A Systematic Review of Mental Health Monitoring and Intervention Using Unsupervised Deep Learning on EEG Data

Abstract

Electroencephalography (EEG) is a widely used non-invasive method for capturing brain activity, offering valuable insights into cognitive and emotional states relevant to mental health. With the growing complexity and volume of EEG data, machine learning (ML) techniques—particularly deep learning—have become integral in extracting meaningful patterns. While much of the current literature focuses on supervised learning methods that rely on labeled data, unsupervised learning offers an alternative approach capable of discovering hidden structures and novel biomarkers without requiring predefined labels. This systematic review aimed to identify and synthesize recent peer-reviewed research that applied unsupervised or self-supervised learning techniques to EEG data in the context of mental health monitoring, diagnosis, or analysis. A comprehensive search was conducted across six major databases, including PubMed, Scopus, Web of Science, IEEE Xplore, PsycINFO, and Google Scholar, covering literature from January 2018 to March 2025. Following PRISMA guidelines, predefined inclusion and exclusion criteria were applied to screen and assess the relevance and quality of studies. From 512 initial records, 403 unique articles were screened, and 20 underwent full-text review. Ultimately, no studies met all the inclusion criteria. Most were excluded for employing only supervised methods, being review articles, or focusing on non-mental-health applications. The absence of eligible studies highlights a significant gap in current research and emphasizes the need for future empirical work exploring unsupervised techniques in EEG-based mental health applications. Such efforts could pave the way for more scalable, label-free approaches to understanding brain dynamics in psychological conditions.

1. Introduction

1.1. Background: EEG and Mental Health

Mental health disorders constitute a major global health challenge, contributing significantly to the worldwide burden of disease. According to the Institute for Health Metrics and Evaluation (IHME) and the World Health Organization (WHO), mental disorders affected nearly 14% of the global population in 2021 and remain among the top ten leading causes of disability-adjusted life years (DALYs) worldwide, with conditions like depression and anxiety being particularly prevalent (GBD 2019 Mental Disorders Collaborators, 2022). This substantial burden underscores the urgent need for objective, scalable, and accessible tools for the early detection, diagnosis, and monitoring of mental health conditions.

Electroencephalography (EEG) emerges as a promising candidate in this context. As a non-invasive technique, EEG measures the brain’s electrical activity with high temporal resolution, capturing the dynamic neural oscillations (summarized in

Table 1. Summary of common EEG frequency bands and associated mental states.

EEG Band	Frequency Range (Hz)	Associated Mental States/Cognitive Processes	Example Source(s)
Delta	<4	Deep sleep, certain brain injuries, and some continuous attention tasks	(Newson & Thiagarajan, 2019)
Theta	4–8	Drowsiness, light sleep, memory processing, creativity, and meditative states	(Newson & Thiagarajan, 2019)
Alpha	8–12	Relaxed wakefulness (eyes closed), attention modulation, and inhibition control	(Newson & Thiagarajan, 2019)
Beta	13–30	Active thinking, alertness, concentration, anxiety, and motor control	(Newson & Thiagarajan, 2019)
Gamma	>30	Higher cognitive functions, sensory processing, attention, and memory formation	(Newson & Thiagarajan, 2019)

) and event-related potentials that reflect underlying cognitive and emotional processes (Gao et al., 2021; Liu & Zhao, 2025). Alterations in these EEG patterns, such as changes in frequency-band power (e.g., alpha, theta, and gamma bands),connectivity between brain regions, or specific event-related potential components, have been increasingly investigated as potential biomarkers for various neuropsychiatric conditions, including depression, schizophrenia, Alzheimer’s disease, and anxiety disorders (Newson & Thiagarajan, 2019). For instance, increased beta-wave activity has often been correlated with heightened states of anxiety in EEG-based mental health research (Newson & Thiagarajan, 2019). Quantitative EEG (qEEG) methods, which apply sophisticated numerical and statistical analyses to the EEG signal, are particularly crucial for uncovering subtle patterns that may not be apparent through visual inspection, thereby enhancing the potential for biomarker discovery (Huidobro et al., 2025). The relative affordability, portability, and non-invasiveness of EEG compared to other neuroimaging techniques like fMRI or PET further bolsters its potential for widespread clinical application and mental health research.

1.2. Limitations of Supervised Learning in EEG Mental Health Analysis

Machine learning (ML), particularly supervised learning (SL), has become a cornerstone in analyzing EEG signals for mental health applications. Traditional ML approaches often utilize handcrafted features (e.g., power spectral density and connectivity measures) fed into classifiers like Support Vector Machine (SVM) or k-Nearest Neighbors (k-NN) (Langer et al., 2022). More recently, deep learning (DL) models, such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), have gained prominence due to their ability to automatically learn hierarchical features from raw or minimally processed EEG data, often achieving state-of-the-art performance in specific classification tasks (Roy et al., 2019; Vafaei & Hosseini, 2025).

Despite these advances, the reliance on supervised paradigms presents significant limitations, particularly in the complex domain of mental health (Craik et al., 2019; Vafaei & Hosseini, 2025). A primary challenge is the requirement for large, accurately labeled datasets for training. Obtaining reliable labels in mental health is notoriously difficult due to the following:

• Subjectivity and Heterogeneity: Psychiatric diagnoses often rely on subjective clinical assessments and symptom reporting, leading to variability and overlap between conditions. Diagnostic categories themselves represent heterogeneous groups, and labels may not fully capture the underlying neurobiological diversity (Vafaei & Hosseini, 2025).
• Label Scarcity and Cost: Acquiring large datasets with precise clinical labels or continuous state annotations is resource-intensive, requiring significant clinical expertise and time.
• High Dimensionality and Noise: EEG data are inherently high-dimensional, noisy, and susceptible to artifacts. While DL models can handle complexity better than traditional ML, they still require substantial data to learn robust representations and avoid overfitting (Al-Saegh et al., 2021).
• Generalizability Issues: Models trained on specific datasets often struggle to generalize to new populations, recording equipment, or experimental conditions due to inter-subject variability and domain-shift issues (Gupta et al., 2021).

Furthermore, by definition, supervised methods are constrained to identifying patterns related to the predefined labels. They may fail to discover novel, clinically relevant subtypes or subtle state transitions within the data that do not align with the existing labeling scheme. This limits their potential for exploratory analysis and uncovering fundamentally new insights into the neurobiology of mental health conditions.

1.3. The Potential of Unsupervised Learning for EEG Mental Health Analysis

Given the inherent challenges of supervised learning in the mental health domain, unsupervised learning (UL) offers a compelling alternative paradigm. UL algorithms aim to discover inherent structures, patterns, relationships, or anomalies within data without relying on predefined labels (Vafaei & Hosseini, 2025; J. Wang & Biljecki, 2022). This makes them particularly well-suited for exploratory analysis of complex, high-dimensional datasets like EEG data, where ground-truth labels may be scarce, noisy, or inadequate to capture the full spectrum of neural dynamics.

Several families of unsupervised techniques hold promise for EEG-based mental health analysis:

• Dimensionality Reduction: Techniques like Principal Component Analysis (PCA) and autoencoders (a type of neural network) can learn compressed representations of high-dimensional EEG data, potentially highlighting the most salient features or variations related to different mental states while reducing noise (Bank et al., 2021; Klein et al., 2025). Autoencoders, in particular, can learn complex non-linear mappings.
• Clustering: Algorithms such as K-Means and Gaussian Mixture Models (GMMs) can group similar EEG patterns or segments together based on intrinsic features (Alhagry et al., 2017; Ezugwu et al., 2022). This could potentially identify distinct neurophysiological subtypes within a diagnostic category, discover different brain states over time, or segment continuous EEG recordings based on underlying activity patterns. Furthermore, prior studies suggest that combining features from different EEG bands, such as the alpha and beta bands, may enhance unsupervised pattern recognition in datasets of emotional or cognitive states (Drzewiecki & Fox, 2024).
• Generative Models: Models like Generative Adversarial Networks (GANs) learn to generate synthetic data that mimics the distribution of the real data (Rathee et al., 2021; K. Wang et al., 2020). In the context of EEG, GANs could be used for data augmentation, anomaly detection (by identifying samples that the generator struggles to create), or learning latent representations that capture the underlying factors of variation in brain activity.

By leveraging these approaches, unsupervised learning can potentially be used for the following purposes:

• Discovery of Novel Biomarkers: Previously unknown EEG patterns or features associated with specific mental health conditions or states can be identified, independent of existing diagnostic criteria.
• Identification of Patient Subgroups: Heterogeneity within diagnostic groups can be uncovered based on distinct neurophysiological profiles, potentially leading to more personalized treatment strategies.
• Modeling of Baseline Activity and Detection of Deviations: Normative patterns of brain activity can be learned, and subtle deviations indicative of emerging mental health issues or treatment response can be detected.
• Overcoming of Label Scarcity: Large amounts of unlabeled EEG data can be used, which are often more readily available than labeled datasets.

Therefore, exploring the application of unsupervised learning to EEG data is crucial for the advancement of our understanding of the neural underpinnings of mental health and the development of more objective and data-driven assessment tools.

1.4. Scope and Structure of This Review

Despite the potential of unsupervised learning for analyzing EEG data in mental health, a systematic overview of its application in this specific context is lacking. Previous reviews have often focused broadly on machine learning for EEG or deep learning for mental health, sometimes including unsupervised methods but not as the primary focus (Vafaei & Hosseini, 2025). This systematic review aims to fill this gap by specifically investigating the use of unsupervised learning algorithms applied to EEG data for the purpose of understanding, monitoring, or assessing mental health conditions. We sought to identify studies published between 2018 and 2025 that employed unsupervised techniques (such as clustering, dimensionality reduction, autoencoders, GANs, etc.) to EEG signals to explore patterns related to mental health, without relying on diagnostic labels or predefined states during the core learning process. The following sections detail the methodology employed for this systematic search (Section 2), the results of the literature screening process (Section 3), a discussion of the findings and the current landscape (Section 4), and concluding remarks on the limitations of this study and future directions (Section 5).

2. Materials and Methods

This systematic review adhered to a predefined protocol to identify, screen, and synthesize research on the application of unsupervised learning techniques to EEG data for mental health monitoring. The methodology was designed to address the specific requirements of a systematic review, ensuring transparency and reproducibility.

2.1. Search Strategy

A comprehensive literature search was conducted across multiple electronic databases renowned for their coverage of biomedical, psychological, and engineering research. The selected databases included PubMed, Scopus, Web of Science, IEEE Xplore, PsycINFO, and Google Scholar. These databases were specifically chosen for their complementary coverage: PubMed and PsycINFO for biomedical and psychological research, Scopus and Web of Science for broad academic coverage, IEEE Xplore for engineering and computer science publications, and Google Scholar to capture potential gray literature with subsequent checks for peer review status. This study was preregistered on the Open Science Framework (Registration: dg2e6, https://doi.org/10.17605/OSF.IO/XUYAD).

The search strategy used combinations of keywords related to the following core concepts: EEG, unsupervised learning, and mental health. Boolean operators (AND, OR) were utilized to combine terms effectively. The primary search string structure involved the following variations:

Asterisks (*) were used for truncation to capture variations in terms. The search terms were adapted as necessary to comply with the specific syntax requirements of each database. The search was restricted to articles published between 1 January 2018 and 15 March 2025 (the date of search execution).

2.2. Inclusion and Exclusion Criteria

The studies were selected based on predefined criteria.

Inclusion criteria:

• Peer-reviewed journal articles;
• Published in English;
• Published between 2018 and 2025;
• Utilized unsupervised learning or unsupervised deep learning techniques;
• Analyzed electroencephalography (EEG) data;
• Related to mental health, mental state, mood, emotion, or psychological trait monitoring/analysis in humans.

Exclusion criteria:

• Non-peer-reviewed materials (e.g., conference abstracts, editorials, and preprints);
• Non-English publications;
• Review articles, meta-analyses, or theoretical papers without original empirical data analysis;
• Animal studies;
• Studies using only supervised learning methods;
• Studies not applying computational analysis to EEG data or using EEG for applications unrelated to mental health (e.g., seizure detection without mental health context).

2.3. Screening and Selection Process (PRISMA)

The selection process followed stages inspired by the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines:

• Identification: Search results from all databases were aggregated. Duplicates were removed using reference management software.
• Screening: Titles and abstracts of the unique articles were screened against the inclusion and exclusion criteria. This step was performed systematically.
• Eligibility: Full texts of articles deemed potentially relevant during the screening phase were retrieved and assessed in detail for final eligibility.
• Inclusion: Studies that met all inclusion criteria after full-text review were selected for data extraction and synthesis.

The results of this selection process, including the number of articles identified, selected, assessed for eligibility, and included/excluded (with reasons for exclusion), are detailed in the Results section and visually summarized in the PRISMA flow diagram (Figure 1).

2.4. Data Extraction Plan

For studies that met the inclusion criteria, the data were planned to be extracted in a structured format. Key information targeted for extraction included the following: author(s), year, study objective, participant characteristics, EEG acquisition details, unsupervised learning method(s) used, key EEG features, application/mental state studied, main findings/results, reported metrics, and dataset details. The extracted information was intended to be synthesized thematically to address the objectives of the review. However, as detailed in the Results section, no studies ultimately met all inclusion criteria.

3. Results

3.1. PRISMA Flow Diagram

The systematic literature search and selection process is summarized in the PRISMA flow diagram (Figure 1). The search yielded 512 initial records. After removing duplicates, 403 unique articles remained for screening. The selection of titles and abstracts led to the exclusion of 383 articles. The full texts of the remaining 20 articles were retrieved and assessed for eligibility. Inter-rater reliability was ensured through independent screening by two researchers, with disagreements resolved through discussion with a third reviewer.

3.2. Characteristics of Excluded Studies

Following the detailed full-text review, all 20 assessed articles were excluded. The primary reasons for exclusion are summarized below and detailed in

Table 2. Summary of Excluded Studies After Full-Text Review.

Study ID Citation/URL	Reason for Exclusion
S1#1 Frontiers Unsupervised EEG Artifact Detection (https://doi.org/10.3389/fdgth.2020.608920)	Incorrect application focus
S1#2 Sp64n3r Feature engineering of EEG applied to mental disorders… (https://doi.org/10.1007/s10489-023-04702-5)	Supervised learning only
S1#3 Clin Pract Epidemiol Ment Health Machine Learning Techniques to Predict Mental Health Diagnoses… (https://doi.org/10.2174/0117450179315688240607052117)	Supervised learning only
S1#4 Frontiers Machine learning in biosignals processing for mental health… (https://doi.org/10.3389/fpsyg.2022.1066317)	Review article
S1#5 arXiv Stress Monitoring Using Low-Cost EEG Devices… (https://arxiv.org/abs/2403.05577)	Preprint
S2#1 PMC Electroencephalography-Based Depression Detection… (https://doi.org/10.3390/s23115079)	Supervised learning only
S2#2 PubMed Depression Detection and Diagnosis Based on EEG Analysis… (PMID: 39857094)	Supervised learning only
S2#3 PMC Leveraging deep learning for robust EEG analysis… (https://doi.org/10.3389/fninf.2024.1494970)	Supervised learning only
S2#4 PubMed EEG-based Signatures of Schizophrenia, Depression… (PMID: 39248267)	Supervised learning only
S2#5 PubMed EEG-derived brainwave patterns for depression diagnosis… (PMID: 39879638)	Supervised learning only
S2#6 PubMed Machine-Learning-Based Depression Detection Model from EEG… (PMID: 39595870)	Supervised learning only
S2#7 PubMed Machine Learning-Based EEG Phenotypes of Schizophrenia… (PMID: 34721112)	Supervised learning only
S2#8 PubMed Comparing resting state and task-based EEG using ML… (PMID: 37156879)	Supervised learning only
S3#1 ScienceDirect Psychiatric disorders from EEG signals through deep learning… (https://doi.org/10.1016/j.biopsycho.2021.108117)	Supervised learning only
S3#2 Wiley Automated Detection of Neurological and Mental Health Disorders… (https://doi.org/10.1002/widm.70002)	Review article
S3#3 ScienceDirect Deep learning-based feature extraction for EEG signals… (https://doi.org/10.1016/j.neucom.2023.126805)	Supervised learning only
S3#4 Frontiers Reproducible machine learning research in mental workload… (https://doi.org/10.3389/fnrgo.2024.1346794)	Review article
S3#5 PMC Enhancing EEG-Based Emotion Detection with Hybrid Models… (https://doi.org/10.3390/s24072217)	Supervised learning only
S3#6 Nature Predicting treatment response using EEG in major depressive disorder… (https://doi.org/10.1038/s41398-022-02064-z)	Supervised learning only
S3#7 Research Square Mental Health Monitoring And Intervention Using Unsupervised… (https://doi.org/10.21203/rs.3.rs-5014270/v1)	Preprint

.

• Use of Supervised Learning Only: The majority of excluded studies (12 articles) were removed because they exclusively employed supervised learning methods, failing to meet the core inclusion criterion of utilizing unsupervised techniques.
• Review Articles: Several studies (six articles) were identified as review articles (systematic, narrative, or guideline reviews) rather than original empirical research, leading to their exclusion.
• Incorrect Application Focus: One study focused solely on EEG artifact detection, without a direct link to mental health state analysis.
• Inaccessible Full Text: One article could not be accessed due to a broken link.
• Preprints: Two articles were identified as preprints and, thus, excluded.

3.3. Summary: No Eligible Studies Found

Ultimately, zero (0) studies were found to meet all the predefined inclusion criteria for this systematic review. No peer-reviewed original research articles published between 2018 and 2025 were identified that specifically utilized unsupervised learning techniques for the analysis of EEG data in the context of human mental health monitoring.

4. Discussion

This systematic review aimed to map the landscape of recent (2018–2025) peer-reviewed research applying unsupervised learning techniques to EEG data for mental health analysis. The striking result is the complete absence of eligible studies meeting our specific inclusion criteria.

4.1. Interpretation of Findings: Why the Gap?

The lack of identified studies suggests a significant gap between the theoretical potential of unsupervised learning and its empirical application in recent peer-reviewed EEG mental health research. Several factors likely contribute:

• Dominance of Supervised Paradigms: The field of machine learning applied to EEG, particularly in clinical contexts, remains heavily focused on supervised approaches. As highlighted by reviews, much research prioritizes the development and application of supervised models (e.g., SVMs and CNNs) for tasks such as classification and decoding based on predefined labels (Al-Saegh et al., 2021). This focus is understandable, given the clear clinical need for diagnostic classification, prediction of treatment outcomes, or identification of specific brain states corresponding to known conditions. Supervised methods offer a direct path to evaluating performance against established ground truths, aligning well with traditional clinical research goals and validation frameworks (Krol et al., 2023). This inherent alignment with classification-driven objectives may lead researchers to prioritize supervised techniques over the more exploratory nature of unsupervised learning.
• Technical Challenges: Applying unsupervised methods to noisy, high-dimensional EEG data is complex. Discovering meaningful, replicable patterns without labels requires sophisticated algorithms and robust validation strategies, which may be less developed or standardized than supervised counterparts.
• Validation Hurdles: Establishing the clinical relevance of patterns found via unsupervised learning (e.g., novel clusters) is difficult, requiring extensive correlation with external measures or longitudinal tracking.
• Publication Bias and Reporting: Negative or exploratory results from unsupervised analyses might be less likely to be published. Furthermore, unsupervised techniques might be used adjunctively within larger studies focused on supervised outcomes and, thus, not highlighted or indexed appropriately.
• Terminology and Indexing: Our search terms, though broad, might have missed studies using niche terminology for specific unsupervised algorithms.

4.2. Role and Limitations of Supervised Methods in EEG Research

While our focus was unsupervised learning, the screening process underscored the prevalence of supervised approaches in EEG-based mental health research. These methods have achieved notable success in specific, well-defined tasks. For instance, supervised classifiers, ranging from traditional machine learning algorithms like Support Vector Machine (SVM) and k-Nearest Neighbors (kNN) to deep learning models like convolutional neural networks (CNNs), have been widely applied to classify mental states such as stress levels (Al-Saegh et al., 2021), detect neurological disorders like epilepsy, or decode cognitive and emotional states based on labeled EEG data (Gao et al., 2021). The strength of supervised learning lies in its ability to learn direct mappings from EEG features to predefined clinical labels or behavioral outcomes, providing quantifiable performance metrics (e.g., accuracy and F1 score) that are readily interpretable in a clinical validation context.

However, the reliance on labeled data constitutes a significant limitation. Acquiring large, accurately labeled EEG datasets for diverse mental health conditions is challenging and expensive. Labels often rely on subjective clinical assessments or self-reports, which can be noisy or inconsistent or fail to capture the full spectrum of a condition (Krol et al., 2023). This dependence on potentially limited or biased labeled data restricts the ability of supervised models to generalize to unseen populations or capture the inherent heterogeneity within complex mental health disorders. Furthermore, supervised models are typically trained for specific tasks and may not easily discover novel, data-driven insights or biomarkers hidden within the complex EEG signals.

4.3. Potential of Unsupervised Learning in EEG Mental Health: Neuroscience Foundations

Despite the gap found in our systematic search, the theoretical potential of unsupervised learning for advancing EEG-based mental health research remains substantial. These methods offer a powerful lens to explore the inherent complexity and heterogeneity within psychiatric conditions, moving beyond predefined labels. To understand why unsupervised learning approaches are particularly well-suited for EEG-based mental health analysis, it is essential to ground these technical approaches in established neuroscience theories and brain models.

4.3.1. Predictive Coding and Free Energy Principle: Theoretical Foundations

Unsupervised learning approaches align remarkably well with prominent neuroscience theories of brain function, particularly the predictive coding framework and the free energy principle. According to predictive coding theory, the brain operates as a prediction machine that continuously generates and updates internal models of the world based on sensory inputs (Caucheteux et al., 2023; Ficco et al., 2021). The brain’s hierarchical structure enables predictions to flow from the top–down while prediction errors propagate in a bottom–up manner, with each level attempting to minimize prediction error (Marais et al., 2025). This process bears striking resemblance to how unsupervised learning algorithms function—extracting patterns and building internal representations without explicit external guidance.

The free energy principle extends this concept, proposing that biological systems, including the brain, act to minimize surprise (or free energy) by either updating their internal models or changing their interactions with the environment (Friston et al., 2023; Isomura et al., 2023). Mental health disorders can be conceptualized within this framework as disturbances in predictive processing, where the brain’s ability to generate accurate predictions or appropriately weight prediction errors becomes compromised (Schilling et al., 2023).

Unsupervised learning methods mirror these neurobiological processes by

• Extracting latent patterns from complex data without predefined labels, similar to how the brain identifies regularities in sensory input;
• Building hierarchical representations that capture increasingly abstract features, paralleling the brain’s cortical hierarchy;
• Minimizing various forms of reconstruction error or statistical divergence, analogous to the brain’s effort to minimize prediction error or free energy.

This theoretical alignment suggests that unsupervised approaches may be particularly well-suited for modeling the brain’s intrinsic activity as captured by EEG, potentially offering more neurobiologically plausible insights than supervised methods that impose external categorical frameworks (Hodson et al., 2024).

4.3.2. Neuroscience-Informed Unsupervised Learning for Mental Health

Mental health conditions involve complex alterations in brain dynamics that often defy simple categorization. Recent neuroscience research suggests that psychiatric disorders may be better understood as dimensional disturbances across multiple neural systems rather than discrete categories (Baydili et al., 2025; Uyanik et al., 2025). Unsupervised learning approaches are uniquely positioned to capture this complexity by

• Identifying natural groupings in neural activity that may correspond to biologically meaningful subtypes within and across diagnostic boundaries;
• Detecting subtle deviations from typical brain dynamics that might represent early markers of mental health conditions;
• Characterizing the dimensional nature of psychopathology by mapping continuous variations in neural patterns.

General conceptual frameworks illustrate how raw EEG signals can be processed using various unsupervised techniques like clustering, dimensionality reduction, and generative models to derive insights (Figure 2).

Furthermore, specific architectures for real-time mental health monitoring using unsupervised learning on EEG data have been conceptualized, although not yet widely implemented or validated in peer-reviewed studies within our search scope. One such conceptual approach involves the preprocessing of EEG data; performing exploratory visualization and quality checks; applying unsupervised methods like PCA for dimensionality reduction, followed by clustering algorithms (e.g., K-Means and GMM) to identify patterns; and evaluating the results. This systematic strategy, adapted from concepts outlined in preliminary work, aims to identify latent structures or anomalies in EEG signals that could correspond to different mental states or potential issues, facilitating timely intervention.

Applying such unsupervised approaches holds promise in several key areas:

• Discovering Novel Subtypes: Unsupervised clustering techniques hold significant promise for identifying data-driven subtypes within traditional diagnostic categories based on neurophysiological signatures. For example, Y. Zhang et al. (2020) successfully applied unsupervised and supervised machine learning to resting-state EEG functional connectivity patterns (power envelope connectivity) to identify two distinct and clinically relevant subtypes within both PTSD and Major Depressive Disorder (MDD) across multiple datasets (Yan et al., 2023). Crucially, these neurobiologically defined subtypes showed differential responses to treatment (psychotherapy for PTSD and medication for MDD), suggesting their potential utility in personalizing interventions, a goal often difficult to achieve with symptom-based diagnoses alone. Visualizing the results of clustering algorithms applied to dimensionally reduced EEG data can provide insights into the separability of potential groups (Figure 3).
• Dimensionality Reduction and Feature Extraction: The high dimensionality of EEG data presents a challenge. Unsupervised dimensionality reduction methods, such as Principal Component Analysis (PCA) and Independent Component Analysis (ICA) or more complex non-linear techniques like autoencoders, can effectively reduce data complexity, extract salient features representing underlying neural dynamics, denoise signals, and facilitate visualization (Roy et al., 2019). These extracted features can then potentially be used in subsequent supervised or unsupervised analyses.
• Generative Models and Anomaly Detection: Generative models, like generative adversarial networks (GANs) or Variational Autoencoders (VAEs), can learn the distribution of normative brain activity from large EEG datasets. Once trained, these models could be used for anomaly detection, potentially identifying subtle deviations from typical patterns that might indicate early stages of mental health decline or treatment response (K. Wang et al., 2020). They can also be used to generate realistic synthetic EEG data to augment limited datasets for the training of more robust supervised models.
• Exploring Latent Structures: Techniques borrowed from other domains and commonly used in natural language processing, such as topic modeling (e.g., latent Dirichlet allocation), might be adapted to EEG data to uncover latent ‘neural topics’ or recurring patterns of brain activity associated withdifferent mental states or cognitive processes, without prior hypotheses.

4.4. Implications for Clinical Research and Practice

The apparent underutilization of unsupervised learning in the published EEG mental health literature, as revealed by this review, signifies missed opportunities for advancing clinical research and practice. While supervised methods excel at confirming hypotheses based on existing labels, unsupervised approaches offer a complementary path to discovery, potentially yielding insights with significant clinical implications:

• Objective Biomarker Discovery: Unsupervised learning can analyze complex EEG data without prior assumptions, potentially uncovering novel, objective neurophysiological biomarkers associated with mental health conditions, treatment response, or risk prediction (Yun, 2024). Such biomarkers could augment or even eventually replace subjective symptom scales, leading to more reliable diagnostics and monitoring.
• Patient Stratification and Personalized Medicine: Unsupervised clustering can reveal underlying biological heterogeneity masked by broad diagnostic labels. This approach has shown promise in related neuroimaging domains, such as fMRI-based studies that identified distinct neurobiological subtypes within major depressive disorder (Langer et al., 2022). Similar applications to EEG data could enable more personalized treatment selection, predicting which patients are likely to respond to specific therapies (e.g., medication vs. psychotherapy vs. neuromodulation) and improving overall treatment efficacy (Simmatis et al., 2023).
• Understanding Pathophysiology: Data-driven patterns discovered through unsupervised analysis (e.g., altered connectivity networks and specific oscillatory patterns) can generate new hypotheses about the neurobiological mechanisms underlying mental illness, guiding further basic and clinical research. For instance, recent work in genomics has demonstrated how unsupervised clustering of gene expression data can reveal novel disease subtypes (Cai et al., 2022), suggesting that parallel approaches could be valuable in EEG analysis.
• Early Detection and Monitoring: Unsupervised anomaly detection models trained on normative EEG data could potentially identify subtle, early signs of mental health decline or relapse before symptoms become clinically apparent, enabling proactive interventions (Al-Saegh et al., 2021). Similarly, they could provide objective measures for tracking treatment progress over time. This approach has shown promise in other medical domains, such as in detecting anomalies in cardiac signals (Demirezen et al., 2024).

Integrating unsupervised findings with supervised models could lead to more robust, nuanced, and clinically actionable decision support systems, ultimately improving patient outcomes in mental health care (Al-Saegh et al., 2021). The lack of focus on these approaches represents a significant bottleneck in translating EEG’s potential into tangible clinical tools.

4.5. Recommendations for Future Research

Addressing the identified gap and harnessing the potential of unsupervised learning in EEG mental health research requires a concerted, multi-faceted effort. Based on the discussed challenges and opportunities, future research should prioritize the following:

• Multimodal Integration: While this review focused on EEG data, future research should explore multimodal approaches that combine EEG with complementary neuroimaging techniques such as functional near-infrared spectroscopy (fNIRS), functional magnetic resonance imaging (fMRI), or physiological measures. Multimodal data can provide more comprehensive insights into brain function and potentially enhance the performance of unsupervised learning algorithms by leveraging complementary information (Uyanik et al., 2025).
• Expanded Definition of Unsupervised Learning: Future studies should consider a broader spectrum of unsupervised and self-supervised learning approaches. Self-supervised learning, which creates supervisory signals from the data itself without external labels, represents a promising middle ground between fully supervised and unsupervised approaches (X. Zhang et al., 2022). Similarly, semi-supervised methods that leverage small amounts of labeled data alongside larger unlabeled datasets could help bridge the gap between these paradigms (Krol et al., 2023; Sosulski et al., 2021).
• Hybrid Approaches: Research should explore hybrid models that combine unsupervised components with supervised fine-tuning. For example, using unsupervised methods for feature learning or dimensionality reduction followed by supervised classification has shown promise in other domains and could be particularly valuable for EEG analysis (Kinahan et al., 2024).
• Model Interpretability: As unsupervised methods often discover complex patterns, research should prioritize interpretable models or post hoc explanation techniques that can translate mathematical findings into clinically meaningful insights. This is essential for clinical adoption and scientific advancement (Géron, 2022).
• Standardized Evaluation Frameworks: Developing robust validation frameworks for unsupervised EEG analysis is crucial. This includes creating benchmark datasets with diverse populations, standardized preprocessing pipelines, and metrics for evaluating the clinical relevance of discovered patterns (Khan et al., 2022).
• Algorithm Development and Validation: There is a critical need to develop and rigorously validate unsupervised algorithms specifically tailored for the unique characteristics of EEG data, such as high dimensionality, a low signal-to-noise ratio, non-stationarity, and inter-individual variability (Al-Saegh et al., 2021). Methods should be robust to artifacts and capable of handling longitudinal data. Validation should move beyond technical metrics and focus on establishing the clinical relevance and replicability of discovered patterns. For instance, while various unsupervised techniques like PCA, K-means, and GMM have been conceptualized for EEG analysis, their empirical validation specifically for mental health monitoring in large, diverse cohorts remains limited.
• Exploration of Specific EEG Features: While frequency-band analysis is common, certain features, like gamma-band activity, remain relatively underutilized in unsupervised EEG research for mental health, suggesting a critical gap for future exploration (Newson & Thiagarajan, 2019). Investigating how different unsupervised models capture information from various frequency bands and connectivity measures is essential.
• Standardized Benchmarks and Protocols: The lack of standardized benchmarks hinders the comparison and advancement of unsupervised methods. The field needs to establish common datasets, preprocessing pipelines, evaluation metrics (both technical and clinical), and reporting standards specifically for unsupervised EEG analysis in mental health (Demirezen et al., 2024; Kinahan et al., 2024; Langer et al., 2022). This would facilitate objective comparison of algorithms and improve research reproducibility.
• Large-Scale, Well-Characterized Datasets: Progress in both supervised and unsupervised learning is often limited by data availability. Collaborative initiatives are needed to create and share large-scale, diverse, and richly annotated EEG datasets for mental health research (Khan et al., 2022; X. Zhang et al., 2022). Addressing limitations in dataset labeling and ensuring diverse representation are crucial for developing generalizable models.
• Hybrid and Semi-Supervised Approaches: Given the challenges of purely unsupervised learning and the limitations of purely supervised methods (especially regarding labeled data scarcity), investigating hybrid or semi-supervised approaches is crucial (Cai et al., 2022; Sosulski et al., 2021). These methods aim to leverage the strengths of both paradigms, using limited labeled data to guide the learning process on larger unlabeled datasets, potentially improving model performance, generalizability, and interpretability.
• Cross-Domain Adaptation: EEG signals vary significantly across subjects, sessions, and hardware. Unsupervised domain adaptation techniques are needed to develop models that can generalize effectively across these variations, enabling broader clinical applicability (Sartipi & Cetin, 2024; C. Xu et al., 2025; T. Xu et al., 2022).
• Explainable AI (XAI) for Unsupervised Models: A major barrier to the clinical translation of complex machine learning models, including unsupervised ones, is their often ‘black-box’ nature. Developing and applying XAI methods tailored for unsupervised EEG analysis is essential to understanding which patterns the models are learning and why they group certain data points together or identify specific anomalies (Apicella et al., 2022; Farahani et al., 2022; Ieracitano et al., 2023; Katmah et al., 2021). This interpretability is crucial for building clinical trust, validating findings, and generating actionable insights.

This theoretical illustration in Figure 4 demonstrates typical rhythmic oscillations that might be analyzed in EEG signals across different frequency bands. This is a conceptual representation and does not depict actual study results.

Figure 5 shows that EEG signal data provides a complete perspective of brain activity by measuring a variety of parameters across multiple channels. Every row denotes a particular observation or incident, and every column denotes a different EEG channel. These channels record electrical activity from certain brain regions, offering important information about mental health conditions. Parameters like Fp1, AF3, F3, F7, FC5, FC1, C3, T7, CP5, CP1, and so forth are among the 33 columns of features that are taken from EEG signals.

5. Conclusions

This systematic review investigated the application of unsupervised learning techniques to EEG data for mental health monitoring between 2018 and 2025. Despite the theoretical advantages of unsupervised methods for exploring complex, unlabeled neurophysiological data, our comprehensive search across six major databases yielded zero eligible peer-reviewed studies meeting the specific inclusion criteria. The results highlight a significant gap in the current literature, with research predominantly focusing on supervised learning paradigms. While supervised methods have shown success in specific classification tasks, their reliance on potentially scarce or subjective labels limits their ability to uncover novel biomarkers or address the inherent heterogeneity of mental health conditions. The absence of empirical studies utilizing unsupervised approaches represents a missed opportunity to leverage these techniques for data-driven discovery, patient stratification, and understanding of the neurobiological underpinnings of mental illness. Future research should prioritize the development and validation of robust unsupervised algorithms for EEG, the creation of standardized benchmarks and large-scale datasets, the exploration of hybrid/semi-supervised methods, and the integration of explainable AI to bridge the gap between potential and practice, ultimately unlocking the full potential of EEG for the advancement of mental health care.