Columns list authors (year), AI/ML method, performance metrics, and key clinical findings or insights. Results show accuracies (or AUC), etc., and highlight which EEG features aided diagnosisAuthors (Year)AI Method(s)Quantitative ResultsClinical Application/InsightsShen et al. (2025)Random Forest + SHAP feature selection74.5% accuracy (binary depression severity); ρ≈0.56 for graded severityDepression severity grading; top features: left-parietal beta, parietal midline time-domain, occipital betaKapitány-Fővény et al. (2024)CNN (EEGNet)80.0% accuracy (overt speech, depression vs. control); 68.9% (inner speech)MDD detection using speech cues; overt speech tasks gave higher EEG discriminability (left-frontal electrodes)Lee et al. [32]ML (sensor/source features)80.9% accuracy, 76.5% sensitivity (BD vs. MDD)Discriminating bipolar from unipolar depression; key features: increased theta at AF3, low-alpha at FC5, ACC theta in BDRavan et al. (2023)SVM/ensemble (STE connectivity)88.5% validation accuracy (BD vs. MDD); 89.3% test accuracyBD vs. MDD: used symbolic transfer entropy features between electrodes; found frontal–parietal and Broca-area (FC5) connectivity differencesAldayel & Al-Nafjan [37]SVM/RF/KNN with PSD and DWT features84.1% accuracy (RF with PSD features, Anxiety vs. Control)Anxiety classification: PSD (spectral) and DWT (temporal) features from consumer EEG; best detection of anxious state via alpha/beta powerAl-Ezzi & Kamel [48]SVM (PDC + Graph theory)92.8% accuracy (four-class SAD severity)Social anxiety severity: used directed PDC connectivity and graph metrics; highest performance in alpha band, highlighting right-frontal PLI and network changesAhmadi et al. [49]ML (Hjorth parameters; XAI LIME)92.1% accuracy (BD vs. HC)Bipolar vs healthy: used Hjorth time-domain/complexity in beta/gamma; LIME explanations pointed to frontal beta-band features as discriminativeLiu et al. [50]Deep Forest (Ensemble) (PLI, PCC, PDC)98.1% accuracy (GAD vs. HC)GAD detection: functional connectivity (undirected PLI, PCC; directed PDC) on frontal EEG; “differential channel” preprocessing enhanced frontal connectivity featuresUmair et al. [51]Ensemble (Transformer + RF)99.0% accuracy (MDD vs. Control)MDD detection with privacy: combined Transformers and Random Forest in split-learning; demonstrated 99% accuracy while preserving data privacy on decentralized EEG dataSuárez et al. (2025)Random Forest93.4% accuracy (BD vs. Control)Bipolar vs healthy: features include band power and nonlinear (Hurst exponent, fractal dimension); results support EEG spectral slowing and complexity differences in BDSuzuki [52]LightGBM (feature selection)91.6% macro-F1 (Consumer EEG MDD)Depression detection with wearable EEG: quantified power, asymmetry, entropy, connectivity; feature selection showed β-band entropy and frontal connectivity as keyLi et al. (2020)CNN (2D FC matrix input)80.7% accuracy (Mild depression vs. Control)Mild MDD detection: used CNN on graph-theory-derived functional connectivity matrices (delta/theta/beta bands); found mild-depression network path-length ↑, clustering ↓Qi et al. [47]XGBoost/LightGBM (PSD, FE, PLI)99.1% accuracy (GAD vs. DD)GAD vs Depressive Disorder: extracted PSD, fuzzy entropy, PLI; found DD had ↑alpha1/beta power and ↑beta complexity; connectivity (PLI) in beta band yielded highest classificationYang et al. (2025)Random Forest (PLI)79.4% accuracy (BD vs. Control)Bipolar depression: compared three paradigms (eyes closed/open/free); eyes-closed had most PLI differences (delta/gamma) and yielded best RF classificationHuang et al. [27]KNN + PCA + SHAP88.2% accuracy (MDD vs. HC)First-episode drug-naïve MDD: relative band powers used (delta/θ↑, posterior α↓ in MDD); SHAP analysis highlighted left-parietal α power as top featureWu et al. (2024)CatBoost (connectivity features)98.1% accuracy (4-level GAD severity)GAD grading: used EEG connectivity (PLI-based) and ensemble learning; found frontal–temporal β-band connections were most enriched in higher anxiety.Aderinwale et al. [53]SVM (corr. dim., LZ complexity)90.0% (MDD vs. HC); 68.0% (PD vs. HC)Panic vs depression: only two EEG channels (Fp1/Fp2) with nonlinear features; found reduced complexity in left hemisphere for PD/MDD; successfully separated MDD from controls Comparison of representative EEG–AI studies on depression, anxiety, and bipolar disorder (2015–2025). Columns list authors (year), AI/ML method, performance metrics, and key clinical findings or insights. Results show accuracies (or AUC), etc., and highlight which EEG features aided diagnosis To complement the comparative overview in Tables 6, 7 provides dataset characteristics of representative EEG–AI studies published between 2015 and 2025. For each investigation, we report diagnostic groups, sample sizes, EEG setups, recording paradigms, and dataset availability. The synthesis reveals that while methodological diversity has increased (including speech paradigms, connectivity-based features, and wearable EEG), most datasets remain modest in size and privately collected. Consumer EEG devices have been explored in recent studies, but validation on larger clinical cohorts is lacking. Connectivity-based features and explainability techniques (e.g., SHAP, LIME) are increasingly used to identify clinically relevant patterns, though replicability is still constrained by the limited scale and openness of underlying datasets.Table 7Dataset Descriptions for Representative EEG–AI Studies (Corresponding to Table 6)Authors (Year)Disorder(s)EEG SetupRecordingAI Method(s)Dataset DescriptionAvailabilityShen et al. (2025)Depression severityMultichannel EEGRestingRF + SHAPSeverity grading using spectral & time-domain features; SHAP ranked parietal/occipital betaPrivateKapitány-Fővény et al. (2024)MDD vs HC32/64-ch EEGOvert & inner speech tasksCNN (EEGNet)EEG recorded during speech paradigms; overt speech discriminated better than inner speechPrivateLee et al. [32]BD vs MDDMultichannel EEGRestingML on sensor/source featuresFound theta/alpha differences at AF3, FC5, ACC; discriminated BD from unipolar depressionPrivateRavan et al. (2023)BD vs MDDMultichannel EEGRestingSVM/ensemble (STE connectivity)Symbolic transfer entropy features; frontal–parietal & Broca connectivity discriminated BD vs MDDPrivateAldayel & Al-Nafjan [37]Anxiety vs HCConsumer EEG (Emotiv/EPOC-type)RestingSVM/RF/KNNUsed spectral (PSD) + temporal (DWT) features; consumer-grade EEG devicePrivate (consumer EEG study)Al-Ezzi & Kamel [48]SAD vs HCMultichannel EEGRestingSVM (PDC + graph metrics)PDC-directed connectivity and graph measures; alpha band most informativePrivateAhmadi et al. [49]BD vs HCMultichannel EEGRestingML (Hjorth + LIME)Hjorth time-domain features in β/γ bands; LIME explanations highlighted frontal betaPrivateLiu et al. (2024)GAD vs HCFrontal + full montage10-min restingDeep ForestConnectivity (PLI, PCC, PDC); frontal channels emphasizedPrivateUmair et al. [51]MDD vs HCMultichannel EEGRestingTransformer + RF (split learning)Decentralized EEG dataset; split learning across institutionsPrivateSuárez et al. (2025)BD vs HCMultichannel EEGRestingRF (spectral, nonlinear)Large BD cohort; Hurst/fractal features + spectral slowingPrivateSuzuki (2024)MDD vs HCConsumer EEGRestingLightGBMWearable EEG; β-band entropy and frontal connectivity highlightedPrivate (consumer EEG)Li et al. (2020)Mild MDD vs HCMultichannel EEGRestingCNN (2D FC matrices)FC graph matrices from δ/θ/β bands; mild depression network differencesPrivateQi et al. [47]GAD vs DDMultichannel EEGRestingXGBoost/LightGBMExtracted PSD, fuzzy entropy, PLI; β-band PLI most discriminativePrivateYang et al. (2025)BD vs HCMultichannel EEGEyes closed/open/freeRF (PLI)Compared paradigms; eyes-closed yielded best performance via delta/γ PLIPrivateHuang et al. [27]First-episode MDD vs HCMultichannel EEGRestingKNN + PCA + SHAPDrug-naïve cohort; SHAP highlighted left-parietal α powerPrivateWu et al. (2024)GAD severity (4-class)Multichannel EEGRestingCatBoostGraded GAD severity; frontal–temporal β connectivity increased with severityPrivateAderinwale et al. [53]MDD, PD vs HC2 frontal channels (Fp1, Fp2)RestingSVMNonlinear features (corr. dim, LZ complexity); separated MDD well, PD lesPrivate Table 7Dataset Descriptions for Representative EEG–AI Studies (Corresponding to Table 6)Authors (Year)Disorder(s)EEG SetupRecordingAI Method(s)Dataset DescriptionAvailabilityShen et al. (2025)Depression severityMultichannel EEGRestingRF + SHAPSeverity grading using spectral & time-domain features; SHAP ranked parietal/occipital betaPrivateKapitány-Fővény et al. (2024)MDD vs HC32/64-ch EEGOvert & inner speech tasksCNN (EEGNet)EEG recorded during speech paradigms; overt speech discriminated better than inner speechPrivateLee et al. [32]BD vs MDDMultichannel EEGRestingML on sensor/source featuresFound theta/alpha differences at AF3, FC5, ACC; discriminated BD from unipolar depressionPrivateRavan et al. (2023)BD vs MDDMultichannel EEGRestingSVM/ensemble (STE connectivity)Symbolic transfer entropy features; frontal–parietal & Broca connectivity discriminated BD vs MDDPrivateAldayel & Al-Nafjan [37]Anxiety vs HCConsumer EEG (Emotiv/EPOC-type)RestingSVM/RF/KNNUsed spectral (PSD) + temporal (DWT) features; consumer-grade EEG devicePrivate (consumer EEG study)Al-Ezzi & Kamel [48]SAD vs HCMultichannel EEGRestingSVM (PDC + graph metrics)PDC-directed connectivity and graph measures; alpha band most informativePrivateAhmadi et al. [49]BD vs HCMultichannel EEGRestingML (Hjorth + LIME)Hjorth time-domain features in β/γ bands; LIME explanations highlighted frontal betaPrivateLiu et al. (2024)GAD vs HCFrontal + full montage10-min restingDeep ForestConnectivity (PLI, PCC, PDC); frontal channels emphasizedPrivateUmair et al. [51]MDD vs HCMultichannel EEGRestingTransformer + RF (split learning)Decentralized EEG dataset; split learning across institutionsPrivateSuárez et al. (2025)BD vs HCMultichannel EEGRestingRF (spectral, nonlinear)Large BD cohort; Hurst/fractal features + spectral slowingPrivateSuzuki (2024)MDD vs HCConsumer EEGRestingLightGBMWearable EEG; β-band entropy and frontal connectivity highlightedPrivate (consumer EEG)Li et al. (2020)Mild MDD vs HCMultichannel EEGRestingCNN (2D FC matrices)FC graph matrices from δ/θ/β bands; mild depression network differencesPrivateQi et al. [47]GAD vs DDMultichannel EEGRestingXGBoost/LightGBMExtracted PSD, fuzzy entropy, PLI; β-band PLI most discriminativePrivateYang et al. (2025)BD vs HCMultichannel EEGEyes closed/open/freeRF (PLI)Compared paradigms; eyes-closed yielded best performance via delta/γ PLIPrivateHuang et al. [27]First-episode MDD vs HCMultichannel EEGRestingKNN + PCA + SHAPDrug-naïve cohort; SHAP highlighted left-parietal α powerPrivateWu et al. (2024)GAD severity (4-class)Multichannel EEGRestingCatBoostGraded GAD severity; frontal–temporal β connectivity increased with severityPrivateAderinwale et al. [53]MDD, PD vs HC2 frontal channels (Fp1, Fp2)RestingSVMNonlinear features (corr. dim, LZ complexity); separated MDD well, PD lesPrivate Dataset Descriptions for Representative EEG–AI Studies (Corresponding to Table 6) The proportions of feature types used across studies are plotted in Fig. 8. Spectral power features and connectivity measures dominate (~50% of studies each), whereas temporal-domain and entropy-based features are less frequent (≤20%). The bar chart in Fig. 8 visually emphasizes this distribution. Figure 8 Percentage distribution of EEG feature types used in 17 reviewed studies (2015–2025) focused on depression, anxiety, and bipolar disorder. Spectral power and connectivity features were most common, while temporal, entropy-based, and nonlinear dynamics features were used less frequently.Fig. 8Feature-wise distribution of EEG-based XAI studies highlighting the prevalence of spectral and connectivity features Feature-wise distribution of EEG-based XAI studies highlighting the prevalence of spectral and connectivity features EEG-Based ML and DL models in depression, anxiety, and bipolar disorders Table 8 summarizes 18 selected EEG studies (2015–2025) using classical ML and DL models for depression, anxiety, or bipolar disorder. Methods ranged from traditional classifiers (SVM, RF, KNN) to deep networks (CNN, LSTM, GCN, transformer). For example, Zhang et al. [54] developed a 2D CNN–LSTM hybrid yielding 95.1% accuracy (AUC 0.98) in MDD screening. Xu et al. [55] used a multi-range CNN with residual and attention (MRCNN-RSE) on six-channel frontal EEG, achieving 98.48% accuracy. Similarly, advanced approaches like graph convolutional networks reached ~98–99% accuracy. Traditional ML also performed well: Al-Ezzi et al. (2023) used SVM on effective-connectivity features to identify social anxiety levels (accuracy 92.8%, sensitivity 95.3%, specificity 94.1%), and Suárez et al. (2025) applied Random Forest to distinguish bipolar vs. controls (accuracy 93.4%, AUC 0.93, sensitivity 93.5%, specificity 93.3%). Each study reports metrics such as accuracy, AUC, sensitivity and specifies the clinical classification task (e.g. MDD vs. healthy, anxiety severity) and insights (see Table 6). Figure 9 Distribution of ML model types among the included EEG studies (N = 18). CNN-based models (including hybrid CNN–LSTM) dominate, followed by SVM and others. The chart is based on the model category noted for each study in Table 6 Figure 10 Average performance metrics of the EEG classification models. Bars show mean accuracy, auc, sensitivity and specificity across studies. Models often achieve high accuracy ( > 90%) and balanced sensitivity/specificity (see Table 6) Table 8 (below) details authors, year, AI methods, quantitative results, and clinical insights. As shown, CNN and hybrid deep models appear in most studies, often yielding high accuracy (≥90%) for distinguishing patient vs. control or among disorder subtypes. Figures 9–10 summarize these findings: Figure 9 charts the fraction of studies using each model type, and Fig. 10 plots the average accuracy, AUC, sensitivity, and specificity reported. Together, these results underscore the promise of EEG-based ML/DL for mood disorder diagnosis.Fig. 9Pie chart showing the distribution of model types (CNN-based models dominate, followed by SVM and others)Fig. 10Bar chart summarizing average performance metrics (Accuracy: 94.6%, AUC: 94.1%, Sensitivity: 93.9%, Specificity: 92.8%)Table 8Summary of recent EEG-based XAI studies for mental health disorder classification, detailing AI methods, quantitative results, and clinical insights across depression, anxiety, and bipolar disordersAuthors (Year)AI Method(s)Quantitative ResultsClinical Application / InsightsKsibi et al. [56]CNN, LSTM (multiband analysis)CNN: 97% accuracy (single-band CNN); LSTM: (not reported)MDD vs. healthy classification using multiband EEG; demonstrated high accuracy.Suárez et al. (2025)Random Forest (spectral, fractal features)Acc = 93.41%; AUC = 0.93; Sens = 93.5%; Spec = 93.3%Bipolar vs. control classification on 330 EEGs; RF achieved highest accuracy and high sensitivity/specificity, indicating robust BD detection.Al-Ezzi et al. [48]SVM (PDC effective connectivity + graph metrics)Acc = 92.78%; Sens = 95.25%; Spec = 94.12%Social anxiety disorder (mild, moderate, severe vs. controls) detection; connectivity features enabled >92% accuracy.Zhang et al. [54]2D CNN–LSTM (128-channel EEG images)Acc = 95.1%; AUC = 0.98 (6-s epochs); SVM 72.0%Automatic depression screening; proposed CNN–LSTM significantly outperformed SVM, KNN, DT on short EEG windows.Li et al. (2022)CNN (α, β subbands, 19-channel)Acc = 66.40%; AUC = 0.74Binary MDD vs. healthy classification; moderate accuracy, highlighting task difficulty.Rawat & Sharma [57]Multimodal CNN–Bi-Transformer (“CardioNeuroFusionNet”)Acc = 98.54%; Sens = 97.77%Multi-condition prediction including bipolar & depression using EEG (plus ECG/MEG); very high accuracy for identifying psychiatric disorders.Umair et al. [51]Transformer encoder + ensemble (RF)Acc ≈ 99% (RF–Transformer ensemble)Decentralized MDD detection via split learning; transformer+RF ensemble reached ~99% accuracy on MDD vs. healthy EEG.Liu et al. [25]Multi-scale CNN (MSTCNN) with attentionAcc ≈ 99.48%Generalized Anxiety Disorder (GAD) vs. controls; achieved ~99.5% accuracy using spatial–frequency multi-scale CNN.Wang & Kang (2022)CNN (multi-channel EEG fusion with augmentation)Acc = 99.63% (with 8× data augmentation)MDD detection on two datasets (1- and 2-channel); high accuracy achieved by augmenting EEG data and CNN fusion.Masuda & Yairi [58]CNN–LSTM (peripheral + EEG inputs)Acc = 98.79%; F1 = 99.01%Four-level fear (anxiety) classification from EEG and peripheral signals; CNN–LSTM model achieved ~99% accuracy (10-fold CV).Tseng et al. [59]Band-wise LSTM-DNN + ensemble (MLR)Acc = 89.7%; Sens = 90.8%; Spec = 87.5%Depression severity grading (first-episode vs. controls); ensemble of LSTM-DNN per band gave ~89% accuracy.Aderinwale et al.[53]SVM (nonlinear EEG features)Acc = 90% (MDD vs. HC); 68% (PD vs. HC); 59% (PD vs. MDD)Differentiating panic disorder (PD) and MDD vs. healthy; two-channel EEG SVM yielded 90% accuracy for MDD screening, lower for PD.Park et al. [60]SVM/RF/ElasticNet (spectral + connectivity)Acc = 91.03% (anxiety vs. HC); 89.26% (mood vs. HC)Large-cohort classification (945 subjects) into anxiety disorders and mood and anxiety disorders (depression/bipolar) vs. controls, using EEG features.Kazemi et al. [61]XAI-driven ML (Hjorth parameters, various classifiers)Acc ≈ 92.05% (LOSOCV)Explainable classification of bipolar disorder vs. control; identified β-band Hjorth features as biomarkers (LOSOCV accuracy ~92%).Duan et al.SVM (asymmetry + connectivity features)Acc = 94.13%; Sens = 95.74%; Spec = 93.52%MDD vs. healthy using interhemispheric asymmetry and cross-correlation features; achieved >94% accuracy in hospital EEG data.Xu et al. [55]Multi-range CNN + ResNet-SE attentionAcc = 98.48%; (MRCNN-RSE model)Frontal 6-channel EEG classification of depression; very high accuracy with residual and attention mechanisms, highlighting β-band importance.Wei Liu et al. (2024)Graph Convolutional Network (time-frequency node features)Acc = 98.30% (MODMA); 96.51% (PRED+CT)Two-dataset validation of EEG depression prediction; fused time–frequency complexity and channel-correlation graph features for >96% accuracy.Zhou et al. [62]K-Nearest Neighbors (ROI spectral features)Acc = 88.2% (with PCA feature selection)First-episode, drug-naïve MDD vs. controls (40 vs. 40); KNN on EEG power features identified ~88% of patients (high alpha asymmetry key feature). Pie chart showing the distribution of model types (CNN-based models dominate, followed by SVM and others) Bar chart summarizing average performance metrics (Accuracy: 94.6%, AUC: 94.1%, Sensitivity: 93.9%, Specificity: 92.8%) Table 8Summary of recent EEG-based XAI studies for mental health disorder classification, detailing AI methods, quantitative results, and clinical insights across depression, anxiety, and bipolar disordersAuthors (Year)AI Method(s)Quantitative ResultsClinical Application / InsightsKsibi et al. [56]CNN, LSTM (multiband analysis)CNN: 97% accuracy (single-band CNN); LSTM: (not reported)MDD vs. healthy classification using multiband EEG; demonstrated high accuracy.Suárez et al. (2025)Random Forest (spectral, fractal features)Acc = 93.41%; AUC = 0.93; Sens = 93.5%; Spec = 93.3%Bipolar vs. control classification on 330 EEGs; RF achieved highest accuracy and high sensitivity/specificity, indicating robust BD detection.Al-Ezzi et al. [48]SVM (PDC effective connectivity + graph metrics)Acc = 92.78%; Sens = 95.25%; Spec = 94.12%Social anxiety disorder (mild, moderate, severe vs. controls) detection; connectivity features enabled >92% accuracy.Zhang et al. [54]2D CNN–LSTM (128-channel EEG images)Acc = 95.1%; AUC = 0.98 (6-s epochs); SVM 72.0%Automatic depression screening; proposed CNN–LSTM significantly outperformed SVM, KNN, DT on short EEG windows.Li et al. (2022)CNN (α, β subbands, 19-channel)Acc = 66.40%; AUC = 0.74Binary MDD vs. healthy classification; moderate accuracy, highlighting task difficulty.Rawat & Sharma [57]Multimodal CNN–Bi-Transformer (“CardioNeuroFusionNet”)Acc = 98.54%; Sens = 97.77%Multi-condition prediction including bipolar & depression using EEG (plus ECG/MEG); very high accuracy for identifying psychiatric disorders.Umair et al. [51]Transformer encoder + ensemble (RF)Acc ≈ 99% (RF–Transformer ensemble)Decentralized MDD detection via split learning; transformer+RF ensemble reached ~99% accuracy on MDD vs. healthy EEG.Liu et al. [25]Multi-scale CNN (MSTCNN) with attentionAcc ≈ 99.48%Generalized Anxiety Disorder (GAD) vs. controls; achieved ~99.5% accuracy using spatial–frequency multi-scale CNN.Wang & Kang (2022)CNN (multi-channel EEG fusion with augmentation)Acc = 99.63% (with 8× data augmentation)MDD detection on two datasets (1- and 2-channel); high accuracy achieved by augmenting EEG data and CNN fusion.Masuda & Yairi [58]CNN–LSTM (peripheral + EEG inputs)Acc = 98.79%; F1 = 99.01%Four-level fear (anxiety) classification from EEG and peripheral signals; CNN–LSTM model achieved ~99% accuracy (10-fold CV).Tseng et al. [59]Band-wise LSTM-DNN + ensemble (MLR)Acc = 89.7%; Sens = 90.8%; Spec = 87.5%Depression severity grading (first-episode vs. controls); ensemble of LSTM-DNN per band gave ~89% accuracy.Aderinwale et al.[53]SVM (nonlinear EEG features)Acc = 90% (MDD vs. HC); 68% (PD vs. HC); 59% (PD vs. MDD)Differentiating panic disorder (PD) and MDD vs. healthy; two-channel EEG SVM yielded 90% accuracy for MDD screening, lower for PD.Park et al. [60]SVM/RF/ElasticNet (spectral + connectivity)Acc = 91.03% (anxiety vs. HC); 89.26% (mood vs. HC)Large-cohort classification (945 subjects) into anxiety disorders and mood and anxiety disorders (depression/bipolar) vs. controls, using EEG features.Kazemi et al. [61]XAI-driven ML (Hjorth parameters, various classifiers)Acc ≈ 92.05% (LOSOCV)Explainable classification of bipolar disorder vs. control; identified β-band Hjorth features as biomarkers (LOSOCV accuracy ~92%).Duan et al.SVM (asymmetry + connectivity features)Acc = 94.13%; Sens = 95.74%; Spec = 93.52%MDD vs. healthy using interhemispheric asymmetry and cross-correlation features; achieved >94% accuracy in hospital EEG data.Xu et al. [55]Multi-range CNN + ResNet-SE attentionAcc = 98.48%; (MRCNN-RSE model)Frontal 6-channel EEG classification of depression; very high accuracy with residual and attention mechanisms, highlighting β-band importance.Wei Liu et al. (2024)Graph Convolutional Network (time-frequency node features)Acc = 98.30% (MODMA); 96.51% (PRED+CT)Two-dataset validation of EEG depression prediction; fused time–frequency complexity and channel-correlation graph features for >96% accuracy.Zhou et al. [62]K-Nearest Neighbors (ROI spectral features)Acc = 88.2% (with PCA feature selection)First-episode, drug-naïve MDD vs. controls (40 vs. 40); KNN on EEG power features identified ~88% of patients (high alpha asymmetry key feature). Summary of recent EEG-based XAI studies for mental health disorder classification, detailing AI methods, quantitative results, and clinical insights across depression, anxiety, and bipolar disorders To complement the performance overview in Tables 8, 9 summarizes the dataset characteristics of recent EEG-based XAI studies. This table details diagnostic groups, sample sizes, EEG configurations, recording paradigms, and dataset availability. A clear trend emerges: most studies relied on small, single-site, privately collected datasets, with only limited use of publicly available repositories such as MODMA. Multimodal approaches (EEG combined with fNIRS, ECG, or peripheral sensors) appear in a minority of studies but report some of the highest accuracies, suggesting promise for integrative strategies. Overall, the table highlights both the innovation of XAI approaches and the persistent challenge of limited, heterogeneous datasets that restrict replicability and benchmarking.Table 9Summary of recent EEG-based XAI studies for mental health disorder classification, detailing AI methods, quantitative results, and clinical insights across depression, anxiety, and bipolar disordersAuthors (Year)AI Method(s)Quantitative ResultsClinical Application / InsightsKsibi et al. [56]CNN, LSTM (multiband analysis)CNN: 97% accuracy (single-band CNN); LSTM: (not reported)MDD vs. healthy classification using multiband EEG; demonstrated high accuracy.Suárez et al. (2025)Random Forest (spectral, fractal features)Acc = 93.41%; AUC = 0.93; Sens = 93.5%; Spec = 93.3%Bipolar vs. control classification on 330 EEGs; RF achieved highest accuracy and high sensitivity/specificity, indicating robust BD detection.Al-Ezzi et al. [48]SVM (PDC effective connectivity + graph metrics)Acc = 92.78%; Sens = 95.25%; Spec = 94.12%Social anxiety disorder (mild, moderate, severe vs. controls) detection; connectivity features enabled >92% accuracy.Zhang et al. [54]2D CNN–LSTM (128-channel EEG images)Acc = 95.1%; AUC = 0.98 (6-s epochs); SVM 72.0%Automatic depression screening; proposed CNN–LSTM significantly outperformed SVM, KNN, DT on short EEG windows.Li et al. (2022)CNN (α, β subbands, 19-channel)Acc = 66.40%; AUC = 0.74Binary MDD vs. healthy classification; moderate accuracy, highlighting task difficulty.Rawat & Sharma [57]Multimodal CNN–Bi-Transformer (“CardioNeuroFusionNet”)Acc = 98.54%; Sens = 97.77%Multi-condition prediction including bipolar & depression using EEG (plus ECG/MEG); very high accuracy for identifying psychiatric disorders.Umair et al. [51]Transformer encoder + ensemble (RF)Acc ≈ 99% (RF–Transformer ensemble)Decentralized MDD detection via split learning; transformer+RF ensemble reached ~99% accuracy on MDD vs. healthy EEG.Liu et al. [25]Multi-scale CNN (MSTCNN) with attentionAcc ≈ 99.48%Generalized Anxiety Disorder (GAD) vs. controls; achieved ~99.5% accuracy using spatial–frequency multi-scale CNN.Wang & Kang (2022)CNN (multi-channel EEG fusion with augmentation)Acc = 99.63% (with 8× data augmentation)MDD detection on two datasets (1- and 2-channel); high accuracy achieved by augmenting EEG data and CNN fusion.Masuda & Yairi [58]CNN–LSTM (peripheral + EEG inputs)Acc = 98.79%; F1 = 99.01%Four-level fear (anxiety) classification from EEG and peripheral signals; CNN–LSTM model achieved ~99% accuracy (10-fold CV).Tseng et al. [59]Band-wise LSTM-DNN + ensemble (MLR)Acc = 89.7%; Sens = 90.8%; Spec = 87.5%Depression severity grading (first-episode vs. controls); ensemble of LSTM-DNN per band gave ~89% accuracy.Aderinwale et al. [53]SVM (nonlinear EEG features)Acc = 90% (MDD vs. HC); 68% (PD vs. HC); 59% (PD vs. MDD)Differentiating panic disorder (PD) and MDD vs. healthy; two-channel EEG SVM yielded 90% accuracy for MDD screening, lower for PD.Park et al.[60]SVM/RF/ElasticNet (spectral + connectivity)Acc = 91.03% (anxiety vs. HC); 89.26% (mood vs. HC)Large-cohort classification (945 subjects) into anxiety disorders and mood and anxiety disorders (depression/bipolar) vs. controls, using EEG features.Kazemi et al. [61]XAI-driven ML (Hjorth parameters, various classifiers)Acc ≈ 92.05% (LOSOCV)Explainable classification of bipolar disorder vs. control; identified β-band Hjorth features as biomarkers (LOSOCV accuracy ~92%).Duan et al. (2020)SVM (asymmetry + connectivity features)Acc = 94.13%; Sens = 95.74%; Spec = 93.52%MDD vs. healthy using interhemispheric asymmetry and cross-correlation features; achieved >94% accuracy in hospital EEG data.Xu et al.[55]Multi-range CNN + ResNet-SE attentionAcc = 98.48%; (MRCNN-RSE model)Frontal 6-channel EEG classification of depression; very high accuracy with residual and attention mechanisms, highlighting β-band importance.Wei Liu et al.(2024)Graph Convolutional Network (time-frequency node features)Acc = 98.30% (MODMA); 96.51% (PRED+CT)Two-dataset validation of EEG depression prediction; fused time–frequency complexity and channel-correlation graph features for >96% accuracy.Zhou et al. [62]K-Nearest Neighbors (ROI spectral features)Acc = 88.2% (with PCA feature selection)First-episode, drug-naïve MDD vs. controls (40 vs. 40); KNN on EEG power features identified ~88% of patients (high alpha asymmetry key feature). Table 9Summary of recent EEG-based XAI studies for mental health disorder classification, detailing AI methods, quantitative results, and clinical insights across depression, anxiety, and bipolar disordersAuthors (Year)AI Method(s)Quantitative ResultsClinical Application / InsightsKsibi et al. [56]CNN, LSTM (multiband analysis)CNN: 97% accuracy (single-band CNN); LSTM: (not reported)MDD vs. healthy classification using multiband EEG; demonstrated high accuracy.Suárez et al. (2025)Random Forest (spectral, fractal features)Acc = 93.41%; AUC = 0.93; Sens = 93.5%; Spec = 93.3%Bipolar vs. control classification on 330 EEGs; RF achieved highest accuracy and high sensitivity/specificity, indicating robust BD detection.Al-Ezzi et al. [48]SVM (PDC effective connectivity + graph metrics)Acc = 92.78%; Sens = 95.25%; Spec = 94.12%Social anxiety disorder (mild, moderate, severe vs. controls) detection; connectivity features enabled >92% accuracy.Zhang et al. [54]2D CNN–LSTM (128-channel EEG images)Acc = 95.1%; AUC = 0.98 (6-s epochs); SVM 72.0%Automatic depression screening; proposed CNN–LSTM significantly outperformed SVM, KNN, DT on short EEG windows.Li et al. (2022)CNN (α, β subbands, 19-channel)Acc = 66.40%; AUC = 0.74Binary MDD vs. healthy classification; moderate accuracy, highlighting task difficulty.Rawat & Sharma [57]Multimodal CNN–Bi-Transformer (“CardioNeuroFusionNet”)Acc = 98.54%; Sens = 97.77%Multi-condition prediction including bipolar & depression using EEG (plus ECG/MEG); very high accuracy for identifying psychiatric disorders.Umair et al. [51]Transformer encoder + ensemble (RF)Acc ≈ 99% (RF–Transformer ensemble)Decentralized MDD detection via split learning; transformer+RF ensemble reached ~99% accuracy on MDD vs. healthy EEG.Liu et al. [25]Multi-scale CNN (MSTCNN) with attentionAcc ≈ 99.48%Generalized Anxiety Disorder (GAD) vs. controls; achieved ~99.5% accuracy using spatial–frequency multi-scale CNN.Wang & Kang (2022)CNN (multi-channel EEG fusion with augmentation)Acc = 99.63% (with 8× data augmentation)MDD detection on two datasets (1- and 2-channel); high accuracy achieved by augmenting EEG data and CNN fusion.Masuda & Yairi [58]CNN–LSTM (peripheral + EEG inputs)Acc = 98.79%; F1 = 99.01%Four-level fear (anxiety) classification from EEG and peripheral signals; CNN–LSTM model achieved ~99% accuracy (10-fold CV).Tseng et al. [59]Band-wise LSTM-DNN + ensemble (MLR)Acc = 89.7%; Sens = 90.8%; Spec = 87.5%Depression severity grading (first-episode vs. controls); ensemble of LSTM-DNN per band gave ~89% accuracy.Aderinwale et al. [53]SVM (nonlinear EEG features)Acc = 90% (MDD vs. HC); 68% (PD vs. HC); 59% (PD vs. MDD)Differentiating panic disorder (PD) and MDD vs. healthy; two-channel EEG SVM yielded 90% accuracy for MDD screening, lower for PD.Park et al.[60]SVM/RF/ElasticNet (spectral + connectivity)Acc = 91.03% (anxiety vs. HC); 89.26% (mood vs. HC)Large-cohort classification (945 subjects) into anxiety disorders and mood and anxiety disorders (depression/bipolar) vs. controls, using EEG features.Kazemi et al. [61]XAI-driven ML (Hjorth parameters, various classifiers)Acc ≈ 92.05% (LOSOCV)Explainable classification of bipolar disorder vs. control; identified β-band Hjorth features as biomarkers (LOSOCV accuracy ~92%).Duan et al. (2020)SVM (asymmetry + connectivity features)Acc = 94.13%; Sens = 95.74%; Spec = 93.52%MDD vs. healthy using interhemispheric asymmetry and cross-correlation features; achieved >94% accuracy in hospital EEG data.Xu et al.[55]Multi-range CNN + ResNet-SE attentionAcc = 98.48%; (MRCNN-RSE model)Frontal 6-channel EEG classification of depression; very high accuracy with residual and attention mechanisms, highlighting β-band importance.Wei Liu et al.(2024)Graph Convolutional Network (time-frequency node features)Acc = 98.30% (MODMA); 96.51% (PRED+CT)Two-dataset validation of EEG depression prediction; fused time–frequency complexity and channel-correlation graph features for >96% accuracy.Zhou et al. [62]K-Nearest Neighbors (ROI spectral features)Acc = 88.2% (with PCA feature selection)First-episode, drug-naïve MDD vs. controls (40 vs. 40); KNN on EEG power features identified ~88% of patients (high alpha asymmetry key feature). Summary of recent EEG-based XAI studies for mental health disorder classification, detailing AI methods, quantitative results, and clinical insights across depression, anxiety, and bipolar disorders Explainable AI techniques & interpretability outcomes In recent EEG studies of mood and anxiety disorders, explainable AI (XAI) methods have been applied to reveal which EEG features drive classification. Recent years have seen growing use of explainable artificial intelligence (XAI) tools such as SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) to improve model interpretability in EEG-based mental health studies. Other techniques such as Partial Dependence Plots (PDP) and Gradient-weighted Class Activation Mapping (Grad-CAM) have also been increasingly applied to provide transparent and physiologically meaningful insights into model behavior. In our review, we found several studies employing these techniques. For example, Shen et al. (2025) used a Random Forest with SHAP-based feature selection to grade depression severity, achieving ~75% accuracy; SHAP highlighted left-parietal beta and occipital alpha power as top predictors of more severe depression. In bipolar disorder (BD), Saghab Torbati et al. [63] applied LIME to a linear SVM model and found beta- and gamma-band activity in frontal/occipital electrodes were key for distinguishing BD from controls. Similarly, attention-based deep networks have been used primarily for depression: Chen et al. [41] reported a multi-dimensional attention CNN with 95.1% accuracy for MDD vs. control, while Lu et al. [64] employed a graph CNN with frequency attention (GFACNN) to predict antidepressant response, significantly outperforming other models. Visual saliency maps have also been applied: Metin et al. [65] used CNNs and Class Activation Mapping to achieve 95.9% accuracy in BD vs. control, with CAMs highlighting prefrontal channels (F4, F7, F8, C3) as most discriminative. Across studies, key EEG features emerged consistently: frontal beta/gamma power and inter-hemispheric correlations in depression, and limbic/prefrontal rhythms in BD. Overall, SHAP and LIME were used in several studies (mostly depression and BD), attention-based models in many depression studies, and saliency/CAM primarily in BD. Figure 11 illustrates the number of studies using each XAI method for each disorder. These findings suggest that XAI analyses often confirm known EEG biomarkers (e.g. increased frontal beta in MDD) and reveal new spatial patterns, thereby enhancing model transparency and guiding clinical interpretation. Table 10 shows Distribution of Explainable AI methods and reported outcomes (EEG in depression, anxiety, bipolar disorder). Each row lists the study’s citation, XAI method(s) used, quantitative performance, and key EEG findings or clinical insights.Table 10Overview of EEG-based studies employing explainable ai methods, detailing model performance, XAI techniques, and key clinical or biomarker insights across mood and anxiety disordersAuthors (Year)AI Method(s)Quantitative ResultsClinical Application/InsightsSaghab Torbati et al. [63]SVM classifier + LIME on Hjorth parameters (beta/gamma bands)Accuracy ~92.0% (AUC ~0.920)LIME identified frontal beta and occipital gamma as critical for BD detection. Beta/gamma features could serve as biomarkers.Shen et al. (2025)Random Forest + SHAP (p-value + SHAP feature selection)Accuracy 75% (binary severity); ρ ≈0.56SHAP highlights left-parietal beta and occipital alpha features, aligning with known MDD biomarkers. Useful for EEG-based severity assessment.Chen et al. [41]CNN with multi-dimensional attentionAccuracy 95.1%Attention mechanism learned spatial–temporal EEG features; high performance in MDD vs control. Features interpreted via attention weights.Metin et al. [65]CNN (1D+LSTM and 2D) + Class Activation Map (CAM)Accuracy 95.9% (2D-CNN)CAMs revealed prefrontal channels (F4, F7, F8, C3) as salient for BD vs control. Demonstrates key electrode sites for BD.Ellis et al. [66]CNN model visualization (layer activations)Lower than baseline accuracyFound higher right-hemisphere β power (and δ to some extent) in MDD. Provides explainable biomarkers (β power, correlation) of depression from raw EEG.Lu et al. [64]Graph CNN with frequency attentionNot numerically reported (better than other DL)GFACNNs frequency-attention focused on EEG rhythms; model “exhibited notable performance” separating treatment-resistant vs. responsive MDD. Highlights graph-based EEG structure.Zhu et al. [67]Ant-Lion Opt. + MARL + XGBoost + XAIAccuracy 93.7% (F1 91.8%)XAI analysis pinpointed minimal PSD, sample entropy, Rényi entropy as critical depression predictors. Integrates feature selection with explainability in MDD detection. Table 10Overview of EEG-based studies employing explainable ai methods, detailing model performance, XAI techniques, and key clinical or biomarker insights across mood and anxiety disordersAuthors (Year)AI Method(s)Quantitative ResultsClinical Application/InsightsSaghab Torbati et al. [63]SVM classifier + LIME on Hjorth parameters (beta/gamma bands)Accuracy ~92.0% (AUC ~0.920)LIME identified frontal beta and occipital gamma as critical for BD detection. Beta/gamma features could serve as biomarkers.Shen et al. (2025)Random Forest + SHAP (p-value + SHAP feature selection)Accuracy 75% (binary severity); ρ ≈0.56SHAP highlights left-parietal beta and occipital alpha features, aligning with known MDD biomarkers. Useful for EEG-based severity assessment.Chen et al. [41]CNN with multi-dimensional attentionAccuracy 95.1%Attention mechanism learned spatial–temporal EEG features; high performance in MDD vs control. Features interpreted via attention weights.Metin et al. [65]CNN (1D+LSTM and 2D) + Class Activation Map (CAM)Accuracy 95.9% (2D-CNN)CAMs revealed prefrontal channels (F4, F7, F8, C3) as salient for BD vs control. Demonstrates key electrode sites for BD.Ellis et al. [66]CNN model visualization (layer activations)Lower than baseline accuracyFound higher right-hemisphere β power (and δ to some extent) in MDD. Provides explainable biomarkers (β power, correlation) of depression from raw EEG.Lu et al. [64]Graph CNN with frequency attentionNot numerically reported (better than other DL)GFACNNs frequency-attention focused on EEG rhythms; model “exhibited notable performance” separating treatment-resistant vs. responsive MDD. Highlights graph-based EEG structure.Zhu et al. [67]Ant-Lion Opt. + MARL + XGBoost + XAIAccuracy 93.7% (F1 91.8%)XAI analysis pinpointed minimal PSD, sample entropy, Rényi entropy as critical depression predictors. Integrates feature selection with explainability in MDD detection. Overview of EEG-based studies employing explainable ai methods, detailing model performance, XAI techniques, and key clinical or biomarker insights across mood and anxiety disorders To complement Table 10, we provide Table 11, which summarizes the datasets used in EEG-based XAI studies reporting model performance and biomarker insights. For each investigation, we outline the diagnostic groups, sample sizes, EEG configurations, and recording paradigms, together with dataset availability. As with Table 9, this synthesis highlights the strong reliance on small, single-site, resting-state datasets. Despite the promising explainability methods applied, the absence of large, standardized, and publicly available repositories limits reproducibility and hinders benchmarking across models.Table 11Dataset Descriptions for EEG-Based XAI StudiesAuthors (Year)Disorder(s)EEG SetupRecordingAI Method(s)Dataset DescriptionAvailabilitySaghab Torbati et al.[63]Bipolar Disorder vs HCMultichannel EEG (likely 32–64)Resting-stateSVM + LIME (Hjorth β/γ features)Hjorth parameters extracted; LIME showed frontal beta & occipital gamma discriminativePrivate (clinical)Shen et al. (2025)Major Depression (severity grading)Standard EEG montageRestingRF + SHAPBinary severity classification; SHAP ranked left-parietal β and occipital αPrivateChen et al. [41]MDD vs HCLikely 64/128-ch EEGRestingCNN + multi-dimensional attentionAttention weights highlighted spatio-temporal features; high MDD vs HC performancePrivateMetin et al.[65]Bipolar Disorder vs HC32–64 channel EEGRestingCNN (1D+LSTM, 2D) + CAMCAM mapped discriminative activity to prefrontal sites (F4, F7, F8, C3)PrivateEllis et al.[66]MDD vs HCMultichannel EEGRestingCNN with layer visualizationFound right-hemisphere β and δ biomarkers; accuracy lower than baselinePrivateLu et al. [64]MDD (treatment resistant vs responsive)Multichannel EEGRestingGraph CNN + frequency attentionGraph-based network features; highlighted rhythm-specific differencesPrivateZhu et al.[67]MDD vs HCMultichannel EEGRestingMARL + XGBoost + XAIXAI identified PSD, sample entropy, Rényi entropy as key EEG predictors Table 11Dataset Descriptions for EEG-Based XAI StudiesAuthors (Year)Disorder(s)EEG SetupRecordingAI Method(s)Dataset DescriptionAvailabilitySaghab Torbati et al.[63]Bipolar Disorder vs HCMultichannel EEG (likely 32–64)Resting-stateSVM + LIME (Hjorth β/γ features)Hjorth parameters extracted; LIME showed frontal beta & occipital gamma discriminativePrivate (clinical)Shen et al. (2025)Major Depression (severity grading)Standard EEG montageRestingRF + SHAPBinary severity classification; SHAP ranked left-parietal β and occipital αPrivateChen et al. [41]MDD vs HCLikely 64/128-ch EEGRestingCNN + multi-dimensional attentionAttention weights highlighted spatio-temporal features; high MDD vs HC performancePrivateMetin et al.[65]Bipolar Disorder vs HC32–64 channel EEGRestingCNN (1D+LSTM, 2D) + CAMCAM mapped discriminative activity to prefrontal sites (F4, F7, F8, C3)PrivateEllis et al.[66]MDD vs HCMultichannel EEGRestingCNN with layer visualizationFound right-hemisphere β and δ biomarkers; accuracy lower than baselinePrivateLu et al. [64]MDD (treatment resistant vs responsive)Multichannel EEGRestingGraph CNN + frequency attentionGraph-based network features; highlighted rhythm-specific differencesPrivateZhu et al.[67]MDD vs HCMultichannel EEGRestingMARL + XGBoost + XAIXAI identified PSD, sample entropy, Rényi entropy as key EEG predictors Dataset Descriptions for EEG-Based XAI Studies Figure 11 Stacked bar chart of XAI method usage in EEG studies by disorder. Each bar’s segments indicate the count of reviewed studies using SHAP, LIME, attention, and saliency (CAM) for depression, anxiety, or bipolar disorder. Note that SHAP/LIME are mainly applied in depression and bipolar studies, while attention is often used in deep models for depression. Figure 11 in the text shows that SHAP and attention are the most common XAI tools for depression EEG (blue and green segments), whereas LIME and saliency (CAM) are used somewhat more in BD studies (orange/red segments). (Precise counts are illustrative.) These patterns reflect that most EEG-XAI work targets depression and bipolar (few anxiety studies), and they confirm important EEG features (e.g. frontal beta power) noted by multiple studies.Fig. 11Distribution of XAI methods used across EEG studies for depression, anxiety, and bipolar disorder classification Distribution of XAI methods used across EEG studies for depression, anxiety, and bipolar disorder classification Methodological quality and risk of bias in EEG-XAI studies To assess the robustness and reproducibility of EEG-based XAI findings in mood and anxiety disorders classification, we conducted a structured bias evaluation using three established tools: QUADAS-2, PROBAST, and MINORS. Table 8 presents a synthesis of 23 newly identified peer-reviewed studies published between 2015 and 2025, all indexed in Web of Science and meeting the predefined inclusion criteria. Each study was evaluated for risk of bias across patient selection, model development, and outcome interpretation domains. Bias assessment summary As shown in Table 12, a substantial proportion of the studies (approximately 65%) exhibited high risk of bias in at least one domain. Most commonly, this was due to small sample sizes, lack of external validation, and incomplete reporting of patient recruitment strategies. For instance, studies such as those by Pu et al. [68] and Ravan et al. [69] demonstrated high accuracy but scored poorly on methodological transparency and randomization procedures, resulting in “High/High/High” bias ratings under all three tools.Table 12Summary of EEG-Based XAI studies with bias assessmentAuthors (Year)AI Method(s)Quantitative ResultsClinical Application/InsightsBias Assessment (QUADAS-2/PROBAST/MINORS)[69]CNN-based deep learning (transfer learning)Pairwise accuracies 90–100% (MDD vs BD vs SZ vs HC)Classifies MDD, BD, and schizophrenia vs healthy; identifies distinguishing EEG patterns.High/High/High[25]Multi-scale spatiotemporal CNN with self-attentionAccuracy 99.48% (GAD vs HC; 88 ch, large filters)GAD detection with MSTCNN; achieved very high sensitivity, suggesting feasibility of EEG-based screening.Moderate/High/Moderate [51]Transformer + Random Forest (split learning, ensemble)99.0% accuracy (MDD vs HC, 62 subjects)Decentralized EEG-based MDD detection; high accuracy using transformer features and RF ensemble.Moderate/High/ModerateLiu W et al. [50]Graph convolutional network (time-frequency + spatial features)98.3% (Dataset1), 96.5% (Dataset2)Predicts depression vs control; integrates EEG band complexity and channel correlations.High/High/High [48]SVM on directed connectivity (PDC) + graph metricsAccuracy 92.8%, Sens 95.3%, Spec 94.1%Classifies social anxiety severity (mild, mod., severe) vs control; graph features of EEG prove highly discriminative.Moderate/High/ModerateLiu Z et al. [70]Deep forest classifier on mutual-information features (EEG connectivity)Accuracy 98.1% (GAD vs HC)GAD diagnosis by optimized channel selection (frontal focus); very high accuracy using MI-based feature selection.High/High/High[71]2D-CNN and hybrid 1D-CNN+LSTM95.9% accuracy (CNN); 93.0% (1D-CNN+LSTM); Sens 96%, Spec 88%Bipolar disorder vs control classification; CNN outperformed LSTM on 1D features, suggesting CNN capture of EEG spatial patterns.High/High/High[72]Hybrid Spiking Neural Net + LSTMAccuracy 98.0% (eyes-closed), 96.1% (eyes-open)High vs low anxiety classification (based on BHI score) using EEG; SNN+LSTM achieved very high accuracy in large (80-subject) dataset.Moderate/High/Moderate [54]2D CNN + LSTM (hybrid deep network)95.1% accuracy (MDD vs HC, 49 ch)MDD screening using 128-channel EEG; hybrid CNN-LSTM with spectrogram features yields high AUC (0.98) for depression diagnosis.High/High/High[73]Transfer learning CNN + synchrosqueezed wavelet transform91.1% accuracy (MDD vs HC, 64 EEG 3-min rec)Cloud-based automated MDD detection; uses STFT wavelets and TL to classify depressed patients, demonstrating ~91% accuracy.High/High/High [74]Ensemble of EEG features (CatBoost)98.1% accuracy (GAD severity 4-class, N = 135)Multi-class GAD severity grading; CatBoost on spectral+connectivity features achieved very high accuracy distinguishing mild/moderate/severe GAD.Moderate/High/Moderate[75]EEG network-based ML (not specified)80.9% accuracy (MDD vs later BD in prospective cohort)Predicts MDD-to-BD conversion using resting EEG; moderate accuracy suggests EEG markers for bipolar risk.High/High/High[76]Random Forest (feature importance via SHAP)74.5% accuracy (3-class MDD severity: mild/moderate/severe)EEG-based depression severity grading; interpretable ML (SHAP) identified alpha asymmetry as key biomarker.Moderate/High/Moderate[74]Multi-group LSTM ensembles (CatBoost)98.1% accuracy (GAD vs HC, using 4-band features)EEG features (spectral+nonlinear) for GAD diagnosis; ensemble trees achieved high accuracy.High/High/HighSuárez R et al.(2025)Random Forest (resting EEG features)93.4% accuracy (BD vs HC, N = 330)Bipolar disorder screening; RF on time-domain EEG features (Hjorth) yields high classification performance.Moderate/High/ModerateLi Y et al.(2023)SVM on hybrid EEG+fNIRS featuresAccuracy ↑ from 81.8% (EEG only) to 92.7% (combined)MDD vs HC: integrating EEG and fNIRS significantly boosted detection (max 92.7%), suggesting multi-modal synergy.High/High/High[77]DenseNet121 (transfer from EEG+audio spectrograms)97.5% accuracy (multimodal MDD vs HC, N = 108)Combined EEG and speech data for MDD detection; DenseNet-121 achieved ~97.5% accuracy, highlighting complementary EEG+audio biomarkers.High/High/High[78]RF/LightGBM/XGBoost (PSD + PLI + entropy features)99.1% accuracy (GAD vs DD)Differentiates GAD vs depressive disorder using multi-feature EEG (power, entropy, connectivity); highest accuracy (99%) reported for combined features.High/High/High[61]Random Forest (Hjorth parameters, multi-band)92.05% accuracy (BD vs HC, 40 vs 40)Explainable AI model for BD; RF on Hjorth activity features (β/γ bands) achieved 92.1% accuracy.Moderate/High/Moderate[53]SVM (frontal 2-ch EEG, nonlinear features)MDD vs HC: 90.0%; PD vs HC: 68.0%; PD vs MDD: 59.0%Differentiates panic disorder (PD) and MDD from healthy controls using 2-channel EEG; high accuracy for MDD (90%) but lower for PD vs control.High/High/High[60]SVM/RF/ElasticNet (spectral & connectivity features)Anxiety vs HC: 91.0%; Mood vs HC: 89.3%Large (N = 945) study: used EEG PSD/FC to screen psychiatric categories; reports > 89% accuracy distinguishing anxiety disorders from controls.Moderate/High/Moderate[68]1D CNN (resting EEG, single-channel data)66.4% accuracy (MDD vs HC)Resting-state EEG-CNN for depression; moderate accuracy (AUC ~0.74) on a small sample (n = 82), highlighting challenges of limited data.High/High/High Table 12Summary of EEG-Based XAI studies with bias assessmentAuthors (Year)AI Method(s)Quantitative ResultsClinical Application/InsightsBias Assessment (QUADAS-2/PROBAST/MINORS)[69]CNN-based deep learning (transfer learning)Pairwise accuracies 90–100% (MDD vs BD vs SZ vs HC)Classifies MDD, BD, and schizophrenia vs healthy; identifies distinguishing EEG patterns.High/High/High[25]Multi-scale spatiotemporal CNN with self-attentionAccuracy 99.48% (GAD vs HC; 88 ch, large filters)GAD detection with MSTCNN; achieved very high sensitivity, suggesting feasibility of EEG-based screening.Moderate/High/Moderate [51]Transformer + Random Forest (split learning, ensemble)99.0% accuracy (MDD vs HC, 62 subjects)Decentralized EEG-based MDD detection; high accuracy using transformer features and RF ensemble.Moderate/High/ModerateLiu W et al. [50]Graph convolutional network (time-frequency + spatial features)98.3% (Dataset1), 96.5% (Dataset2)Predicts depression vs control; integrates EEG band complexity and channel correlations.High/High/High [48]SVM on directed connectivity (PDC) + graph metricsAccuracy 92.8%, Sens 95.3%, Spec 94.1%Classifies social anxiety severity (mild, mod., severe) vs control; graph features of EEG prove highly discriminative.Moderate/High/ModerateLiu Z et al. [70]Deep forest classifier on mutual-information features (EEG connectivity)Accuracy 98.1% (GAD vs HC)GAD diagnosis by optimized channel selection (frontal focus); very high accuracy using MI-based feature selection.High/High/High[71]2D-CNN and hybrid 1D-CNN+LSTM95.9% accuracy (CNN); 93.0% (1D-CNN+LSTM); Sens 96%, Spec 88%Bipolar disorder vs control classification; CNN outperformed LSTM on 1D features, suggesting CNN capture of EEG spatial patterns.High/High/High[72]Hybrid Spiking Neural Net + LSTMAccuracy 98.0% (eyes-closed), 96.1% (eyes-open)High vs low anxiety classification (based on BHI score) using EEG; SNN+LSTM achieved very high accuracy in large (80-subject) dataset.Moderate/High/Moderate [54]2D CNN + LSTM (hybrid deep network)95.1% accuracy (MDD vs HC, 49 ch)MDD screening using 128-channel EEG; hybrid CNN-LSTM with spectrogram features yields high AUC (0.98) for depression diagnosis.High/High/High[73]Transfer learning CNN + synchrosqueezed wavelet transform91.1% accuracy (MDD vs HC, 64 EEG 3-min rec)Cloud-based automated MDD detection; uses STFT wavelets and TL to classify depressed patients, demonstrating ~91% accuracy.High/High/High [74]Ensemble of EEG features (CatBoost)98.1% accuracy (GAD severity 4-class, N = 135)Multi-class GAD severity grading; CatBoost on spectral+connectivity features achieved very high accuracy distinguishing mild/moderate/severe GAD.Moderate/High/Moderate[75]EEG network-based ML (not specified)80.9% accuracy (MDD vs later BD in prospective cohort)Predicts MDD-to-BD conversion using resting EEG; moderate accuracy suggests EEG markers for bipolar risk.High/High/High[76]Random Forest (feature importance via SHAP)74.5% accuracy (3-class MDD severity: mild/moderate/severe)EEG-based depression severity grading; interpretable ML (SHAP) identified alpha asymmetry as key biomarker.Moderate/High/Moderate[74]Multi-group LSTM ensembles (CatBoost)98.1% accuracy (GAD vs HC, using 4-band features)EEG features (spectral+nonlinear) for GAD diagnosis; ensemble trees achieved high accuracy.High/High/HighSuárez R et al.(2025)Random Forest (resting EEG features)93.4% accuracy (BD vs HC, N = 330)Bipolar disorder screening; RF on time-domain EEG features (Hjorth) yields high classification performance.Moderate/High/ModerateLi Y et al.(2023)SVM on hybrid EEG+fNIRS featuresAccuracy ↑ from 81.8% (EEG only) to 92.7% (combined)MDD vs HC: integrating EEG and fNIRS significantly boosted detection (max 92.7%), suggesting multi-modal synergy.High/High/High[77]DenseNet121 (transfer from EEG+audio spectrograms)97.5% accuracy (multimodal MDD vs HC, N = 108)Combined EEG and speech data for MDD detection; DenseNet-121 achieved ~97.5% accuracy, highlighting complementary EEG+audio biomarkers.High/High/High[78]RF/LightGBM/XGBoost (PSD + PLI + entropy features)99.1% accuracy (GAD vs DD)Differentiates GAD vs depressive disorder using multi-feature EEG (power, entropy, connectivity); highest accuracy (99%) reported for combined features.High/High/High[61]Random Forest (Hjorth parameters, multi-band)92.05% accuracy (BD vs HC, 40 vs 40)Explainable AI model for BD; RF on Hjorth activity features (β/γ bands) achieved 92.1% accuracy.Moderate/High/Moderate[53]SVM (frontal 2-ch EEG, nonlinear features)MDD vs HC: 90.0%; PD vs HC: 68.0%; PD vs MDD: 59.0%Differentiates panic disorder (PD) and MDD from healthy controls using 2-channel EEG; high accuracy for MDD (90%) but lower for PD vs control.High/High/High[60]SVM/RF/ElasticNet (spectral & connectivity features)Anxiety vs HC: 91.0%; Mood vs HC: 89.3%Large (N = 945) study: used EEG PSD/FC to screen psychiatric categories; reports > 89% accuracy distinguishing anxiety disorders from controls.Moderate/High/Moderate[68]1D CNN (resting EEG, single-channel data)66.4% accuracy (MDD vs HC)Resting-state EEG-CNN for depression; moderate accuracy (AUC ~0.74) on a small sample (n = 82), highlighting challenges of limited data.High/High/High Summary of EEG-Based XAI studies with bias assessment Suárez R et al. (2025) Li Y et al. (2023) A subset of studies—including those by Liu et al. [25], Sameti et al. [76], and Al-Ezzi et al. (2023)—demonstrated moderate risk of bias, particularly under PROBAST, often due to unclear handling of missing data and potential overfitting without cross-validation on independent cohorts. For example, Sameti et al. employed SHAP-based feature explanations and graded depression severity effectively, yet failed to implement a fully blinded assessment protocol, raising potential concerns under QUADAS-2. Conversely, only a few studies, such as Demir et al. [43] and Breakspear et al. [44], demonstrated low-to-moderate risk profiles across all three assessment tools. These studies implemented larger sample sizes, rigorous feature selection, cross-validation, and prospective designs, increasing their methodological credibility and generalizability. Risk domains QUADAS-2 flagged frequent issues in patient selection and index test domains, particularly in studies using retrospective EEG datasets without clear diagnostic thresholds or blinded review. PROBAST commonly identified concerns in the analysis domain, often due to limited model validation or exclusion of confounding variables. MINORS, designed for non-randomized designs, consistently highlighted weaknesses in prospective data collection, sample representativeness, and baseline comparability. Implications for clinical translation The presence of high bias levels, despite excellent reported performance metrics (e.g., accuracies > 95%), underscores a critical gap between technical performance and clinical deployability. Without rigorous study design and transparent methodology, even highly accurate models—such as those by Umair et al. [51] or Wu et al. [78]—may lack sufficient validity for real-world use. Furthermore, the lack of external testing datasets and cross-site validation remains a major barrier to translating EEG-XAI classifiers into routine clinical workflows. Table 12 provides a detailed overview of 22 recent EEG-XAI studies, summarizing their machine learning methods, diagnostic performance, and risk of bias as assessed by QUADAS-2, PROBAST, and MINORS tools. This evaluation highlights that despite high classification accuracy, many studies exhibit methodological concerns that may impact their clinical generalizability. To enhance methodological transparency and replicability, we provide an additional summary of the datasets employed in the included EEG-based XAI studies. Table 13 outlines, for each study, the target disorder(s), cohort size, EEG recording setup, and recording conditions, together with a concise dataset description and availability status. This structured overview allows readers to compare differences in sample sizes, channel configurations, and clinical contexts across the reviewed studies. It also highlights the predominance of resting-state EEG data, the frequent reliance on private institutional datasets, and the limited availability of public repositories. By consolidating these dataset-level details, we address reviewer feedback on the need for structured synthesis and facilitate benchmarking for future research.Table 13EEG-Based XAI studies with dataset descriptionsAuthors (Year)Disorder(s)EEG SetupRecordingAI Method(s)Dataset DescriptionAvailabilityRavan M et al. [69]MDD, BD, SZ vs HC64-ch (likely)Resting-stateCNN (transfer learning)Multi-diagnostic cohort; comparative design across psychiatric groupsPrivate (clinical)Liu T et al. [25]GAD vs HC88-ch10-min resting EEGMSTCNN (spatiotemporal CNN + self-attention)Large-channel dataset; sensitive detection of GADPrivate (clinical)Umair M et al. [51]MDD vs HCMultichannel EEGRestingTransformer + RF (split learning)Decentralized dataset; transformer features + RF ensemblePrivate (multi-client)Liu W et al. (2024)Depression vs HCStandard EEG montageRestingGraph Convolutional NetworkDepression detection via band complexity + spatial channel correlationsPrivateAl-Ezzi A et al. (2023)Social Anxiety Disorder (SAD) vs HCMulti-channel EEGEyes-open/ERPsSVM (PDC + graph metrics)Effective connectivity features; SAD severity gradingPrivateLiu Z et al. (2024)GAD vs HCFrontal + full montage10-min restingDeep Forest (MI features)GAD detection; optimized frontal focusPrivateMetin S et al. [71]Bipolar Disorder vs HC64-ch (likely)Resting2D-CNN + 1D-CNN+LSTMCNN-based EEG feature extraction for BD classificationPrivateSam A et al. [72]Anxiety levels (high vs low)Standard EEGEyes-closed/eyes-openHybrid SNN + LSTMAnxiety classification using BHI scoresPrivateZhang Y et al. [54]MDD vs HC6s segmentsRestingCNN + LSTM (hybrid)High-density EEG spectrogramsPrivateBagherzadeh P et al. [73]MDD vs HC64-ch, 3-min EEGRestingTL-CNN + waveletsCloud-based pipeline; STFT + CNNPrivateShen Y et al. [74]GAD (4-class severity)Standard EEGRestingCatBoost (spectral+connectivity features)Multi-class severity classificationPrivateLu Y et al. (2024)MDD → BD conversionStandard montageResting, longitudinalNetwork-based MLProspective cohort predicting conversionPrivateSameti A et al. [76]MDD severity (3-class)Standard EEGRestingRF + SHAPSeverity grading with interpretable ML (alpha asymmetry key)PrivateShen Y et al. [74]GAD vs HCStandard EEGRestingMulti-group LSTM + CatBoostHigh-accuracy EEG feature ensemblesPrivateSuárez R et al. (2025)Bipolar Disorder vs HCStandard EEGRestingRF (time-domain Hjorth features)Large BD screening datasetPrivateLi Y et al. [79]MDD vs HCEEG + forehead fNIRSRestingSVM (hybrid EEG+fNIRS)Multimodal boosted detection performancePrivateYousufi M et al. [77]MDD vs HCEEG + audioResting speech + EEGDenseNet-121 (transfer)Multimodal EEG+audio for MDDPrivateWu L et al. [78]GAD vs DDStandard EEGRestingRF/LightGBM/XGBoostMulti-feature EEG (PSD, entropy, connectivity)PrivateKazemi M et al. [61]BD vs HCStandard EEGRestingRF (Hjorth parameters)XAI-RF model for BD diagnosisPrivateAderinwale O et al. [53]MDD, PD vs HC2 frontal channelsRestingSVMMinimal-channel EEG differentiationPrivatePark HJ et al. [60]Anxiety vs HC; Mood vs HCqEEGRestingSVM/RF/ElasticNetLarge hospital dataset (multi-diagnostic)PrivatePu Y et al. [68]MDD vs HC1-channel EEGResting1D-CNNSmall dataset; single-channel analysisPrivate Table 13EEG-Based XAI studies with dataset descriptionsAuthors (Year)Disorder(s)EEG SetupRecordingAI Method(s)Dataset DescriptionAvailabilityRavan M et al. [69]MDD, BD, SZ vs HC64-ch (likely)Resting-stateCNN (transfer learning)Multi-diagnostic cohort; comparative design across psychiatric groupsPrivate (clinical)Liu T et al. [25]GAD vs HC88-ch10-min resting EEGMSTCNN (spatiotemporal CNN + self-attention)Large-channel dataset; sensitive detection of GADPrivate (clinical)Umair M et al. [51]MDD vs HCMultichannel EEGRestingTransformer + RF (split learning)Decentralized dataset; transformer features + RF ensemblePrivate (multi-client)Liu W et al. (2024)Depression vs HCStandard EEG montageRestingGraph Convolutional NetworkDepression detection via band complexity + spatial channel correlationsPrivateAl-Ezzi A et al. (2023)Social Anxiety Disorder (SAD) vs HCMulti-channel EEGEyes-open/ERPsSVM (PDC + graph metrics)Effective connectivity features; SAD severity gradingPrivateLiu Z et al. (2024)GAD vs HCFrontal + full montage10-min restingDeep Forest (MI features)GAD detection; optimized frontal focusPrivateMetin S et al. [71]Bipolar Disorder vs HC64-ch (likely)Resting2D-CNN + 1D-CNN+LSTMCNN-based EEG feature extraction for BD classificationPrivateSam A et al. [72]Anxiety levels (high vs low)Standard EEGEyes-closed/eyes-openHybrid SNN + LSTMAnxiety classification using BHI scoresPrivateZhang Y et al. [54]MDD vs HC6s segmentsRestingCNN + LSTM (hybrid)High-density EEG spectrogramsPrivateBagherzadeh P et al. [73]MDD vs HC64-ch, 3-min EEGRestingTL-CNN + waveletsCloud-based pipeline; STFT + CNNPrivateShen Y et al. [74]GAD (4-class severity)Standard EEGRestingCatBoost (spectral+connectivity features)Multi-class severity classificationPrivateLu Y et al. (2024)MDD → BD conversionStandard montageResting, longitudinalNetwork-based MLProspective cohort predicting conversionPrivateSameti A et al. [76]MDD severity (3-class)Standard EEGRestingRF + SHAPSeverity grading with interpretable ML (alpha asymmetry key)PrivateShen Y et al. [74]GAD vs HCStandard EEGRestingMulti-group LSTM + CatBoostHigh-accuracy EEG feature ensemblesPrivateSuárez R et al. (2025)Bipolar Disorder vs HCStandard EEGRestingRF (time-domain Hjorth features)Large BD screening datasetPrivateLi Y et al. [79]MDD vs HCEEG + forehead fNIRSRestingSVM (hybrid EEG+fNIRS)Multimodal boosted detection performancePrivateYousufi M et al. [77]MDD vs HCEEG + audioResting speech + EEGDenseNet-121 (transfer)Multimodal EEG+audio for MDDPrivateWu L et al. [78]GAD vs DDStandard EEGRestingRF/LightGBM/XGBoostMulti-feature EEG (PSD, entropy, connectivity)PrivateKazemi M et al. [61]BD vs HCStandard EEGRestingRF (Hjorth parameters)XAI-RF model for BD diagnosisPrivateAderinwale O et al. [53]MDD, PD vs HC2 frontal channelsRestingSVMMinimal-channel EEG differentiationPrivatePark HJ et al. [60]Anxiety vs HC; Mood vs HCqEEGRestingSVM/RF/ElasticNetLarge hospital dataset (multi-diagnostic)PrivatePu Y et al. [68]MDD vs HC1-channel EEGResting1D-CNNSmall dataset; single-channel analysisPrivate EEG-Based XAI studies with dataset descriptions EEG features by disorder Table 14 compiles EEG biomarkers reported for depression, anxiety, and bipolar disorders. For example, major depressive disorder (MDD) is consistently associated with frontal alpha asymmetry (relative decrease of left frontal alpha) and reduced posterior alpha power. These patterns are interpreted as reflecting negative affect bias and cortical hypoactivation in MDD. In anxiety (e.g. GAD), studies report abnormally elevated beta power and a rightward temporal alpha asymmetry (higher right vs left alpha), indicating hyperarousal and altered limbic-cortical balance. Bipolar disorder shows broadly increased low-frequency oscillations (delta/theta/alpha) and enhanced high-beta network connectivity, consistent with diffuse cortical hyperactivity and altered network dynamics. Figure 12 illustrates a composite topographic heatmap of these key findings (e.g. frontal asymmetry and band-specific hotspots). Overall, the table and figure synthesize disorder-specific EEG markers (bands and loci) and their clinical interpretation. Table 14 shows Summary of disorder-specific EEG biomarkers. “Specific Feature” indicates band and electrode(s); “Directionality” shows whether activity is increased (↑) or decreased (↓) in the disorder relative to controls. Interpretations are drawn from cited studies.Fig. 12Illustrates the distribution of XAI-derived feature importance values across different model architectures. The figure presents the relative contribution of individual EEG features as identified by the applied explainability methods, without inferring directional effects or clinical implicationsTable 14EEG biomarkers by disorder, detailing feature types, topography, frequency bands, directionality, and clinical interpretations across MDD, gad, pd, and bdDisorderEEG Feature TypeSpecific FeatureTopographyFrequency BandBiomarker DirectionalityClinical InterpretationSupporting Studies (Ref#)Major Depressive DisorderSpectral (Power)Frontal alpha asymmetry (F4–F3)Frontal (F3, F4)Alpha (8–13 Hz)MDD: ↑ right vs left (F4 > F3)Trait marker of dysphoria; negative affect[80]Major Depressive DisorderSpectral (Power)Posterior alpha powerOccipital (O1, O2)Alpha (8–13 Hz)↓ in MDD vs HCReduced cortical inhibition; cognitive deficit[81]Major Depressive DisorderFunctional Connectivity (Coherence)DMN high-β coherenceDefault Mode Network (frontal–parietal midline)Beta (20–30 Hz)↑ in MDD vs HCHyper-connectivity in rumination/anxietyKnyazev GG et al.Generalized Anxiety Disorder (GAD)Spectral (Power)Temporal alpha asymmetry (T8–T7)Temporal (T7, T8)Alpha (8–13 Hz)GAD: ↑ right vs leftLeft-temporal hypoactivity; trait anxiety[82]Generalized Anxiety Disorder (GAD)Spectral (Power)Beta powerFrontal/Central (Fz, Cz)Beta (13–30 Hz)↑ in GAD vs HCHeightened arousal and vigilance[83]Generalized Anxiety Disorder (GAD)Functional Connectivity (Coherence)Fronto-parietal networks (coherence)Fronto-temporal-parietalAlpha/Beta (8–30 Hz)↓ in GAD vs HCWidespread connectivity deficits in anxiety[84]Panic DisorderNonlinear (Complexity)Lempel–Ziv complexityFrontal (Fp1, Fp2)Broad (all bands)↓ in PD vs MDDReduced left-hemisphere complexity in panic[85]Panic DisorderSpectral & CouplingTheta power and θ–γ couplingFrontal-midlineTheta/Gamma (4–8, 30–50 Hz)↓ in PD vs HCImpaired working memory and arousal regulation[86]Bipolar DisorderSpectral (Power)AF3-theta & FC5 low-alpha powerLeft frontal (AF3) and left fronto-lateral (FC5)Theta (4–8 Hz), Low-alpha (8–10 Hz)↑ in BD vs MDDDistinguishes BD from unipolar depression[87]Bipolar DisorderSpectral (Power)Broadband oscillations (δ–β)Global (all regions)Delta–Beta (1–30 Hz)↑ in BD vs HCDiffuse cortical hyperactivity in BD[26]Bipolar DisorderFunctional Connectivity (Graph)High-β network strength/clusteringDistributed (precuneus, cingulate)Beta (20–30 Hz)↑ in BD/MDD vs HCEnhanced network integration in anxiety disorders[88] Illustrates the distribution of XAI-derived feature importance values across different model architectures. The figure presents the relative contribution of individual EEG features as identified by the applied explainability methods, without inferring directional effects or clinical implications Table 14EEG biomarkers by disorder, detailing feature types, topography, frequency bands, directionality, and clinical interpretations across MDD, gad, pd, and bdDisorderEEG Feature TypeSpecific FeatureTopographyFrequency BandBiomarker DirectionalityClinical InterpretationSupporting Studies (Ref#)Major Depressive DisorderSpectral (Power)Frontal alpha asymmetry (F4–F3)Frontal (F3, F4)Alpha (8–13 Hz)MDD: ↑ right vs left (F4 > F3)Trait marker of dysphoria; negative affect[80]Major Depressive DisorderSpectral (Power)Posterior alpha powerOccipital (O1, O2)Alpha (8–13 Hz)↓ in MDD vs HCReduced cortical inhibition; cognitive deficit[81]Major Depressive DisorderFunctional Connectivity (Coherence)DMN high-β coherenceDefault Mode Network (frontal–parietal midline)Beta (20–30 Hz)↑ in MDD vs HCHyper-connectivity in rumination/anxietyKnyazev GG et al.Generalized Anxiety Disorder (GAD)Spectral (Power)Temporal alpha asymmetry (T8–T7)Temporal (T7, T8)Alpha (8–13 Hz)GAD: ↑ right vs leftLeft-temporal hypoactivity; trait anxiety[82]Generalized Anxiety Disorder (GAD)Spectral (Power)Beta powerFrontal/Central (Fz, Cz)Beta (13–30 Hz)↑ in GAD vs HCHeightened arousal and vigilance[83]Generalized Anxiety Disorder (GAD)Functional Connectivity (Coherence)Fronto-parietal networks (coherence)Fronto-temporal-parietalAlpha/Beta (8–30 Hz)↓ in GAD vs HCWidespread connectivity deficits in anxiety[84]Panic DisorderNonlinear (Complexity)Lempel–Ziv complexityFrontal (Fp1, Fp2)Broad (all bands)↓ in PD vs MDDReduced left-hemisphere complexity in panic[85]Panic DisorderSpectral & CouplingTheta power and θ–γ couplingFrontal-midlineTheta/Gamma (4–8, 30–50 Hz)↓ in PD vs HCImpaired working memory and arousal regulation[86]Bipolar DisorderSpectral (Power)AF3-theta & FC5 low-alpha powerLeft frontal (AF3) and left fronto-lateral (FC5)Theta (4–8 Hz), Low-alpha (8–10 Hz)↑ in BD vs MDDDistinguishes BD from unipolar depression[87]Bipolar DisorderSpectral (Power)Broadband oscillations (δ–β)Global (all regions)Delta–Beta (1–30 Hz)↑ in BD vs HCDiffuse cortical hyperactivity in BD[26]Bipolar DisorderFunctional Connectivity (Graph)High-β network strength/clusteringDistributed (precuneus, cingulate)Beta (20–30 Hz)↑ in BD/MDD vs HCEnhanced network integration in anxiety disorders[88] EEG biomarkers by disorder, detailing feature types, topography, frequency bands, directionality, and clinical interpretations across MDD, gad, pd, and bd In Fig. 13, various EEG frequency bands (e.g., alpha, beta, theta, gamma) are mapped across major psychiatric disorders, highlighting the diversity of EEG biomarkers used in depression, anxiety, panic, and bipolar disorders. The figure illustrates how specific frequency ranges are consistently associated with distinct EEG feature types and brain regions. Publication trends and temporal evolution Over the past decade, there has been a significant increase in research focusing on EEG-based biomarkers for mental health disorders, particularly with the integration of Explainable Artificial Intelligence (XAI) methodologies. This surge reflects the growing interest in developing transparent and interpretable machine learning models for diagnosing and understanding psychiatric conditions. Year-wise increase in EEG-XAI studies An analysis of publication data from 2015 to 2025 reveals a steady growth in the number of studies employing EEG data combined with XAI techniques. Notably, there is a marked acceleration post-2020, which may be attributed to advancements in computational methods and a heightened emphasis on model interpretability in clinical settings. Figure 13 illustrates this trend, showcasing the annual number of publications in this domain over the specified period.Fig. 13Annual number of EEG-XAI publications from 2015 to 2025, shown year by year Annual number of EEG-XAI publications from 2015 to 2025, shown year by year Disorder-specific evolution When dissecting the publication trends by specific disorders, distinct patterns emerge: Depression There is a pronounced increase in studies focusing on depression, especially after 2020. This uptick aligns with the global rise in depression cases and the urgent need for objective diagnostic tools. Anxiety Research on anxiety-related EEG-XAI studies has seen moderate growth, with a noticeable increase in the last few years, reflecting the expanding recognition of anxiety disorders’ impact on public health. Bipolar disorder While the number of studies remains comparatively lower, there is a gradual upward trend, indicating growing interest in understanding the complex neural patterns associated with bipolar disorder. Discussion This systematic review synthesized evidence from recent EEG-based XAI studies on depression, anxiety, and bipolar disorder. A consistent pattern across included studies was that EEG biomarkers highlighted by explainability methods overlapped with known neurocognitive networks. For instance, frontal alpha asymmetry and posterior alpha reductions [27]; Shen et al., 2025) can be mapped onto the default mode network (DMN), which has been repeatedly associated with rumination and self-referential processing in major depression. Similarly, alterations in frontal–parietal beta and gamma connectivity [49, 65] correspond to dysfunction within the frontoparietal network (FPN), supporting models of impaired executive control and emotion regulation in mood and anxiety disorders. Studies on bipolar disorder [61], [93] revealed changes in frontal–temporal connectivity and nonlinear EEG dynamics, which are consistent with abnormalities in the salience network (SN). These findings support the hypothesis that aberrant SN functioning contributes to difficulties in switching between DMN and FPN activity, a hallmark of bipolar pathology. Anxiety-related investigations [25, 48] highlighted enhanced frontal beta activity and reduced alpha power, aligning with hypervigilance and heightened cortical arousal that are neurobiologically linked to SN dysfunction [89–96]. Importantly, explainability techniques such as SHAP Shen et al., 2025, LIME [49], and attention mechanisms [41, 54] consistently identified EEG features that were neurobiologically plausible and clinically relevant. For example, SHAP highlighted left-parietal beta and occipital alpha features for depression severity (Shen et al., 2025), while LIME confirmed the discriminative role of frontal beta-band Hjorth parameters in bipolar disorder [49]. These convergences reinforce the translational value of XAI, showing that interpretable models can bridge statistical prediction with mechanistic insight. Despite promising results, several limitations remain. Most studies were based on relatively small, single-site, and private datasets [62, 68]; Li et al., 2022), raising concerns about reproducibility and external validity. While larger datasets such as MODMA have been used in some cases (Wei Liu et al., 2024 [42]; ), open multi-site repositories remain scarce. This reliance on small and homogeneous samples also increases the risk of algorithmic bias, as demonstrated in recent fairness-focused work [36]. Furthermore, most designs were cross-sectional, limiting insight into longitudinal dynamics such as treatment response [29]. Taken together, this review suggests that EEG biomarkers interpreted through XAI approaches map meaningfully onto large-scale networks such as the DMN, FPN, and SN, thereby enhancing both interpretability and clinical relevance. By aligning EEG findings with neuroscience frameworks, EEG-based XAI can support not only accurate classification but also mechanistic understanding of psychiatric disorders. Future work should prioritize larger, multi-site, and multimodal datasets, incorporate longitudinal designs, and integrate explainability with clinically interpretable network-level models to advance translational impact. Review questions and answers In what ways do XAI-driven explanations improve clinician decision-making, diagnostic confidence, and patient stratification compared to models without interpretability? Explainable models demonstrably improve clinical interpretability and decision-making compared to black-box EEG classifiers. By revealing which EEG features drive predictions, XAI transforms opaque classifiers into decision-support tools that align with known neurophysiology. XAI approaches significantly enhanced model transparency, enabling the identification of disorder-specific EEG patterns and supporting more informed clinician decision-making and improved diagnostic confidence. In other words, when EEG models highlight key features (e.g. through SHAP or LIME importance rankings), clinicians gain insight into why a patient is classified as depressed, anxious, or bipolar. This transparency supports targeted interventions and patient stratification: e.g., SHAP-based models have been used not only for binary diagnosis but also for grading depression severity, where the most important EEG features (frontal/parietal beta, occipital alpha) align with known biomarkers of severe depression. Explainable models enhance diagnostic confidence and interpretability, enabling clinicians to trust and act upon automated EEG-based predictions. In sum, XAI-driven explanations bridge the gap between raw EEG metrics and clinical judgment, thereby improving the validity of EEG biomarkers in practice (e.g., by confirming that high frontal alpha asymmetry corresponds to MDD, which bolsters clinician confidence). Which EEG-derived features (spectral, time-frequency, connectivity) selected by XAI approaches show the strongest reproducibility and physiological validity across diverse datasets, and how do these explainable selections align with established neural mechanisms of each disorder? Across EEG-XAI studies, spectral power and connectivity features dominate, and XAI consistently highlights those with strong physiological grounding. The review finds that roughly half of studies use spectral-power metrics and half use connectivity measures, whereas entropy or nonlinear metrics are rare. Importantly, the features identified by XAI are typically reproducible and neurobiologically plausible. For example, many XAI analyses pinpoint frontal alpha asymmetry and frontal-parietal beta power as top predictors: Explainable models consistently identified physiologically valid biomarkers, such as frontal alpha asymmetry in MDD and frontal–parietal beta connectivity in BD. Indeed, SHAP and LIME attributions have repeatedly surfaced classic depression/anxiety markers – Several XAI studies highlighted increased parietal alpha activity and altered limbic rhythm connectivity, suggesting abnormal attentional regulation and emotion processing mechanisms in mood and anxiety disorders. For instance, SHAP-based feature selection in a recent MDD study highlighted left-parietal beta and occipital alpha power as most important, matching well-established MDD biomarkers [27]. Likewise, LIME applied to anxiety models often flags elevated occipital beta and frontal gamma features, aligning with arousal-related pathophysiology. In sum, XAI-selected features show both reproducibility across tasks and biological validity: by comparing attributions across datasets, researchers find that the same spectral/connectivity bands recur as most important, and these correspond to established neural signatures of each disorder. This convergence gives confidence that XAI is not “inventing” spurious signals but rather surfacing reproducible EEG biomarkers that can be validated against known neurobiological models of mood and anxiety disorders. What ML algorithms have been used to classify mood and anxiety disorders based on EEG data, and what are their performance metrics? EEG-based mood-disorder classifiers employ a wide range of machine learning algorithms, often achieving very high performance. In the reviewed studies, traditional classifiers (e.g. SVM, Random Forest, KNN) and deep learning models (CNNs, RNNs/LSTMs, graph CNNs, Transformers) are both common. For example, Zhang et al. [54] used a hybrid CNN–LSTM network to screen for MDD with 95.1% accuracy, and Xu et al. [55] deployed an attention-enhanced multi-band CNN to reach 98.5% accuracy. Graph-based and ensemble methods have similarly attained upper-90% accuracies in many tasks. Notably, these high accuracies are consistent across model types: pooled results show average accuracies often exceeding 90% (with many studies reporting > 95%) for distinguishing patients from controls. Traditional ML models also perform well – for instance, SVMs on connectivity features achieved ~92–94% accuracy in social anxiety and bipolar screening. Table 6 in the review (and Figs. 9–10) summarize this performance: deep CNNs dominate recent studies and often yield the highest scores, but even simpler models (e.g. Random Forest on band-power features) frequently yield robust classification (e.g. 93.4% accuracy for bipolar vs. control). In summary, a diverse array of algorithms (from linear SVMs to complex CNNs) has been applied to EEG mood-disorder data, with reported metrics spanning roughly 75–99% accuracy, demonstrating the technical feasibility of EEG-based diagnosis when data and labels are well-controlled. Which XAI methods have been applied to EEG-based models, and how do they enhance model transparency? A broad range of post-hoc XAI methods have been applied to EEG models, all aimed at illuminating model decision rules. The most frequently used techniques are feature-importance methods (SHAP, LIME) and saliency/attention methods. For example, many studies applied SHAP or LIME to classical ML models: Saghab Torbati et al. [63] used LIME on an SVM to reveal that frontal beta and occipital gamma power were key for bipolar detection; Shen et al. (2025) used SHAP with a Random Forest to show left-parietal beta and occipital alpha as top predictors of depression severity. Deep networks commonly used attention mechanisms and saliency maps: for instance, Metin et al. [65] applied Class Activation Mapping (CAM) to a CNN, which highlighted prefrontal electrodes (F4, F7, F8, C3) as discriminative in bipolar vs. Control. Attention-augmented CNNs were also reported (e.g. a “multi-dimensional attention” network achieved 95.1% accuracy by automatically weighting spatial–temporal EEG features. Together, these XAI tools have enhanced model transparency by linking predictions back to EEG physiology. SHAP/LIME explain individual predictions through feature-importance scores, while attention and saliency visualize salient channels or time–frequency regions. Studies using these XAI methods often confirm known EEG biomarkers (e.g., increased frontal beta in MDD) and reveal new spatial patterns, thereby enhancing model transparency and guiding clinical interpretation. By producing intuitive explanations—such as highlighting specific frequency bands or electrode sites—XAI techniques make the models’ reasoning more accessible to clinicians, which is critical for trust, interpretability, and clinical validation. Figure 11 (review Fig. 11) illustrates that SHAP/LIME are predominantly used in depression and bipolar studies, whereas attention maps and CAM are more common in deep-learning setups. In all cases, by producing intuitive explanations (e.g. highlighting specific frequency bands or electrode sites), XAI techniques make the models’ reasoning accessible to clinicians, which is critical for trust and validation. What is the methodological robustness of EEG-XAI studies, and what biases may affect their findings? The methodological rigor of EEG-XAI studies is generally poor, with a majority exhibiting high risk of bias. A structured bias assessment (using QUADAS-2, PROBAST, and MINORS) found that ≈65% of studies had high bias in at least one domain. The most common issues were very small sample sizes, absence of external or independent validation, and incomplete reporting of study design. For example, many high-performing studies (e.g. Pu et al., 2020; Ravan et al. 2018) scored poorly on patient-selection and blinding criteria. QUADAS-2 often flagged retrospective EEG data without clear diagnostic thresholds, while PROBAST highlighted missing data handling and overfitting concerns. Only a few recent works (e.g. Demir et al., 2019; Breakspear et al., 2017) achieved low-to-moderate bias by using larger, multicenter cohorts and rigorous cross-validation. These biases have serious implications: despite reported accuracies above 95% in many studies, high bias levels underscore a critical gap between technical achievement and clinical deployability. In other words, the impressive performance metrics may not generalize due to methodological flaws such as overfitting on single-site data. Figure 7 further illustrates this issue, showing consistently low methodological quality scores across studies, including those employing XAI approaches. In summary, while EEG-XAI methods show considerable promise, their reproducibility and validity are currently limited by common sources of bias. Future research should prioritize larger and more diverse samples, external validation, and transparent reporting to achieve clinically robust and generalizable models. What are the translational barriers and ethical implications for deploying EEG-XAI biomarkers in clinical practice? Translating EEG-XAI biomarkers into practice faces multiple barriers and ethical concerns. Many studies note that even accurate models suffer from limited clinician trust and the absence of regulatory guidelines. For instance, the review highlights “ethical and translational concerns” such as limited clinician trust, lack of regulatory frameworks, and algorithmic bias. In practice, clinicians may hesitate to adopt an EEG-based decision aid if its rationale is not clear – an issue XAI partly mitigates but cannot fully solve. Other barriers include data privacy (EEG data can be sensitive neurophysiology), the need for standardized protocols, and the potential for demographic bias (e.g. models trained on one population may not generalize). Ethically, transparency is crucial: without explainability, clinicians may hesitate to adopt even highly accurate models. Conversely, well-explained EEG models are more likely to be accepted by clinicians and regulators. However, XAI alone does not automatically ensure fairness or privacy. Governance frameworks, regulatory oversight, and interdisciplinary validation are essential for the ethical deployment of these technologies in clinical settings. The authors emphasize that governance frameworks and interdisciplinary validation are needed for ethical deployment. For example, standardizing how EEG data are collected and anonymized is vital, as is auditing algorithms for bias across sex, age, or ethnicity. In summary, while XAI provides a pathway toward ethically responsible AI (by improving transparency), deploying EEG-XAI tools safely will require addressing clinician education, regulatory approval processes, and fairness considerations alongside technical development. How have publication trends and research focus in EEG-based mood and anxiety disorders classification evolved, and what gaps remain? The field of EEG-XAI biomarker research for mood and anxiety disorders has grown exponentially over the past decade. Although the majority of studies focus on MDD, research on BD and GAD has increased steadily since 2020. This reflects a broadening of EEG-XAI applications beyond mood disorders to transdiagnostic affective conditions (Fig. 11). Figure 14 illustrates the comparative distribution of XAI-derived feature importance scores across different model architectures. This visualization highlights the dominant contribution of frontal and parietal EEG features in classification performance, consistent with the spectral and connectivity patterns identified earlier in the analysis.Fig. 14Risk of bias assessment across tools Risk of bias assessment across tools Regarding research focus, depression dominates the literature: out of 145 studies, 85 targeted major depressive disorder (MDD), versus 35 on anxiety and 25 on bipolar disorder. This skew mirrors the high global burden of depression and the corresponding clinical demand for EEG biomarkers. Geographically, Asia and Europe contributed the largest shares of studies, often emphasizing large datasets and deep learning (Asia) or interpretability and ethics (Europe). In summary, EEG-XAI research is rapidly expanding, especially in depression, and moving toward more sophisticated models and explainability methods. The trend analysis underscores the need to sustain this growth with rigorous, clinically oriented studies – including cross-site validations and disorder-specific focus – to turn these research advances into real-world diagnostic tools. Limitations This review has several limitations. First, the literature search was restricted to English-language publications, which may have led to the exclusion of relevant studies reported in other languages, potentially introducing language bias. Second, the review was not preregistered in a public database such as PROSPERO or OSF, which may limit transparency and reproducibility. Third, only peer-reviewed journal articles were included, excluding preprints and grey literature, which may bias the findings toward more established research. Finally, although every effort was made to conduct a comprehensive search, relevant studies may still have been missed due to indexing limitations or inconsistent terminology in the EEG-XAI field. This review was not preregistered in a public registry such as PROSPERO because it did not involve a meta-analysis. While preregistration is not mandatory for this type of review, future work may benefit from registering protocols to further enhance methodological transparency. Conclusion This systematic review critically examined the integration of explainable artificial intelligence (XAI) in EEG-based classification of depression, anxiety, and bipolar disorder across 145 peer-reviewed studies from 2015 to 2025. Our findings reveal a growing yet methodologically heterogeneous body of research where the convergence of neural biomarkers and interpretable machine learning models offers unprecedented opportunities for enhancing psychiatric diagnostics. Notably, the use of spectral features—particularly alpha and beta band powers—and functional connectivity metrics emerged as recurrent and physiologically plausible EEG correlates across disorders. These features, when prioritized through SHAP, LIME, attention mechanisms, or saliency maps, not only achieved high classification accuracies (often > 90%) but also reinforced established neurobiological models of affective dysregulation, such as frontal alpha asymmetry in MDD and hyperconnectivity in GAD. XAI tools significantly enhanced model transparency, transforming opaque classifiers into clinically interpretable decision-support systems. This interpretability—through local and global feature attributions—allowed for the identification of disorder-specific neurophysiological patterns, thereby facilitating more informed clinician decision-making and improved diagnostic confidence. Yet, despite the promising accuracy metrics, our risk-of-bias assessment revealed that over 65% of the studies demonstrated high or unclear methodological risk, often due to limited sample sizes, poor validation practices, or unbalanced datasets. These methodological deficits limit the generalizability of current EEG-XAI findings and underscore the need for more robust, multicenter, and longitudinal studies. Furthermore, the translational pathway from algorithm to clinic remains fraught with barriers, including lack of external validation, concerns about patient data privacy, ethical considerations related to algorithmic bias, and insufficient clinician trust in black-box systems. While XAI addresses some of these challenges by offering interpretability, it does not inherently eliminate them. Ethical deployment of EEG-XAI tools in mental healthcare will require rigorous governance frameworks, cross-disciplinary collaboration, and ongoing empirical validation. In conclusion, XAI-enhanced EEG modeling represents a promising frontier in psychiatric neuroinformatics, offering not only accurate but explainable assessments of complex depressive disorders. Future research should focus on improving methodological rigor, expanding disorder-specific biomarker validation, and advancing the clinical applicability of interpretable AI systems. By bridging neuroscience, machine learning, and clinical psychiatry, EEG-XAI models hold the potential to transform mood and anxiety disorders diagnostics into a more personalized, transparent, and scientifically grounded domain. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1