Accessibility settings

This paper is in the following e-collection/theme issue:

Digital Health Reviews (3570) Decision Support for Health Professionals (2170) Insomnia and Sleep Hygiene (307) Biomedical Engineering Articles (81) Machine Learning (3105) Artificial Intelligence (4625)

Published on in Vol 28 (2026)

Preprints (earlier versions) of this paper are available at https://preprints.jmir.org/preprint/93378, first published .
Woman sleeping with CPAP mask for sleep apnea treatment

Accuracy of Machine Learning Algorithms Based on Electroencephalogram in Sleep Apnea Detection: Systematic Review and Meta-Analysis

Accuracy of Machine Learning Algorithms Based on Electroencephalogram in Sleep Apnea Detection: Systematic Review and Meta-Analysis

Authors of this article:

Xiangshuo Li1 Author Orcid Image ;   Lulu Wang1 Author Orcid Image ;   Ting Tang2 Author Orcid Image ;   Yuanyuan Chen2 Author Orcid Image ;   Lan Yang1 Author Orcid Image ;   Hao Cai1 Author Orcid Image ;   Chen Wang1 Author Orcid Image ;   Shuxiao Zhang1 Author Orcid Image ;   Ning Ding1 Author Orcid Image ;   Kouying Liu1 Author Orcid Image
Article Authors Cited by Tweetations Metrics

    1School of Nursing, Nanjing Medical University, 101 Longmian Avenue, Jiangning District, Nanjing, Jiangsu, China

    2Department of Respiratory and Critical Care Medicine, Jiangsu Province Hospital, Nanjing, Jiangsu, China

    Corresponding Author:

    Kouying Liu, PhD


    Background: Sleep apnea (SA) is a serious sleep disorder, and its diagnostic gold standard, polysomnography, is costly and time-consuming. Electroencephalogram (EEG) signals, due to their direct correlation with neural activity and ease of extraction, represent a promising tool. Despite increasing research on machine learning (ML) and deep learning for EEG-based SA detection, model performance has not been consistently evaluated.

    Objective: This systematic review evaluated the accuracy of ML in detecting SA from EEG data and provided an evidence base for further clinical application and future research.

    Methods: Following the PRISMA-DTA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy) and PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 expanded checklists, we systematically searched PubMed, Embase, Web of Science, Cochrane Library (CENTRAL), Scopus, IEEE Xplore, and ClinicalTrials.gov databases from inception to April 2026. Studies evaluating the value of ML algorithms for detecting SA based only on EEG data were included. The Quality Assessment of Diagnostic Accuracy Studies-2 and Prediction Model Risk of Bias Assessment Tool for Artificial Intelligence tools were used to assess the risk of bias in each study. Statistical analysis was performed using the mada and metafor packages in R (version 4.6.0; R Foundation for Statistical Computing) and the Meta-DiSc (version 1.4; Hospital Ramón y Cajal) software. We used GRADE (Grading of Recommendations Assessment, Development and Evaluation) to evaluate the certainty of evidence.

    Results: A total of 27 retrospective studies were included. Segment-level analyses showed high diagnostic performance, with a pooled sensitivity of 0.90 (95% CI 0.85‐0.94; 95% prediction interval 0.43‐0.99) and specificity of 0.92 (95% CI 0.87‐0.95; 95% prediction interval 0.46‐0.99). The pooled area under the summary receiver operating characteristic curve was 0.95 (95% CI 0.92‐0.99). Meta-regression identified EEG channel configuration, region, and validation strategy as significant sources of heterogeneity (P=.004, P=.003, and P=.046, respectively). Multichannel EEG, deep learning approaches, and hold-out validation strategies generally demonstrated better diagnostic performance. Only 2 studies evaluated patient-level diagnostic performance, which was summarized qualitatively.

    Conclusions: To our knowledge, this is the first systematic review and meta-analysis specifically focused on the diagnostic accuracy of EEG-based ML models in the detection of SA. This meta-analysis indicates that ML models based on EEG demonstrate good diagnostic accuracy in detecting SA at the segment level and show promise as tools for SA screening and clinical decision support. However, most current studies are retrospective segment-level analyses, which may overestimate the practical value of this technology in real-world clinical settings. To reliably integrate EEG-based ML models into clinical diagnostic workflows, further prospective studies incorporating full-night monitoring and patient-level validation are needed.

    Trial Registration: PROSPERO CRD420251244156; https://www.crd.york.ac.uk/PROSPERO/view/CRD420251244156

    J Med Internet Res 2026;28:e93378

    doi:10.2196/93378

    Keywords

    artificial intelligence; machine learning; deep learning; sleep apnea; electroencephalogram; diagnosis 


    Sleep apnea (SA) is a serious sleep disorder affecting approximately 936 million people worldwide with moderate to severe cases [1]. Characterized by a reduction in airflow to less than 10% of normal levels during sleep, lasting over 10 seconds, it leads to fragmented sleep and triggers a range of clinical symptoms, including daytime sleepiness, fatigue, and cognitive decline. Recurrent SA events disrupt brain neuroelectric activity patterns, increasing the risk of seizures, stroke, and cardiovascular disease. Specifically, SA was associated with an approximately 2.15-fold higher risk of stroke and a 1.92-fold higher risk of all-cause mortality [2]. According to data from the American Academy of Sleep Medicine, approximately 4% to 19% of adults are affected by SA, while the prevalence can reach as high as 49% among the older population [3]. However, approximately 75% of patients with moderate SA fail to receive timely diagnosis and treatment, imposing a heavy burden on public health [4].

    Currently, polysomnography (PSG) serves as the gold standard for diagnosing SA. This assessment involves overnight recording of multiple physiological signals—including electroencephalogram (EEG), electrocardiogram (ECG), and blood oxygen saturation—in a laboratory setting. However, PSG faces limitations such as high costs, lengthy wait times, and dependence on specialized sleep laboratories, which restrict its widespread application [5]. Additionally, manual interpretation of PSG data is time-consuming, labor-intensive, and prone to subjective bias [6]. Therefore, the pursuit of more convenient and efficient auxiliary diagnostic tools holds significant clinical importance.

    Researchers attempted to detect SA based on a single biological signal (such as ECG, blood oxygen saturation, or respiratory signals), and the results demonstrated its feasibility [7-9]. EEG, as a noninvasive technique, records the brain’s electrical activity by placing electrodes on the scalp surface [10]. Although EEG has limitations such as the application of electrodes, the use of conductive gel, and potential discomfort associated with prolonged monitoring, compared to other physiological signals such as ECGs and respiratory airflow, EEG is less susceptible to interference from endogenous physiological factors like irregular breathing and arrhythmia when used for sleep monitoring. It directly reflects neuronal activity and sleep states in the brain, providing richer and more direct physiological insights for the assessment of SA [11]. During SA events, EEG signals often exhibit changes in spectral and temporal characteristics associated with sleep stage transitions, arousal responses, or microawakenings. By analyzing features across different frequency bands, SA events can be identified [12]. Recent advances in artificial intelligence (AI) have further accelerated the application of EEG in SA detection. Machine learning (ML) methods can identify discriminative patterns from handcrafted EEG features, whereas deep learning (DL) models enable automatic extraction of complex spatiotemporal representations directly from raw EEG signals through end-to-end learning [13]. In recent years, there has been a steady increase in the number of studies on the automatic detection of obstructive sleep apnea (OSA) using electroencephalography. However, existing studies have primarily focused on algorithm development, exhibiting significant variations in research design and experimental settings. Therefore, a systematic review and meta-analysis is warranted to assess the heterogeneity among these studies and provide comprehensive test performance results. Although there have been reviews discussing the application of ML in SA detection, most studies have not specifically focused on EEG signals [14-16]. In a previous review, Fathima and Ahmed [11] provided a qualitative summary of studies on EEG-based SA detection, but did not conduct a quantitative meta-analysis of diagnostic accuracy. Therefore, there remains a lack of systematic quantitative evidence regarding the overall accuracy of EEG-based ML models in SA detection.

    This systematic review aims to enhance understanding of the field by thoroughly analyzing the variations in ML algorithms for detecting SA in EEGs and their impact on diagnostic efficacy, thereby providing a reference for future technological development and clinical translation.


    Study Protocol and Registration

    This systematic review was conducted in accordance with the PRISMA-DTA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy) guidelines, and reporting was supplemented using relevant items from the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 expanded checklist. For details, see Checklists 1 and 2. The protocol was registered on the PROSPERO website with registration number CRD420251244156. Minor modifications to the statistical analysis plan were made during the review process to improve methodological rigor, including the incorporation of multilevel random-effects models, prediction intervals, and the Hartung-Knapp-Sidik-Jonkman method. These changes did not affect the study objectives, eligibility criteria, or primary outcomes.

    Search Strategy

    A comprehensive literature search was developed and conducted in accordance with the PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Literature Search Extension) guideline [17]. The following databases were systematically searched: PubMed or MEDLINE via the National Library of Medicine, Embase via Elsevier, Web of Science via Clarivate Analytics, Cochrane CENTRAL via Wiley, Scopus via Elsevier, and IEEE Xplore via IEEE. ClinicalTrials.gov was additionally searched to identify ongoing or unpublished studies. The initial search was conducted in October 2025 and updated in April 2026 using the same strategy. The search strategy combined MeSH terms and free-text keywords included “artificial intelligence,” “machine learning,” “deep learning,” “algorithm,” “sleep apnea syndromes,” “electroencephalography,” “EEG,” “diagnostic accuracy,” “sensitivity,” and “specificity.” Boolean operators (AND/OR) were applied, and reference lists of included studies were manually screened. The literature search was conducted independently by the author. No published search filters were used. The selection of studies was carried out independently by 2 researchers (XL and LW); any disagreements were resolved through discussion, and when necessary, a third reviewer (KL) was consulted for a final decision. Detailed search strategies are provided in Multimedia Appendix 1.

    Eligibility Criteria

    Studies were eligible for inclusion if they met the following criteria: (1) included adults with SA aged ≥18 years; (2) directly detected SA events using EEG signals alone; (3) explicitly applied either traditional ML or DL algorithms, with specific algorithm types reported; (4) used PSG as the reference standard; and (5) reported sufficient data to directly or indirectly construct 2×2 contingency tables (true positives [TPs], false positives [FPs], true negatives [TNs], and false negatives [FNs]). Studies were excluded if they met any of the following criteria: (1) focused on sleep staging rather than SA detection; (2) did not specifically address SA; (3) were reviews, conference abstracts, case reports, or other nonoriginal research; or (4) lacked sufficient data for analysis or full-text availability.

    Data Collection Process

    The data extraction process uses Microsoft Excel spreadsheets, capturing the following information: study ID, publication year, database source, research objective, patient count, internal or external validation, EEG channel names, feature source, cross-validation method, data type, feature extraction method, training set size, test set size, classifier type, algorithm, sensitivity, specificity, accuracy, F1-score, area under the curve (AUC), TP, TN, FP, and FN. For studies reporting multiple sets of diagnostic performance results, the dataset or model explicitly recommended by the authors and most consistent with the primary objective of the study was preferentially selected for analysis. In cases of missing or unclear data, attempts were made to contact the original authors to obtain the required information. When necessary data were not directly available, 2×2 contingency tables were reconstructed based on reported sensitivity, specificity, and total sample size. Specifically, TPs, FPs, TNs, and FNs were derived using standard formulas for diagnostic test accuracy meta-analysis to enable quantitative synthesis. Data extraction was performed independently by 2 researchers, with discrepancies resolved through discussion or negotiation with a third party.

    Unit of Analysis and Data Partitioning

    Studies are categorized into segment-level and patient-level designs based on the unit of analysis. Segment-level studies analyze EEG data using fixed time windows (eg, 30-second epochs and 10-second frames), treating each segment as an independent unit for feature extraction and classification [18]. Patient-level studies aggregate features recorded throughout the night to generate a single diagnostic result for each patient. To maintain sample independence, studies based on subframe analysis typically use nonoverlapping, fixed-length segmentation. However, SA events may span subframe boundaries; some studies use overlapping sliding windows and perform feature aggregation of subframe features at the feature level to generate a single global feature, thereby preserving event integrity while reducing statistical dependencies between samples [19]. At the data partitioning level, the patient-based training-test set division (where all data from a single participant are fully allocated to either the training set or the test set) effectively prevents data leakage and ensures model generalization. Conversely, if EEG segments from the same participant are distributed across both the training and test sets, this may lead to severe information leakage and an overestimation of performance [20].

    Risk of Bias and Applicability

    This study used the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool [21] and the Prediction Model Risk of Bias Assessment Tool for Artificial Intelligence (PROBAST+AI) tool [22] to conduct quality assessments and risk of bias evaluations for the included studies. QUADAS-2 as a quality assessment tool for diagnostic studies, risk of bias and applicability are evaluated across four dimensions: (1) case selection, (2) studies to be evaluated, (3) gold standard, and (4) process and timing. Each dimension is categorized into 3 levels: low risk, high risk, or unclear risk. Given that the included studies used ML algorithms to diagnose SA based on EEG data, we additionally used an AI prediction model to investigate the PROBAST+AI bias risk assessment tool. This tool is an updated version of the Prediction Model Risk of Bias Assessment Tool-2019, specifically designed to evaluate research on predictive models based on regression modeling or AI methods, distinguishing between the model development and model validation phases. The tool assesses risk of bias and applicability across four domains: (1) participants and data sources, (2) predictor variables, (3) outcomes, and (4) analysis. All assessments were independently conducted by 2 reviewers, and disagreements were resolved through discussion or consultation with a third reviewer.

    Data Synthesis

    Statistical analysis was performed using the mada and metafor packages in R (version 4.6.0; R Foundation for Statistical Computing) and the Meta-DiSc (version 1.4; Hospital Ramón y Cajal) software. Spearman correlation coefficients were calculated to assess the presence of threshold effects. When a threshold effect was identified, a summary receiver operating characteristic curve was constructed, and the AUC was reported. Pooled sensitivity, specificity, likelihood ratios, diagnostic odds ratios, and corresponding 95% CIs were synthesized using random-effects models. To address the nonindependence of effect sizes resulting from the reuse of the same publicly available databases across multiple studies and the inclusion of multiple datasets from a single study, a multilevel random-effects model was applied to account for correlations between effect sizes across and within studies, thereby reducing the risk of type I errors arising from violations of the independence assumption [23]. Between-study heterogeneity was assessed using the Cochran Q test, the I2 statistic, and 95% prediction intervals, with prediction intervals used to reflect the potential range of true effects across different study settings and the practical significance of heterogeneity [24]. Between-study variance (τ2) was estimated using restricted maximum likelihood, and the Hartung-Knapp-Sidik-Jonkman method [25] was applied to obtain more robust 95% CIs. Subgroup analyses and meta-regression analyses were conducted to further explore the potential sources of heterogeneity. The subgroup variables were prespecified based on clinical relevance and methodological considerations, including EEG channel, region, feature extraction method, validation strategy, classifier category, detection task, and dataset source. Sensitivity analyses were conducted to evaluate the robustness of the pooled results. Publication bias was assessed using the Deeks funnel plot asymmetry test. In addition, the Fagan nomogram was used to evaluate the clinical utility of EEG-based ML models for SA detection.

    Certainty Assessment

    The certainty of evidence was assessed using the GRADE (Grading of Recommendations Assessment, Development and Evaluation) framework. The GRADEpro Guideline Development Tool was used to facilitate the assessment. Five domains were considered, including risk of bias, inconsistency, indirectness, imprecision, and publication bias. Certainty was rated as high, moderate, low, or very low according to standard GRADE criteria. Assessments were performed independently by 2 reviewers, with disagreements resolved through discussion and consultation with a third reviewer when necessary.


    Study Selection

    A total of 1914 literature records were identified. After deduplication using EndNote (Clarivate Analytics) software, 578 duplicates were removed. Preliminary screening based on titles and abstracts excluded 1179 studies, leaving 157 studies for full-text eligibility assessment. Following full-text review, 130 studies were excluded, resulting in 27 studies ultimately included in the meta-analysis. The study selection process is illustrated in the PRISMA flow diagram (Figure 1).

    ‎
    Figure 1. PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flow diagram illustrating the study selection process for a systematic review and meta-analysis of diagnostic test accuracy studies evaluating machine learning algorithms based on EEG signals for the detection of SA. The literature search included multiple electronic databases from inception to April 2026, and records were screened according to predefined eligibility criteria. EEG: electroencephalogram; SA: sleep apnea.

    Study Characteristics

    The main characteristics of the included studies are summarized in Tables 1 and 2. The studies were published between 2006 and 2026 and were conducted across Asian and Western countries. All 27 studies used retrospective designs. Among them, 3 studies [26-28] performed external validation, while the remaining 24 studies [19,29-51] relied on internal validation. Publicly available datasets were the primary data source, accounting for 80.8% (21/26) of the included studies [19,26-29,31,34-38,41-51]. The most frequently used databases were the MIT-BIH Sleep Database (11/21) and the University College Dublin Sleep Apnea Database (6/21). Notably, the MIT-BIH Sleep Database included exclusively male participants. In addition, 5 studies used self-collected clinical datasets [26,32,33,39,40], and 23 studies reported the specific EEG channels used, with the central region channels being the most frequently used. The proportions of C3-A2 and C4-A1 channels were 47.8% (11/23) [26,28,30,33,34,37-40,45,50] and 56.5% (13/23), respectively [27-29,33,37-41,44,47,49,50]. The included studies applied both ML and DL approaches. Different studies used diverse signal processing and feature extraction techniques, including discrete wavelet transform, variational mode decomposition, fast Fourier transform, Hilbert-Huang transform, and wavelet packet decomposition. These were typically combined to construct multidimensional feature sets incorporating time-domain, frequency-domain, and nonlinear features for model training. Among model evaluation strategies, cross-validation is widely used to obtain more stable assessments of model performance. Specifically, 13 studies used k-fold cross-validation [19,26,29,32,37-40,43,45,49-51], 2 studies used leave-one-out cross-validation [37,47], and 4 studies used the hold-out method to split the data into training and test sets [30,31,33,36]. Detailed information is available in Multimedia Appendix 2 [19,26-51].

    Table 1. Characteristics of the studies included in this systematic review and meta-analysis of diagnostic test accuracy evaluating machine learning algorithms based on electroencephalogram (EEG) signals for the detection of sleep apneaa.
    StudyRegionTargetDatabaseModel typeAlgorithmsEEGFeature extractionValidation strategy
    Cheng et al [26]bAsian countriesDetect OSAcUCDDBd, ISRUCe, Local hospitalDLfEEG-MILgC3-A2AutomaticCross-validation
    Barnes et al [27]bWestern countriesDetect SAhSHHSiDLCNNjC4-A1AutomaticCross-validation
    Mahmud et al [28]bAsian countriesDetect SAUCDDBDLCNNC3-A2, C4-A1AutomaticIndependent dataset validation
    Jiang et al [29]bAsian countriesDetect SAMIT-BIHkDLMSPCNNlO2-A1, C4-A1, C3-O1AutomaticCross-validation
    Emin Tagluk and Sezgin [30]bAsian countriesDetect OSA—mMLnANNoC3-A2Advanced nonlinearHold-out validation
    Lin et al [31]bAsian countriesDetect SAMIT-BIHMLANNC3-O1Traditional handcraftedHold-out validation
    Zhang et al [32]pAsian countriesScreen severe OSAThe Seventh Affiliated Hospital of Sun Yat-sen UniversityDLGCNqF3, F4, C3, C4, O1, O2Advanced nonlinearCross-validation
    Wang et al [33]bAsian countriesDetect SATianjin Chest HospitalMLRFrC3-A2, C4-A1Traditional handcraftedHold-out validation
    Prucnal and Polak [34]bWestern countriesDetect OSA or CSAsUCDDBDLFFNNtC3-A2Traditional handcrafted—
    Delimayanti et al [35]bAsian countriesDetect SACAP Sleep databaseuDLCNNFp1-F3, F3-C3, C3-P3, P3-O1 and/or Fp2-F4, F4-C4, C4-P4, P4-O2Automatic—
    Zhou et al [36]bAsian countriesDetect SAMIT-BIHMLSVMvC3-O1Advanced nonlinearHold-out validation
    Saha et al [37]bAsian countriesDetect SAUCDDBMLKNNwC3-A2, C4-A1Traditional handcraftedCross-validation
    (LOOCV)x
    Gupta et al [38]bAsian countriesDetect SAMIT-BIHMLEnsemble Bagged TreesC3-A2, C4-A1Traditional handcraftedCross-validation
    Wang et al [39]bAsian countriesDetect SATianjin Chest HospitalDLBI-LSTMyC3-A2, C4-A1AutomaticCross-validation
    Zhao et al [40]bAsian countriesDetect OSA
    or CSA    		Tianjin Chest HospitalMLRFC3-A2, C4-A1Traditional handcraftedCross-validation
    Bonner et al [41]bWestern countriesDetect SAMIT-BIHDLTCNNzC3-O1, C4-A1, O2-A1——
    Gurrala et al [42]bAsian countriesDetect SAMIT-BIHMLEnsemble Bagged Tree—Traditional handcrafted—
    Taran et al [43]bAsian countriesDetect SAMIT-BIHMLKNN—AutomaticCross-validation
    Khan et al [44]pAsian countriesDetect SASHHSMLSVMC4-A1Traditional handcrafted—
    Prucnal and Polak [45]bWestern countriesDetect OSA
    or CSA    		UCDDBMLSVMC3-A2Advanced nonlinearCross-validation
    Wijaya et al [46]bAsian countriesDetect OSA
    or CSA    		MGH 2018aaDLCNN-GRUabC3-M2, C4-M1Automatic—
    Bhalerao and Pachori [47]bAsian countriesDetect SAMIT-BIHDLCNNC3-O1, C4-A1, O2-A1Traditional handcraftedCross-validation (LOOCV)
    Shahnaz et al [19]bAsian countriesDetect SAMIT-BIHMLSVM—Traditional handcraftedCross-validation
    Taran et al [48]bAsian countriesDetect SAMIT-BIHMLLS-SVMac—Advanced nonlinear—
    Sharifi and Fakharzadeh [49]bAsian countriesDetect SAMIT-BIHMLRFC4-A1Traditional handcraftedCross-validation
    Saha et al [50]bAsian countriesDetect SAUCDDBMLKNNC3-A2, O2-A1, C4-A1, and C3- O1Traditional handcraftedCross-validation
    Band and Deshmukh [51]bAsian countriesDetect SASleep EDF DatasetadDL1D-CNNae(FpzCz)Traditional handcraftedCross-validation

    aThe table summarizes key study characteristics, including study design, geographic region, target population, data sources, model types, machine learning algorithms, EEG channel information, feature extraction methods, and validation strategies.

    bEvent-level studies.

    cOSA: obstructive sleep apnea.

    dUCDDB: University College Dublin Sleep Apnea Database.

    eISRUC: Institute of Systems and Robotics, University of Coimbra Sleep Dataset.

    fDL: deep learning.

    gEEG-MIL: EEG multi-instance learning network.

    hSA: sleep apnea.

    iSHHS: Sleep Heart Health Study.

    jCNN: convolutional neural network.

    kMIT-BIH: MIT-BIH Polysomnographic Database.

    lMSPCNN: multiscale parallel convolutional neural network.

    mNot applicable.

    nML: machine learning.

    oANN: artificial neural network.

    pPatient-level study.

    qGCN: graph convolutional network.

    rRF: random forest.

    sCSA: central sleep apnea.

    tFFNN: feed-forward neural network.

    uCAP Sleep database: Cyclic Alternating Pattern Sleep Database.

    vSVM: support vector machine.

    wKNN: k-nearest neighbor.

    xLOOCV: leave-one-out cross-validation.

    yBI-LSTM: bidirectional long short-term memory.

    zTCNN: temporal convolutional neural network.

    aaMGH 2018: a publicly available dataset derived from polysomnographic recordings collected at Massachusetts General Hospital and released through PhysioNet.

    abGRU: gated recurrent unit.

    acLS-SVM: least squares support vector machine.

    adSleep EDF Dataset: Sleep European Data Format Database.

    ae1D-CNN: one-dimensional convolutional neural network.

    Table 2. Data extracted from the included studiesa.
    StudySample size (subjects/segments), nSensitivity (%)Specificity (%)Accuracy (%)
    Cheng et al [26]25/987780.3567.2170.76
    Cheng et al [26]61/23,72876.4162.564.9
    Cheng et al [26]35/13,17182.1970.3074.10
    Barnes et al [27]2691/30,08781.6267.5669.92
    Mahmud et al [28]12/287,43790.0682.9486.38
    Jiang et al [29]16/264093.0883.9089.09
    Emin Tagluk and Sezgin [30]20/470094.1398.1796.15
    Lin et al [31]N/Ab/28369.6444.4454.42
    Zhang et al [32]88/N/A80.7783.8782.95
    Wang et al [33]30/40693.1095.0794.33
    Prucnal and Polak [34]N/A/19886.3683.3388.76
    Prucnal and Polak [34]N/A/19874.2487.8889.63
    Delimayanti et al [35]5/12100.0083.3092.00
    Zhou et al [36]12/14493.298.6095.10
    Saha et al [37]5/170689.6891.7990.74
    Gupta et al [38]5/170693.2097.2095.10
    Wang et al [39]N/A/139088.4690.0789.14
    Zhao et al [40]30/34790.2487.9584.34
    Zhao et al [40]30/34768.2996.2383.33
    Bonner et al [41]15/121955.4691.2184.50
    Gurrala et al [42]18/940194.1998.7397.69
    Taran et al [43]N/A/214295.6896.2296.00
    Khan et al [44]547/N/A56.8263.4160.15
    Prucnal and Polak [45]25/411962.5281.1074.92
    Prucnal and Polak [45]25/411962.2081.2574.90
    Wijaya et al [46]N/A/516998.9599.7999.52
    Wijaya et al [46]N/A/516999.3699.6299.54
    Bhalerao and Pachori [47]14/799897.5896.7296.74
    Shahnaz et al [19]14/272088.5285.1486.84
    Taran et al [48]16/2124100.0091.9996.46
    Sharifi and Fakharzadeh [49]N/A/364085.1989.1687.18
    Saha et al [50]16/300080.9780.1380.55
    Saha et al [50]25/470091.5785.1988.38
    Band and Deshmukh [51]N/A/40092.5086.0089.25

    aThe table presents the included studies, sample sizes (subjects/segments), and reported diagnostic performance metrics, including sensitivity, specificity, and accuracy. Multiple rows within the same study represent different datasets or distinct diagnostic objectives reported in the original paper.

    bN/A: not applicable.

    Risk of Bias and Applicability

    Quality assessments of included studies were conducted using the QUADAS-2 and PROBAST+AI tools (Figures 2 and 3). The QUADAS-2 assessment results showed that, in the area of patient selection, 2 (7.4%) studies were judged to be at high risk of bias due to inappropriate subject selection and artificial balancing of groups. Another 10 (37.0%) studies did not adequately describe the subject selection process, and their risk of bias was judged to be unclear; the remaining 15 (55.6%) studies were classified as having a low risk of bias. In the areas of flow and timing, the interval between the test under evaluation and the reference standard was unclear in 5 (18.5%) studies, and the risk of bias was judged to be unclear. All other areas were classified as having a low risk of bias. According to the PROBAST+AI assessment criteria, during the model development phase, 8 (29.6%) studies had a high risk of bias, and 1 (3.7%) study had significant concerns regarding clinical applicability. During the model validation phase, 6 (22.2%) studies were assessed as having a high risk of bias, and 1 (3.7%) study had significant issues regarding clinical applicability. The analysis domain was the primary source of overall bias risk; most studies did not address potential overfitting issues or did not specify how missing data were handled. Overall, the majority of the included studies had a low to moderate risk of bias.

    ‎
    Figure 2. Risk of bias and applicability assessment of included studies using the QUADAS-2 tool in this systematic review and meta-analysis of diagnostic test accuracy evaluating machine learning algorithms based on electroencephalogram signals for sleep apnea detection. The figure summarizes domain-level judgments of risk of bias and applicability concerns. Green, yellow, and red indicate low, unclear, and high risk of bias, respectively. QUADAS-2: Quality Assessment of Diagnostic Accuracy Studies-2.
    ‎
    Figure 3. Methodological quality and risk of bias assessment of included studies using the PROBAST+AI tool in this systematic review and meta-analysis of diagnostic test accuracy evaluating machine learning models based on electroencephalogram signals for sleep apnea detection. (A) The assessment results for model development studies. (B) Results for external validation studies. Green, yellow, and red indicate low, unclear, and high risk of bias, respectively. PROBAST+AI: Prediction Model Risk of Bias Assessment Tool for Artificial Intelligence.

    Synthesis of Results

    A segment-level meta-analysis included 27 studies comprising 32 sets of 4-cell table data. Spearman correlation analysis demonstrated a significant correlation between logit-transformed sensitivity and logit(1−specificity) (r=−0.685; P<.001), indicating a pronounced threshold effect among the included studies and suggesting that between-study heterogeneity was primarily attributable to differences in classification thresholds. Therefore, a bivariate random-effects model was used to construct the summary receiver operating characteristic curve. The pooled AUC was 0.95 (95% CI 0.92‐0.99; Figure 4), showing a promising overall diagnostic performance of EEG-based ML models for SA detection. The pooled sensitivity was 0.90 (95% CI 0.85‐0.94; I2=99.5%; Figure 5), and the pooled specificity was 0.92 (95% CI 0.87‐0.95; I2=99.8%; Figure 6). The diagnostic odds ratio was 88.58 (95% CI 35.76‐219.41). Although the pooled estimates suggested high diagnostic accuracy for both positive and negative SA cases, substantial between-study heterogeneity was observed, indicating considerable variability in model performance across different study settings. Furthermore, the 95% prediction intervals were wide for both sensitivity (0.43‐0.99) and specificity (0.46‐0.99), suggesting that diagnostic performance may vary substantially in external validation scenarios or future real-world applications, thereby reflecting limited model stability and generalizability. To further evaluate potential clinical utility, Fagan nomograms were constructed. When the pretest probability was set at 50%, the positive likelihood ratio was 10.85 (95% CI 6.39‐17.83), corresponding to a posttest probability of 91%, whereas the negative likelihood ratio was 0.11 (95% CI 0.07‐0.17), corresponding to a posttest probability of 11% (Figure 7). However, the pooled likelihood ratio scatterplot showed that most studies were located in the lower-right quadrant (positive likelihood ratio: <10 and negative likelihood ratio: >0.1), indicating limited ability of these models to independently confirm or exclude SA (Figure 8). In addition, substantial dispersion across studies further suggested considerable variability in diagnostic performance. Sensitivity analyses demonstrated that sequential exclusion of individual studies resulted in changes of less than 5 percentage points in pooled sensitivity and specificity estimates, indicating good robustness and stability of the pooled findings (Multimedia Appendix 3) [19,26-31,33-43,45-51].

    Among patient-level SA detection studies, only 2 studies met the inclusion criteria. Due to the limited number of eligible studies, reliable estimation of pooled effect sizes was not feasible, precluding formal meta-analysis. Descriptive analysis demonstrated substantial variability in diagnostic performance, with sensitivity ranging from 56.82% to 80.77% and specificity ranging from 63.41% to 83.87%.

    ‎
    Figure 4. The SROC graph for the studies. The AUC of electroencephalogram-based machine learning algorithms for detecting sleep apnea was 0.95 (95% CI 0.92‐0.99). AUC: area under the curve; FPR: false positive rate; Se: sensitivity; Sp: specificity; SROC: summary receiver operating characteristic; TPR: true positive rate.
    ‎
    Figure 5. Forest plot of pooled sensitivity for segment-level machine learning and deep learning models. The pooled sensitivity was 0.90 (95% CI 0.85‐0.94), with a 95% prediction interval of 0.43‐0.99 [19,26-31,33-43,45-51]. HKSJ: Hartung-Knapp-Sidik-Jonkman.
    ‎
    Figure 6. Forest plot of pooled specificity for segment-level machine learning and deep learning models. The pooled specificity was 0.92 (95% CI 0.87‐0.95), with a 95% prediction interval of 0.46‐0.99 [19,26-31,33-43,45-51]. HKSJ: Hartung-Knapp-Sidik-Jonkman.
    ‎
    Figure 7. Fagan nomogram of machine learning or deep learning models for the diagnosis of sleep apnea. The first column of the nomogram represents pretest probability, the second column represents likelihood ratio, and the third column shows posttest probability. LR+: positive likelihood ratio; LR−: negative likelihood ratio.
    ‎
    Figure 8. Likelihood ratio scatter plot of electroencephalogram-based machine learning or deep learning diagnostic models. The summary point for machine learning or deep learning models is in the lower-right quadrant (LR+<10 and LR−>0.1; cannot rule out or confirm sleep apnea). LLQ: lower left quadrant; LR+: positive likelihood ratio; LR−: negative likelihood ratio; LUQ: left upper quadrant; RLQ: right lower quadrant; RUQ: right upper quadrant.

    Additional Analysis

    At the segment level, this study investigated potential sources of heterogeneity across studies through subgroup analysis and meta-regression. The results of the subgroup analysis showed that only the pooled sensitivity differences among the region subgroups were statistically significant (P=.01), while the differences in pooled sensitivity and specificity for the remaining subgroups were not statistically significant (sensitivity: P=.94, P=.70, P=.11, P=.83, and P=.88, respectively; specificity: P=.46, P=.71, P=.12, P=.23, and P=.35, respectively). Furthermore, this study originally planned to further evaluate the impact of different detection tasks on diagnostic performance; however, due to the limited number of studies in some subgroups and unstable model convergence, reliable subgroup estimates could not be obtained, and therefore, the results of this analysis are not reported. Detailed subgroup analysis results are presented in Tables 3 and 4. To further identify sources of heterogeneity, this study conducted a meta-regression analysis using prespecified covariates. The results showed that the number of EEG channels, region, and validation method were all significantly associated with pooled diagnostic performance (P=.004, P=.003, and P=.046, respectively), suggesting that differences in EEG channel configuration, region, and validation strategies across studies may represent important sources of heterogeneity and may partially explain the variability among study results. Detailed results are provided in Multimedia Appendix 3.

    Table 3. The diagnostic performance of electroencephalogram (EEG)-based machine learning (ML) models across different subgroups, including sensitivity and 95% CIa.
    CategoryStudies, nSensitivity (%)95% CI (%)I2 (%)τ2QbTest for subgroup differences, P value
    Region.01
    Asian260.920.88‐0.9598.91.23622224.83
    Western60.710.52‐0.8599.20.3792659.56
    Database.94
    Publicly available260.900.83‐0.9499.51.69815489.33
    Self-collected50.860.72‐0.9391.70.427548.42
    Feature extraction.70
    Automatic learning120.900.78‐0.9699.71.67163577.17
    Traditional handcrafted150.880.81‐0.9398.10.8644719.98
    EEG.11
    Single channel210.870.79‐0.9398.91.37081770.51
    Multichannel100.940.86‐0.9898.01.3645443.04
    Validation method.83
    Cross-validation190.890.83‐0.9299.00.69681747.14
    Hold-out validation40.880.41‐0.9999.02.1475290.33
    Algorithm.88
    ML170.900.82‐0.9499.01.35371614.25
    DLc150.900.80‐0.9699.61.56163840.28

    aSubgroups encompass region, diverse data sources, feature extraction, EEG channel configurations, validation methods, and algorithm types. P values indicate the statistical significance of differences in sensitivity between subgroups. Subgroup counts represent independent effect sizes rather than the number of included studies.

    bP value<.001.

    cDL: deep learning.

    Table 4. The diagnostic performance of electroencephalogram (EEG)-based machine learning (ML) models across different subgroups, including specificity and 95% CIa.
    CategoryStudies, nSpecificity (%)95% CI (%)I2 (%)τ2QbTest for subgroup differences, P value
    Region.15
     Asian260.930.88‐0.9699.71.62368195.2
     Western60.830.67‐0.9299.30.4462674.15
    Database.46
     Publicly available260.910.85‐0.9599.81.675815,506.47
     Self-collected50.900.74‐0.9798.10.8351210.64
    Feature extraction .71
     Automatic learning120.900.74‐0.9799.92.459712,331.54
     Traditional handcrafted150.920.86‐0.9598.50.9649964.99
    EEG.12
     Single channel210.900.83‐0.9499.51.23994211.78
     Multichannel100.920.87‐0.9999.01.9297938.56
    Validation method.23
     Cross-validation190.890.83‐0.9399.50.78813697.45
     Hold-out validation40.950.55‐1.0098.12.6500154.79
    Algorithm.35
     ML170.930.88‐0.9698.71.10671218.63
     DLc150.900.77‐0.9699.92.050312,901

    aSubgroups encompass region, diverse data sources, feature extraction, EEG channel configurations, validation methods, and algorithm types. P values indicate the statistical significance of differences in sensitivity between subgroups.

    bP value<.001.

    cDL: deep learning.

    Publication Bias

    Publication bias was primarily assessed using the Deeks funnel plot asymmetry test, which is the recommended method for meta-analyses of diagnostic test accuracy [52]. The results indicate that studies included based on segment-level hierarchy were distributed on both sides of the regression line (P=.07), suggesting no evidence of significant small-study effects (Figure 9).

    ‎
    Figure 9. Deeks funnel plot asymmetry test for publication bias in this systematic review and meta-analysis of diagnostic test accuracy evaluating electroencephalogram-based machine learning algorithms for sleep apnea detection. The P value of the asymmetry test was .07 (P>.05), suggesting no evidence of significant small-study effects. ESS: effective sample size.

    Certainty of Evidence

    According to the GRADE framework, the quality of evidence for both the sensitivity and specificity of this measure was rated as low. However, because several included studies were at high risk of bias and substantial residual heterogeneity remained unexplained after subgroup analyses and meta-regression, the evidence grade was downgraded by 2 levels (Table 5).

    Table 5. GRADEa assessment of the certainty of evidence for the diagnostic accuracy of electroencephalogram-based machine learning models in detecting sleep apneab.
    OutcomeStudies, nPatients, nStudy designFactors that may decrease certainty of evidenceEffect per 1000 patients testedTest accuracy CoEc
    Risk of biasIndirectnessInconsistencyImprecisionPublication biasPretest probability of 15%, n (95% CI)Pretest probability of 30%, n (95% CI)Pretest probability of 50%, n (95% CI)
    True positives27705,115Cross-sectional (cohort-type accuracy study)SeriousdNot seriousSeriouseNot seriousNone135 (128-141)270 (255-282)450 (425-470)⨁⨁◯◯ Low
    False negatives27705,115Cross-sectional (cohort-type accuracy study)SeriousdNot seriousSeriouseNot seriousNone15 (9-22)30 (18-45)50 (30-75)⨁⨁◯◯ Low
    True negatives27705,115Cross-sectional (cohort-type accuracy study)SeriousdNot seriousSeriouseNot seriousNone782 (739-808)644 (609-665)460 (435-475)⨁⨁◯◯ Low
    False positives27705,115Cross-sectional (cohort-type accuracy study)SeriousdNot seriousSeriouseNot seriousNone68 (42-111)56 (35-91)40 (25-65)⨁⨁◯◯ Low

    aGRADE: Grading of Recommendations Assessment, Development and Evaluation.

    bA comprehensive assessment is conducted based on 5 factors: risk of bias, inconsistency, indirectness, imprecision, and publication bias. The quality of evidence is classified into 4 levels: high (⊕⊕⊕⊕), moderate (⊕⊕⊕◯), low (⊕⊕◯◯) , and very low (⊕◯◯◯). Sensitivity=0.90 (95% CI 0.85-0.94); specificity=0.92 (95% CI 0.87-0.95).

    cCoE: certainty of evidence.

    dAmong the included studies, 2 had a high risk of bias regarding patient selection, 10 had an unclear risk of bias, and 5 had an unclear risk of bias regarding flow and timing.

    eThe 95% prediction intervals were significantly wider than the corresponding 95% CIs, and there was unexplained heterogeneity.


    Summary of Evidence

    To the best of our knowledge, this study is the first to quantitatively synthesize the diagnostic performance of EEG-based ML models for SA detection via meta-analysis, while previous reviews only provided qualitative summaries [11]. Segment-level analyses demonstrated promising diagnostic performance, with a pooled sensitivity of 0.90 (95% CI 0.85‐0.94; I2=99.5%) and specificity of 0.92 (95% CI 0.87‐0.95; I2=99.8%), representing the average effect levels across all included studies, while the AUC was 0.95 (95% CI 0.92‐0.99). However, wide 95% prediction intervals for sensitivity (0.43‐0.99) and specificity (0.46‐0.99) suggest that the true effect may vary considerably across different research settings, and caution is warranted when extrapolating these findings to specific clinical contexts. Meta-regression analysis showed that EEG electrode configuration was significantly associated with both pooled sensitivity and specificity, whereas regional factors and validation strategies were associated with sensitivity and specificity, respectively. These findings suggest that differences in electrode configuration, regional background, and validation methods across studies may represent potential sources of heterogeneity. Subgroup analysis further showed that the multilead EEG model outperformed the single-lead model (sensitivity: 0.94 vs 0.87; specificity: 0.92 vs 0.90). However, the between-group differences did not reach statistical significance, and these findings should therefore be interpreted with caution. Studies conducted in Asian populations showed slightly better diagnostic performance than those in Western populations, and hold-out validation appeared to outperform cross-validation; however, these findings may be influenced by small subgroup sample sizes. However, the risk of bias in the included studies was low to moderate, and following a GRADE assessment, the overall certainty of the evidence was judged to be low, primarily due to high heterogeneity and risk of bias in some studies. Therefore, although the pooled estimates suggest promising diagnostic performance, confidence in the magnitude and generalizability of these estimates remains limited.

    In exploring heterogeneity sources, the threshold effect was identified as a significant contributor, potentially attributable to differences in diagnostic task definitions and criteria [53]. Several studies did not stratify patients by disease severity or treat multiple condition categories as a single group [26]. In addition, most studies do not report specific thresholds or the rationale behind them, which limits further subgroup analysis. Subgroup analysis revealed that models using public datasets and automatic feature learning showed better performance than those using institution-specific datasets and manual feature extraction, respectively, though none of these differences reached statistical significance (P=.09 and P=.07, respectively), and the relevant conclusions should be interpreted with caution. Nevertheless, these trends suggest that the standardization of datasets and feature extraction strategies may influence model performance and the reproducibility of results. In addition, previous studies comparing portable monitoring systems with standard PSG have reported inconsistencies in signal synchronization and respiratory event recording [54]. This issue may stem from a lack of synchronization between the recording times of portable monitoring devices and PSG. This discrepancy can lead to inaccuracies in event annotation, which in turn can interfere with model training and result in biased performance evaluations [55]. In addition, sleep stage information has a significant impact on the performance of EEG-based SA detection models. The combined sleep stage and OSA screening model developed by Kang et al [56] found that in younger individuals, EEG features (K-complexes, spindles, and beta-band activity) differed significantly between rapid eye movement and nonrapid eye movement sleep, whereas in older adults, these differences were only observed during nonrapid eye movement sleep. This suggests that the interaction between sleep stages and EEG features varies by age, and models that fail to account for sleep stage information may exhibit inconsistent diagnostic performance across populations.

    Multichannel EEG significantly outperformed single-channel EEG in both sensitivity and specificity. This aligns with the principle that multichannel configurations provide more comprehensive spatial information, capturing distributed brain network activity patterns associated with SA [57]. This finding aligns with the conclusions of Prucnal and Polak’s research [58]. Multichannel setups can partially offset noise interference or localized abnormal signals by integrating complementary signals from different functional brain regions [59]. Zhang et al [60] systematically investigated the performance differences between single-channel and multichannel data in automated sleep staging. Results indicated that as the complexity of classification tasks increased, the performance advantage of multichannel data became increasingly pronounced, further supporting the rationale for using multichannel EEG in complex sleep-related recognition tasks. However, constrained by factors such as channel configurations in public datasets, acquisition costs, and signal quality control, most studies favor single-channel EEG [61]. This tendency poses challenges for the practical application and widespread adoption of multichannel EEG. Consequently, channel selection strategies are crucial for balancing diagnostic performance with practical feasibility. Regarding channel selection, central channels—particularly C3-A2 and C4-A1—have been validated as effective choices due to their proximity to the sensorimotor cortex, where signals are closely associated with respiratory effort and microarousal [39,40,62]. Meanwhile, occipital-parietal pathways such as O1-A2 may more strongly reflect visual cortex activity and have also been reported as high-precision pathways [29,41]. Therefore, selectively incorporating occipital channels may provide supplementary information about sleep microstructure, particularly in distinguishing respiratory events associated with rapid eye movement.

    PSG, as the gold standard for diagnosing SA, has reported sensitivity and specificity ranging from 80% to 97% and 85% to 97%, respectively [63]. However, its complex operation and requirement for specialized sleep laboratories limit widespread application [5]. Recent advances in ML offer a promising alternative solution. Gurrala et al [42] focused on EEG-based feature extraction and proposed a low-complexity method that outperformed earlier approaches using multimodal signals such as ECG, blood pressure, and respiratory data. These results support the feasibility of EEG-only ML approaches for SA detection. From a physiological perspective, EEG alterations triggered by respiratory events manifest as transient changes in spectral and temporal characteristics [64]. Segment-level analysis avoids information dilution caused by long-duration window averaging [65], supporting its use as a physiologically plausible approach.

    Despite the favorable segment-level performance observed in this meta-analysis, translating segment-level prediction into clinically meaningful patient-level diagnosis remains a major challenge for EEG-based ML models. In clinical practice, SA severity is primarily determined using the apnea-hypopnea index (AHI), which is calculated based on the total number of apnea and hypopnea events per hour of sleep [66]. Therefore, accurate patient-level diagnosis requires not only reliable detection of respiratory events throughout the entire sleep recording but also accurate estimation of total sleep time and aggregation of detected events into clinically interpretable AHI values. However, many included studies focused primarily on segment-level classification performance and did not evaluate whether segment-level predictions could be reliably translated into patient-level AHI estimation. Furthermore, the included EEG-based studies primarily focus on apnea detection and do not clearly distinguish hypopnea events, which may further limit clinical applicability.

    Although the overall performance of EEG-based ML detection models is satisfactory, pooled likelihood ratio maps suggest that they are insufficient for diagnosing or ruling out SA events on their own. Therefore, these models may be more suitable as adjunctive diagnostic or screening tools rather than standalone methods. Nevertheless, EEG-based AI systems demonstrate substantial potential for community-based and home-based sleep monitoring. Existing research indicates that continuous 5‐7 day community or home EEG monitoring can capture nocturnal variability in sleep patterns [67], avoid single-night PSG misdiagnosis, and enable early screening. Seol et al [68] validated a portable EEG device against PSG in 77 patients with OSA, reporting AUC values of 0.897 and 0.968 for EEG-based arousal index when screening for severe OSA (AHI ≥15 and ≥30, respectively). This indicates that consumer-grade EEG headbands can serve as effective preliminary screening tools in community settings, especially with automated analysis [69]. However, community health care facilities may face limitations due to equipment costs and technical personnel requirements [70]. Developing low-cost, portable simplified EEG devices combined with automated ML algorithms could overcome this bottleneck. Delimayanti et al [35] found that convolutional neural network models trained on raw EEG data outperformed those using fast Fourier transform–processed data, demonstrating the potential of convolutional neural networks for developing low-cost, accessible detection tools. Beyond improving diagnostic accessibility, AI-assisted EEG systems may contribute to integrated digital sleep health management pathways. Combining portable EEG devices with cloud-based AI analysis, remote telemonitoring, and electronic health record systems could establish proactive screening and longitudinal monitoring frameworks, facilitating early identification and timely referral of high-risk individuals [71]. Additionally, remote real-time EEG monitoring has been reported to reduce health care costs for brain disorders [72], though the cost-effectiveness of AI-assisted SA screening still requires further investigation.

    Future research should prioritize establishing a standardized framework for model evaluation, threshold selection, and result reporting to improve consistency and comparability across studies. While leveraging the diagnostic advantages of multichannel EEG, future studies should also develop channel configuration strategies tailored to different clinical and real-world scenarios. In addition, AI-assisted EEG screening systems require further prospective validation in real-world settings, with emphasis on clinical utility, feasibility, and cost-effectiveness. Novel collaborative frameworks such as federated learning may facilitate multicenter data sharing and model training while preserving patient privacy, thereby improving model stability and generalizability across regions, populations, and health care systems[73]. In parallel, explainable AI methods (eg, Shapley Additive Explanations, Local Interpretable Model-Agnostic Explanations, and Gradient-Weighted Class Activation Mapping) may help address the “ black-box” limitation of DL models, improving interpretability and clinician trust in AI-assisted decision-making [74]. Finally, continued external validation, bias monitoring, and dynamic model updating remain essential to ensure fairness, stability, and long-term reliability in clinical practice [75].

    Limitations

    This systematic review and meta-analysis has several limitations. First, during the data extraction phase, we recoded multiclass outcome measures into binary categories, which prevented further subgroup analyses based on diagnostic task types and limited the assessment of heterogeneity across different task scenarios. Second, quantitative synthesis could not be performed for some studies due to nonstandard reporting or missing key data. For studies with incomplete information, we reconstructed 2×2 contingency tables using statistical inference; this approach may have slightly affected the precision of the pooled effect sizes. We recommend that future studies present raw data completely and in a standardized manner to support higher-quality meta-analyses. Third, the majority of included studies used retrospective designs, and data sources were limited to public databases, which may introduce selection bias; consequently, study conclusions are difficult to generalize to broader clinical populations. Furthermore, and most critically, the vast majority of included studies did not undergo external validation. Most models were evaluated using only internal datasets or single-center data without independent external testing. This may lead to an overoptimistic assessment of the models’ diagnostic performance and significantly reduce their generalizability and practical value in real-world clinical settings. Therefore, future studies should prioritize rigorous external validation across multicenter settings and diverse populations to ensure the models’ stability and clinical translational value.

    Conclusions

    Although EEG-based ML models demonstrate high diagnostic performance at the segment level, their ability to detect short apnea events does not necessarily reflect their performance in assessing overall disease severity (eg, AHI) during overnight monitoring. As a result, event-level metrics reported in current studies may overestimate their clinical screening and diagnostic value in real-world settings. Therefore, future research should prioritize overnight or patient-level external validation to better evaluate clinical applicability and generalizability. With advances in wearable EEG technology, EEG-based approaches remain promising as convenient and cost-effective tools for home sleep monitoring.

    Acknowledgments

    The authors would like to thank all members of the research team for their valuable contributions to this study. The authors are particularly grateful to Jinyu Sun and Jin Liu for their guidance and support in statistical analysis. The authors also sincerely thank Kouying Liu and Ning Ding for their overall supervision of the study and their critical review of the manuscript, which greatly contributed to improving its scientific quality and rigor. The authors declare the use of generative artificial intelligence (GAI) in the research and writing process. According to the GAIDeT taxonomy (2025), the following tasks were delegated to GAI tools under full human supervision: proofreading and editing. The GAI tool used was ChatGPT-5.5. Responsibility for the final manuscript lies entirely with the authors. GAI tools are not listed as authors and do not bear responsibility for the final outcomes. The original ChatGPT conversation logs can be found in Multimedia Appendix 4.

    Funding

    This study was supported by the National Natural Science Foundation of China (grant 82570129) and Project of “Nursing Science” Funded by the 4th Priority Discipline Development Program of Jiangsu Higher Education Institutions (Jiangsu Education Department [2023] No 11). The funding agency played no role in the study design, data collection, analysis or interpretation, or manuscript preparation.

    Data Availability

    All data analyzed during this study were extracted from published original studies. The extracted data supporting the findings of this study are available from the corresponding author upon reasonable request.

    Authors' Contributions

    XL conceived and designed the study, performed the statistical analysis, and drafted the manuscript. LW contributed to the literature search and data extraction. LY, HC, CW, SZ, TT, and YC contributed to manuscript review and editing. KL and ND supervised the study and critically revised the manuscript for important intellectual content. KL and ND are cocorresponding authors.

    Conflicts of Interest

    None declared.

    Multimedia Appendix 1

    Details of the search strategy.

    DOC File, 97 KB

    Multimedia Appendix 2

    Full characteristics of the included studies.

    DOCX File, 93 KB

    Multimedia Appendix 3

    Related materials.

    DOCX File, 506 KB

    Multimedia Appendix 4

    ChatGPT.

    DOC File, 80 KB

    Checklist 1

    PRISMA-DTA checklist.

    DOC File, 59 KB

    Checklist 2

    PRISMA 2020 expanded checklist.

    DOCX File, 47 KB
    1. Benjafield AV, Ayas NT, Eastwood PR, et al. Estimation of the global prevalence and burden of obstructive sleep apnoea: a literature-based analysis. Lancet Respir Med. Aug 2019;7(8):687-698. [CrossRef] [Medline]
    2. Wang X, Ouyang Y, Wang Z, Zhao G, Liu L, Bi Y. Obstructive sleep apnea and risk of cardiovascular disease and all-cause mortality: a meta-analysis of prospective cohort studies. Int J Cardiol. Nov 2013;169(3):207-214. [CrossRef] [Medline]
    3. Senaratna CV, Perret JL, Lodge CJ, et al. Prevalence of obstructive sleep apnea in the general population: a systematic review. Sleep Med Rev. Aug 2017;34:70-81. [CrossRef] [Medline]
    4. Abbasi A, Gupta SS, Sabharwal N, et al. A comprehensive review of obstructive sleep apnea. Sleep Sci. 2021;14(2):142-154. [CrossRef] [Medline]
    5. AlGhanim N, Comondore VR, Fleetham J, Marra CA, Ayas NT. The economic impact of obstructive sleep apnea. Lung. 2008;186(1):7-12. [CrossRef] [Medline]
    6. Nigro CA, Malnis S, Dibur E, Rhodius E. How reliable is the manual correction of the autoscoring of a level IV sleep study (ApneaLink™) by an observer without experience in polysomnography? Sleep Breath. Jun 2012;16(2):275-279. [CrossRef] [Medline]
    7. Yu Y, Huang JJ, Yang H, Ren LJ. Automated diagnostic method for sleep apnea and hypopnea using overnight airflow and oxygen saturation. MethodsX. Dec 2025;15:103528. [CrossRef] [Medline]
    8. Thommandram A, Eklund JM, McGregor C. Detection of apnoea from respiratory time series data using clinically recognizable features and kNN classification. Annu Int Conf IEEE Eng Med Biol Soc. 2013;2013:5013-5016. [CrossRef] [Medline]
    9. Al-Abed M, Manry M, Burk JR, Lucas EA, Behbehani K. A method to detect obstructive sleep apnea using neural network classification of time-frequency plots of the heart rate variability. 2007. Presented at: 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society; Aug 22-26, 2007. [CrossRef]
    10. Ji Z, Li L, Zheng M, et al. Conductive hydrogel-enabled electrode for scalp electroencephalography monitoring. Small Methods. Apr 2026;10(7):e01242. [CrossRef] [Medline]
    11. Fathima S, Ahmed M. Sleep apnea detection using EEG: a systematic review of datasets, methods, challenges, and future directions. Ann Biomed Eng. May 2025;53(5):1043-1067. [CrossRef] [Medline]
    12. Stadelmann K, Latshang TD, Tarokh L, et al. Sleep respiratory disturbances and arousals at moderate altitude have overlapping electroencephalogram spectral signatures. J Sleep Res. Aug 2014;23(4):463-468. [CrossRef] [Medline]
    13. Jonna ST, Natarajan K. EEG signal processing in neurological conditions using machine learning and deep learning methods: a comprehensive review. Eur Phys J Spec Top. Oct 2025;234(15):3981-3999. [CrossRef]
    14. Ferreira-Santos D, Amorim P, Silva Martins T, Monteiro-Soares M, Pereira Rodrigues P. Enabling early obstructive sleep apnea diagnosis with machine learning: systematic review. J Med Internet Res. Sep 30, 2022;24(9):e39452. [CrossRef] [Medline]
    15. Tan BKJ, Gao EY, Tan NKW, et al. Machine listening for OSA diagnosis: a Bayesian meta-analysis. Chest. Aug 2025;168(2):520-530. [CrossRef] [Medline]
    16. Abd-Alrazaq A, Aslam H, AlSaad R, et al. Detection of sleep apnea using wearable AI: systematic review and meta-analysis. J Med Internet Res. Sep 10, 2024;26:e58187. [CrossRef] [Medline]
    17. Rethlefsen ML, Kirtley S, Waffenschmidt S, et al. PRISMA-S: an extension to the PRISMA statement for reporting literature searches in systematic reviews. Syst Rev. Jan 26, 2021;10(1):39. [CrossRef] [Medline]
    18. Zhang DH, Zhou J, Wickens JD, Veale AG, Hallum LE. Apnea and hypopnea event detection using EEG, EMG, and sleep stage labels in a cohort of patients with suspected sleep apnea. medRxiv. Preprint posted online on Oct 25, 2024. [CrossRef]
    19. Shahnaz C, Minhaz AT, Ahamed S. Sub-frame based apnea detection exploiting delta band power ratio extracted from EEG signals. In: Shahnaz C, Minhaz AT, Ahamed ST, editors. 2016. Presented at: TENCON 2016—2016 IEEE Region 10 Conference; Nov 22-25, 2016. [CrossRef]
    20. Shamsi H. Alzheimer’s diagnosis from EEG with reliable probabilities: subject-wise, leakage-free evaluation and isotonic calibration. J Eng Appl Sci. Dec 2025;72(1):226. [CrossRef]
    21. Whiting PF, Rutjes AWS, Westwood ME, et al. QUADAS-2: a revised tool for the quality assessment of diagnostic accuracy studies. Ann Intern Med. Oct 18, 2011;155(8):529-536. [CrossRef] [Medline]
    22. Moons KGM, Damen JAA, Kaul T, et al. PROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. BMJ. Mar 24, 2025;388:e082505. [CrossRef] [Medline]
    23. Cheung MWL. A guide to conducting a meta-analysis with non-independent effect sizes. Neuropsychol Rev. Dec 2019;29(4):387-396. [CrossRef] [Medline]
    24. Borenstein M. How to understand and report heterogeneity in a meta-analysis: the difference between I-squared and prediction intervals. Integr Med Res. Dec 2023;12(4):101014. [CrossRef] [Medline]
    25. IntHout J, Ioannidis JPA, Borm GF. The Hartung-Knapp-Sidik-Jonkman method for random effects meta-analysis is straightforward and considerably outperforms the standard DerSimonian-Laird method. BMC Med Res Methodol. Feb 18, 2014;14:25. [CrossRef] [Medline]
    26. Cheng L, Luo S, Li B, Liu R, Zhang Y, Zhang H. Multiple-instance learning for EEG based OSA event detection. Biomed Signal Process Control. Feb 2023;80:104358. [CrossRef]
    27. Barnes LD, Lee K, Kempa-Liehr AW, Hallum LE. Detection of sleep apnea from single-channel electroencephalogram (EEG) using an explainable convolutional neural network (CNN). PLoS One. 2022;17(9):e0272167. [CrossRef] [Medline]
    28. Mahmud T, Aeioub Ansary M, Mahmud TI, Khan IA, Fattah SA. Real time sleep apnea event detection with deep neural network. 2019. Presented at: 2019 IEEE International Conference on Biomedical Engineering, Computer and Information Technology for Health (BECITHCON); Nov 28-30, 2019. [CrossRef]
    29. Jiang D, Ma Y, Wang Y. A multi-scale parallel convolutional neural network for automatic sleep apnea detection using single-channel EEG signals. In: Wang Y, Ma Y, Wang Y, editors. 2018. Presented at: 2018 11th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI 2018); Oct 13-15, 2018. [CrossRef]
    30. Emin Tagluk M, Sezgin N. A new approach for estimation of obstructive sleep apnea syndrome. Expert Syst Appl. May 2011;38(5):5346-5351. [CrossRef]
    31. Lin R, Lee RG, Tseng CL, Zhou HK, Chao CF, Jiang JA. A new approach for identifying sleep apnea syndrome using wavelet transform and neural networks. Biomed Eng Appl Basis Commun. Jun 25, 2006;18(3):138-143. [CrossRef]
    32. Zhang Y, Li Z, Lu Y, Zhang Y, Liu G, Wang C. An EEG screening method for severe obstructive sleep apnea based on limited penetrable difference visibility graph and graph convolutional network. IEEE J Biomed Health Inform. 2025;29(11):8011-8021. [CrossRef]
    33. Wang Y, Ji S, Yang T, Wang X, Wang H, Zhao X. An efficient method to detect sleep hypopnea- apnea events based on EEG signals. IEEE Access. 2021;9:641-650. [CrossRef]
    34. Prucnal MA, Polak AG. Analysis of features extracted from EEG epochs by discrete wavelet decomposition and Hilbert transform for sleep apnea detection. 2018. Presented at: 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC); Jul 17-21, 2018. [CrossRef]
    35. Delimayanti MK, Muharram AT, Pradiptyas A, et al. Automated sleep apnea detection using CNNs: insights into the impact of FFT feature extraction on EEG signals. J Adv Inf Technol. 2025;16(9):1217-1225. [CrossRef]
    36. Zhou J, Wu XM, Zeng WJ. Automatic detection of sleep apnea based on EEG detrended fluctuation analysis and support vector machine. J Clin Monit Comput. Dec 2015;29(6):767-772. [CrossRef] [Medline]
    37. Saha S, Bhattacharjee A, Fattah SA. Automatic detection of sleep apnea events based on inter-band energy ratio obtained from multi-band EEG signal. Healthc Technol Lett. Jun 2019;6(3):82-86. [CrossRef] [Medline]
    38. Gupta R, Zaidi TF, Farooq O. Automatic detection of sleep apnea using sub-band features from EEG signals. In: Gupta R, Zaidi TF, editors. Presented at: 2020 3rd International Conference on Signal Processing and Information Security (ICSPIS); Nov 25-26, 2020. [CrossRef]
    39. Wang Y, Xiao Z, Fang S, Li W, Wang J, Zhao X. BI - Directional long short-term memory for automatic detection of sleep apnea events based on single channel EEG signal. Comput Biol Med. Mar 2022;142:105211. [CrossRef]
    40. Zhao X, Wang X, Yang T, et al. Classification of sleep apnea based on EEG sub-band signal characteristics. Sci Rep. 2021;11(1):5824. [CrossRef]
    41. Bonner M, Nikolai B, Glovier Q, Lamb A, Gambhir A. Deep learning-based EEG analysis for sleep apnea detection. Presented at: 2024 Systems and Information Engineering Design Symposium (SIEDS); May 3-3, 2024. [CrossRef]
    42. Gurrala V, Yarlagadda P, Koppireddi P. Detection of sleep apnea based on the analysis of sleep stages data using single channel EEG. Traitement du Signal. Apr 30, 2021;38(2):431-436. [CrossRef]
    43. Taran S, Bajaj V, Sinha GR, Polat K. Detection of sleep apnea events using electroencephalogram signals. Appl Acoust. Oct 2021;181:108137. [CrossRef]
    44. Khan A, Biswas SK, Chunka C. ESAD: expert system for apnea detection using enhanced DWT feature extraction and machine learning algorithms. In: Khan A, Biswas SK, Chunka C, editors. Presented at: 2023 14th International Conference on Computing Communication and Networking Technologies (ICCCNT); Jul 6-8, 2023. [CrossRef]
    45. Prucnal MA, Polak AG. Single-channel EEG processing for sleep apnea detection and differentiation. Metrol Meas Syst. 2023;30:323-336. [CrossRef]
    46. Wijaya RS, Djamal EC, Kasyidi F. Sleep apnea identification based on EEG signals using hybrid spatio-temporal deep learning. 2024. Presented at: 2024 International Conference on Computer, Control, Informatics and Its Applications (IC3INA); Oct 9-10, 2024. [CrossRef]
    47. Bhalerao SV, Pachori RB. Sparse spectrum based swarm decomposition for robust nonstationary signal analysis with application to sleep apnea detection from EEG. Biomed Signal Process Control. Aug 2022;77:103792. [CrossRef]
    48. Taran S, Bajaj V, Sharma D. TEO separated AM-FM components for identification of apnea EEG signals. In: Sharma D, Bajaj V, Sharma D, editors. Presented at: 2017 IEEE 2nd International Conference on Signal and Image Processing (ICSIP); Aug 4-6, 2017. [CrossRef]
    49. Sharifi P, Fakharzadeh M. Algorithm for EEG—based sleep—wake classification toward sleep apnea detection. Presented at: 2025 32nd National and 10th International Iranian Conference on Biomedical Engineering (ICBME); Nov 19-20, 2025. [CrossRef]
    50. Saha S, Bhattacharjee A, Fattah SA. An apnea detection method based on temporal feature variational pattern of multi-band EEG signal incorporating sleep stage information. Circuits Syst Signal Process. Mar 25, 2026. [CrossRef]
    51. Band NC, Deshmukh C. Heuristic deep learning framework for EEG-based sleep apnea event classification. Int Res J Multidiscip Scope. 2026;07(1):1656-1665. [CrossRef]
    52. van Enst WA, Ochodo E, Scholten R, Hooft L, Leeflang MM. Investigation of publication bias in meta-analyses of diagnostic test accuracy: a meta-epidemiological study. BMC Med Res Methodol. May 23, 2014;14(1):70. [CrossRef] [Medline]
    53. Deeks JJ. Systematic reviews in health care: systematic reviews of evaluations of diagnostic and screening tests. BMJ. Jul 21, 2001;323(7305):157-162. [CrossRef] [Medline]
    54. Ito K, Ikeda T. Accuracy of Type III portable monitors for diagnosing obstructive sleep apnea. Biomed Hub. 2018;3(2):1-10. [CrossRef] [Medline]
    55. Saha S, Kabir M, Montazeri Ghahjaverestan N, et al. Portable diagnosis of sleep apnea with the validation of individual event detection. Sleep Med. May 2020;69:51-57. [CrossRef] [Medline]
    56. Kang C, An S, Kim HJ, et al. Age-integrated artificial intelligence framework for sleep stage classification and obstructive sleep apnea screening. Front Neurosci. Jun 2023;17:1059186. [CrossRef] [Medline]
    57. Zhuravlev M, Kiselev A, Orlova A, et al. Changes in the spatial structure of synchronization connections in EEG during nocturnal sleep apnea. Clocks Sleep. Dec 31, 2024;7(1):1. [CrossRef] [Medline]
    58. Prucnal MA, Polak AG. Effectiveness of sleep apnea detection based on one vs. two symmetrical EEG channels. In: Prucnal MA, Polak AG, editors. 2019. Presented at: 2019 41st Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC); Jul 23-27, 2019. [CrossRef]
    59. Banville H, Wood SUN, Aimone C, Engemann DA, Gramfort A. Robust learning from corrupted EEG with dynamic spatial filtering. Neuroimage. May 1, 2022;251:118994. [CrossRef] [Medline]
    60. Zhang X, He G, Shang T, Fan F. Comparative analysis of single-channel and multi-channel classification of sleep stages across four different data sets. Brain Sci. Nov 28, 2024;14(12):1201. [CrossRef] [Medline]
    61. Li Y, Zeng W, Dong W, et al. A tale of single-channel electroencephalogram: devices, datasets, signal processing, applications, and future directions. IEEE Trans Instrum Meas. Mar 17, 2025;74:1-20. [CrossRef]
    62. Gao S, Bibineyshvili Y, Safavynia SA, Calderón-Martínez J, Grinspan ZM, Calderon DP. Cortical signatures linked to behavior quantitatively track arousal levels. Proc Natl Acad Sci U S A. May 13, 2025;122(19):e2413789122. [CrossRef] [Medline]
    63. Khor YH, Khung SW, Ruehland WR, et al. Portable evaluation of obstructive sleep apnea in adults: a systematic review. Sleep Med Rev. Apr 2023;68:101743. [CrossRef] [Medline]
    64. Zhou G, Pan Y, Yang J, Zhang X, Guo X, Luo Y. Sleep electroencephalographic response to respiratory events in patients with moderate sleep apnea-hypopnea syndrome. Front Neurosci. 2020. [CrossRef] [Medline]
    65. Mouazen B, Bendaouia A, Bellakhdar O, et al. Transparent EEG analysis: leveraging autoencoders, Bi-LSTMs, and SHAP for improved neurodegenerative diseases detection. Sensors (Basel). Sep 12, 2025;25(18):5690. [CrossRef] [Medline]
    66. Berry RB, Budhiraja R, Gottlieb DJ, et al. Rules for scoring respiratory events in sleep: update of the 2007 AASM Manual for the Scoring of Sleep and Associated Events. Deliberations of the Sleep Apnea Definitions Task Force of the American Academy of Sleep Medicine. J Clin Sleep Med. Oct 15, 2012;8(5):597-619. [CrossRef] [Medline]
    67. Lee KO, Bekinschtein TA, Smith IE. Multi-night home testing with portable EEG shows improvement in sleep quality after starting CPAP for obstructive sleep apnoea. Sleep Sci Pract. 2025;9(1). [CrossRef]
    68. Seol J, Chiba S, Kawana F, et al. Validation of sleep-staging accuracy for an in-home sleep electroencephalography device compared with simultaneous polysomnography in patients with obstructive sleep apnea. Sci Rep. Feb 12, 2024;14(1):3533. [CrossRef] [Medline]
    69. Malhotra A, Ayappa I, Ayas N, et al. Metrics of sleep apnea severity: beyond the apnea-hypopnea index. Sleep. Jul 9, 2021;44(7):zsab030. [CrossRef] [Medline]
    70. Ney JP, Nuwer MR, Hirsch LJ, Burdelle M, Trice K, Parvizi J. The cost of after-hour electroencephalography. Neurol Clin Pract. Apr 2024;14(2):e200264. [CrossRef] [Medline]
    71. Yu Z, Cheng W. Enhancing interpretability and clinical integration of machine learning models for home blood pressure monitoring adherence. Hypertens Res. Apr 2026;49(4):1554-1555. [CrossRef] [Medline]
    72. Ney JP, Gururangan K, Parvizi J. Modeling the economic value of ceribell rapid response EEG in the inpatient hospital setting. J Med Econ. 2021;24(1):318-327. [CrossRef] [Medline]
    73. Rajpurkar P, Chen E, Banerjee O, Topol EJ. AI in health and medicine. Nat Med. Jan 2022;28(1):31-38. [CrossRef] [Medline]
    74. Band SS, Yarahmadi A, Hsu CC, et al. Application of explainable artificial intelligence in medical health: a systematic review of interpretability methods. Inform Med Unlocked. 2023;40:101286. [CrossRef]
    75. Davis SE, Greevy RA, Lasko TA, Walsh CG, Matheny ME. Detection of calibration drift in clinical prediction models to inform model updating. J Biomed Inform. Dec 2020;112:103611. [CrossRef] [Medline]

    ‎
    AHI: apnea-hypopnea index
    AI: artificial intelligence
    AUC: area under the curve
    DL: deep learning
    ECG: electrocardiogram
    EEG: electroencephalogram
    FN: false negative
    FP: false positive
    GRADE: Grading of Recommendations Assessment, Development and Evaluation
    ML: machine learning
    OSA: obstructive sleep apnea
    PRISMA: Preferred Reporting Items for Systematic Reviews and Meta-Analyses
    PRISMA-DTA: Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy
    PRISMA-S: Preferred Reporting Items for Systematic Reviews and Meta-Analyses Literature Search Extension
    PROBAST+AI: Prediction Model Risk of Bias Assessment Tool for Artificial Intelligence
    PSG: polysomnography
    QUADAS-2: Quality Assessment of Diagnostic Accuracy Studies-2
    SA: sleep apnea
    TN: true negative
    TP: true positive

    Edited by Stefano Brini; submitted 12.Feb.2026; peer-reviewed by Haoqi Sun, Zekai Yu; final revised version received 11.Jun.2026; accepted 12.Jun.2026; published 31.Jul.2026.

    Copyright

    © Xiangshuo Li, Lulu Wang, Ting Tang, Yuanyuan Chen, Lan Yang, Hao Cai, Chen Wang, Shuxiao Zhang, Ning Ding, Kouying Liu. Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 31.Jul.2026.

    This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research (ISSN 1438-8871), is properly cited. The complete bibliographic information, a link to the original publication on https://www.jmir.org/, as well as this copyright and license information must be included.