This review followed PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines (Checklist 1) [24]. We included original reports that clearly defined SCD or MCI, enrolled a relevant sample with a reported mean or median age of 60 years or older, used EEG or event-related potentials, and reported extractable between-group, classification, screening or diagnostic, and prognostic outcomes. For mixed-age populations, separately reported eligible subgroup data were used. We excluded animal reports, unrelated populations, severe-dementia-only reports, nonoriginal publications, duplicate reports without additional relevant data, and reports lacking relevant EEG outcomes or sufficient methodological information.

We searched PubMed, Web of Science, MEDLINE via Ovid, Scopus, the Cochrane Library, CNKI, Wanfang Data, CQVIP, and Yiigle from inception to September 25, 2025, and updated the search on August 22, 2026. Controlled vocabulary and free-text terms related to cognitive impairment, SCD, and EEG were used. Full search strategies are provided in Tables S1 to S9 in Multimedia Appendix 1. The strategy was developed by SL, checked by JH, and supervised by a librarian (MY). Only full-text papers published in English or Chinese were eligible.

After duplicate removal in EndNote (Clarivate Analytics), 2 reviewers (SL and JH) independently screened titles and abstracts and assessed full-text reports. Disagreements were resolved by consensus or adjudication by a third reviewer (SH). Reasons for full-text exclusion were recorded, and reference lists of included reports were manually searched.

Data were independently extracted in duplicate using a piloted standardized form. Extracted variables included study design, participants, diagnostic definitions and reference standards, EEG acquisition and features, outcomes, comparisons, and effect or performance estimates. For prediction-model reports, we additionally extracted development, validation, and test samples; participant-level vs epoch-level splitting; cross-validation; feature selection; hyperparameter tuning; class-imbalance handling; calibration; external validation; and possible data leakage. Missing information was recorded as “not reported,” and study authors were not contacted. No automation tools were used.

When multiple models were reported, we prioritized external validation, followed by participant-level held-out testing and internal cross-validation. At the same validation level, the prespecified primary or final model, or the most completely reported model, was selected. Reports from overlapping cohorts were retained as separate reports but were not counted as independent participant samples.

Evidence was organized into 4 prespecified bodies: between-group EEG differences, EEG-based classification, EEG as a screening or diagnostic test, and prediction of progression to dementia. Results were synthesized narratively because of substantial heterogeneity in populations, EEG methods, reference standards, outcomes, and validation procedures. Consequently, no meta-analysis, subgroup analysis, meta-regression, or sensitivity analysis was performed.

Two reviewers independently assessed risk of bias, with disagreements resolved by consensus or adjudication by SH. Prediction and AI model reports were assessed using PROBAST+AI (Prediction Model Risk of Bias Assessment Tool and Applicability Assessment for AI) [25], nonmodel prognostic-factor reports using QUIPS (Quality in Prognosis Studies), and other observational reports using the corresponding Joanna Briggs Institute (JBI) checklist [26-29]. Because these tools assess different methodological constructs, their findings were reported separately and were not converted into a common cross-tool quality or risk-of-bias scale. PROBAST+AI and QUIPS overall judgments were reported according to the guidance for the respective tools. JBI checklist findings were reported at the item level without numerical scoring or author-defined overall categories. Domain-level PROBAST+AI judgments with study-specific rationales, item-level JBI findings, and domain-level QUIPS judgments are presented in Tables S10A-S14 in Multimedia Appendix 1.

Two reviewers independently assessed certainty for each of the 4 evidence bodies. Between-group and progression evidence was assessed using GRADE (Grading of Recommendations Assessment, Development, and Evaluation) for observational evidence, whereas classification and screening or diagnostic evidence was assessed using adapted GRADE-diagnostic test accuracy principles [30,31]. Certainty was rated as high, moderate, low, or very low according to risk of bias, inconsistency, indirectness, imprecision, and publication or reporting bias.

Because the classification tasks and performance measures were heterogeneous, certainty concerned the reproducibility and applicability of out-of-sample model performance rather than pooled diagnostic accuracy. No conventional diagnostic test-accuracy study was identified; therefore, the PET-referenced model-development study was treated as indirect screening evidence.

Reporting bias was assessed qualitatively for each evidence body by considering protocol availability, suspected missing results, multiplicity of EEG features and models, and language and publication-status restrictions. Funnel plots and small-study-effect tests were not used because no comparable meta-analysis was performed and each evidence body contained fewer than 10 sufficiently comparable reports.

The paper selection process is presented in Figure 1. Among the 20 included reports, 10 models of prediction or AI models were assessed using PROBAST+AI, 9 nondiagnostic observational reports using design-specific JBI checklists, and 1 prognostic-factor report using QUIPS. Because these tools assess different methodological constructs, their findings were not combined into a single cross-tool risk-of-bias category. Tool-specific assessments are summarized below and reported in detail in Tables S10A to S14 in Multimedia Appendix 1.

The corrected sum of report-level sample sizes was 3954. After the 101 participants shared by the 2 Sibilano reports were counted once, the confirmed-deduplicated upper bound was 3853 participants: 1039 with SCD/subjective cognitive impairment (SCI), 876 with MCI, 887 with AD or dementia, and 125 with other dementias, 926 cognitively unimpaired controls. The main characteristics of each included study are presented in Table 1.

Study-level findings and principal limitations across the 4 prespecified evidence bodies are summarized in Table 2. Nine reports examined group-level differences in spectral power, functional connectivity, signal complexity, or event-related potentials among participants with SCD, MCI, AD, other dementias, and cognitively unimpaired controls [32,37,39,42,44,45,47,50,51]. Across the spectral reports, cognitive impairment was generally associated with greater low-frequency activity and reduced higher-frequency activity. Jeong et al [44] reported a partial increase in frontal delta power and a decrease in occipital alpha-1 power in participants with SCD compared with community controls. Zawislak-Fornagiel et al [45] similarly found lower global relative beta power and higher theta and delta power in participants with amnestic MCI than in cognitively unimpaired participants. Among participants with SCD, Shim et al [47] found that participants with amyloid PET–positive had higher delta and lower alpha-1 activity than participants with amyloid PET–negative. Mazzon et al [50] also identified differences in alpha-band activity and EEG complexity among participants with SCI, MCI, and controls, particularly during a memory task.

Connectivity and task-related reports also identified group-level differences, although the specific measures and directions varied. Lazarou et al [32,37] found differences in clustering coefficient, network strength, and other connectome measures across SCD, MCI, AD, and control groups during resting-state and visual memory paradigms. Participants with SCD generally showed intermediate network organization between cognitively unimpaired controls and participants with MCI or AD. Briels et al [39] reported that functional connectivity measures differed significantly across various diagnostic groups in dimensions sensitive to changes associated with AD pathology. Task-related differences were also reported for the N170 event-related potential across SCI, MCI, AD, and control groups [51]. Cave et al [42] assessed the differences in resting-state EEG between older adults with SCI and HCs, revealing distinct neurophysiological signatures despite normal objective cognitive function, thus highlighting the potential of EEG-derived features for research-stage characterization of SCD.

In summary, EEG measures, including connectivity features and machine learning–based approaches, have shown exploratory potential for differentiating SCD from MCI.

Eight reports developed or evaluated statistical, machine-learning, or deep-learning models for classifying cognitive-status groups [34,35,40,43,46,48,50,51]. The classification tasks varied substantially and included binary discrimination between SCD and MCI, discrimination between MCI and cognitively unimpaired controls, and multiclass classification of SCD, MCI, AD, other dementias, and healthy aging. Predictor sets included spectral power, connectivity, event-related potentials, electrode-specific measures, cognitive-task EEG features, and combinations of EEG with demographic, cognitive, MRI, or biochemical information.

Sibilano et al [34,35] evaluated transformer-based models using the same underlying cohort of participants with SCD and MCI. The delta-band data produced the best performance for SCD vs MCI classification, whereas alpha- and theta-band data contributed to the 3-class classification of HC, SCD, and MCI. Because the 2023 and 2024 reports analyzed overlapping participants, their findings were treated as 2 analyses of one cohort rather than as independent validation reports. Huggins et al [40] reported a mean accuracy of 98.9% (SD 0.4%) for 3-class classification of AD, MCI, and healthy aging using an AlexNet-based deep-learning approach. However, the participant-level independence of the image-based data split was insufficiently documented, and no external validation was reported.

Jiang et al [43] evaluated electrode configurations for classifying MCI and controls and reported sensitivity of up to 96.2% for a 4-electrode occipital configuration, although specificity was substantially lower for some configurations, indicating a risk of false-positive classification. Kim et al [46] also presented findings that linked EEG features in resting and memory-encoding states to cognitive performance across different dementia stages, supporting further evaluation of EEG-derived features as candidate biomarkers rather than established tools for early detection. Other reports combined EEG features with cognitive, MRI, demographic, or biochemical variables to classify MCI, AD, or other dementia categories [38,48,49].

All contributing classification-model reports were judged to be at high risk of bias. Common limitations included small effective sample sizes relative to model complexity, inadequate separation of feature selection or hyperparameter tuning from performance evaluation, possible participant- or epoch-level data leakage, incomplete reporting of calibration, and a lack of independent external validation. A concise summary of the target tasks, validation designs, key performance estimates, and principal limitations is presented in Table 3; full model-development and validation details are provided in Table S17 in Multimedia Appendix 1.

No included report met all criteria for a conventional diagnostic test-accuracy study. Specifically, no study evaluated a prespecified and fixed EEG test or threshold against an independent reference standard in a representative clinical screening population while also defining whether EEG was intended as a replacement, add-on, or triage test. Consequently, sensitivity and specificity were not pooled.

One PET-referenced machine-learning report provided indirect evidence relevant to this question. Kim et al developed an EEG-based model to identify amyloid PET positivity among 311 participants with SCD or MCI. The study reported a sensitivity of up to approximately 90% in its model evaluation. However, the analysis remained a model-development study: the final model and threshold had not been prospectively fixed, model selection was conducted among multiple candidate approaches, independent external validation was absent, and the intended position of EEG in a clinical diagnostic pathway was not specified. The results, therefore, did not establish the accuracy of a deployable EEG screening test [36].

Shim et al [47] also compared quantitative EEG features between amyloid PET-positive and amyloid PET-negative participants with SCD. However, this report assessed group-level biomarker associations and did not evaluate a prespecified diagnostic threshold or provide a complete participant-level diagnostic-accuracy analysis. It was therefore synthesized with the group-difference evidence rather than as a diagnostic test-accuracy study.

Accordingly, the available evidence was insufficient to determine whether EEG could be used as a replacement, an add-on, or a triage test for SCD, MCI, AD, or underlying amyloid pathology.

Three longitudinal reports examined associations between baseline EEG measures and subsequent progression to dementia [33,37,41]. Gouw et al [33] investigated participants with amyloid-positive SCD or MCI who did not have dementia at baseline and found that lower alpha peak frequency and higher delta or theta activity were associated with faster clinical progression. However, these findings represent exploratory group-level associations rather than evidence of individual-level prognostic prediction. Across the 3 studies, sample sizes were limited, and none developed and externally validated a multivariable prediction model or reported model calibration. The certainty of this evidence was rated as very low, and the available findings are insufficient to support individualized risk estimation or clinical decision-making.

Engedal et al [41] evaluated an automated quantitative EEG dementia index for predicting conversion among participants with SCD or MCI over approximately 5 years. The index showed moderate discrimination, with reported sensitivity of 71%, specificity of 68%, accuracy of 69%, and an area under the curve of approximately 0.78 in the reported analyses. The preexisting dementia index was evaluated in a new cohort; however, the operating cutoff was selected using the evaluation data. Newly fitted incremental models received only apparent evaluation, and calibration was not reported.

Lazarou et al [37] examined longitudinal changes or progression in relation to baseline high-density EEG network characteristics across preclinical and clinical stages of cognitive impairment. The study provided evidence that altered network organization may be associated with subsequent cognitive deterioration, but the sample was small and no validated individual-level risk prediction model was developed.

The prognostic findings were limited by small sample sizes, differences in the definition of progression, heterogeneous follow-up periods, incomplete adjustment for confounding, and the absence of independent external validation. Thus, baseline EEG may contain prognostic information, but the magnitude, reproducibility, and clinical applicability of this information remain uncertain.

Of the 20 included reports, 10 reports on prediction or AI models were assessed using PROBAST+AI, 9 nondiagnostic observational reports using design-specific JBI checklists, and 1 prognostic-factor report using QUIPS. All 10 model-development assessments were judged to be at high risk of bias. The analysis domain was the predominant source of bias. The main concerns included small effective sample sizes, data-driven feature or model selection, inadequate separation of model development from performance evaluation, possible participant- or epoch-level data leakage, limited calibration assessment, and the absence of independent external validation. The separate model-evaluation components in Engedal et al [41] and Jiang et al [43] were also judged to be at high risk of bias, primarily because of data-dependent cutoff selection and the very small external evaluation sample, respectively.

For the 9 nonmodel observational reports assessed using JBI checklists, findings were reported at the checklist-item level rather than converted into overall low-, moderate-, or high-risk categories. Among the 5 analytical cross-sectional reports, Briels et al [39] and Mazzon et al [50] met all 8 checklist criteria. Shim et al [47] met 7 criteria, with 1 item rated as unclear. Jeong et al [44] and Lazarou et al [32] had concerns mainly related to the identification or management of confounding and the appropriateness of the statistical analysis. The 3 case-control reports had item-level concerns involving group comparability or matching, control of confounding, and statistical analysis. The cohort report had concerns regarding confounding management, completeness of follow-up, and statistical analysis. No composite overall JBI rating was assigned to these reports.

The prognostic-factor report assessed using the QUIPS tool was judged to be at a high overall risk of bias, driven principally by high attrition and additional concerns regarding study participation, confounding, and statistical analysis or reporting (Table 4). Tool-specific judgments and study-specific rationales are presented in Tables S10A to S14 in Multimedia Appendix 1.

The certainty of the evidence was rated as very low for all 4 prespecified outcome bodies (Table 5). Evidence concerning between-group EEG differences was downgraded because of risk of bias, substantial clinical and methodological inconsistency, indirectness to individual-level diagnosis, imprecision, and suspected reporting bias. Classification evidence was limited by high risk of bias in model development, heterogeneous classification tasks and algorithms, reliance on internal validation, possible data leakage, and limited precision. Evidence concerning EEG as a screening or diagnostic test was based on only 1 indirect, PET-referenced, model-development report and did not establish the accuracy of a deployable EEG test within an intended clinical pathway. Prognostic evidence was limited by small samples, heterogeneous definitions of progression, incomplete control of confounding, and the absence of externally validated and calibrated prediction models.

Reporting bias was strongly suspected for all 4 evidence bodies. Most reports had no publicly available protocol or prespecified analysis plan, and many evaluated multiple EEG features, thresholds, comparisons, or candidate models, increasing the possibility of selective reporting. The restriction to English- and Chinese-language full-text publications and the exclusion of gray literature may also have contributed to missing evidence. Funnel plots and statistical tests for small-study effects were not performed because no meta-analysis using a common effect measure was conducted and fewer than 10 sufficiently comparable reports contributed to each evidence body. Detailed GRADE judgments and downgrade rationales are presented in Table S15 in Multimedia Appendix 1.

This systematic review synthesized evidence across 4 related but distinct questions: between-group differences in EEG features, EEG-based classification of cognitive status, the use of EEG as a screening or diagnostic test, and the prediction of progression to dementia. Several reports described potentially informative group-level EEG differences and preliminary classification performance in MCI and SCD; however, these findings do not establish individual-level screening or diagnostic accuracy [52].

Several studies reported potentially informative differences in spectral power, functional connectivity, event-related potentials, and other quantitative EEG measures. Nevertheless, the direction and magnitude of these findings were not fully consistent. Differences in recording conditions, electrode configurations, frequency-band definitions, cognitive tasks, diagnostic criteria, and adjustment for potential confounders may have contributed to this heterogeneity. In addition, many reports examined multiple EEG features without clearly identifying a prespecified primary feature or threshold, which may have increased the possibility of selective reporting.

There were 8 reports that developed or evaluated EEG-based models for classifying individuals with SCD, MCI, AD, or cognitively unimpaired controls. Although several studies reported encouraging discrimination or classification performance, most estimates were derived from relatively small and selected samples and were primarily based on apparent performance or internal validation. All contributing classification-model reports were judged to be at high risk of bias. These judgments largely reflected limited effective sample sizes, nonnested feature selection or hyperparameter tuning, potential participant-level or epoch-level information leakage, reuse of test data during model selection, limited calibration assessment, and scarce independent external validation [53]. These factors may have resulted in the overestimation of model performance. The findings should therefore be considered preliminary evidence of technical feasibility that requires further validation before clinical implementation.

None of the included reports evaluated a prespecified EEG test using all elements typically expected in a conventional diagnostic test-accuracy study, including a representative clinical population, an appropriate reference standard, and a clearly defined clinical pathway. The PET-referenced machine-learning study provided potentially valuable evidence regarding amyloid-status classification, although its relevance to the clinical diagnosis of SCD, MCI, or AD was indirect. Longitudinal studies also suggested that selected EEG measures may be associated with subsequent cognitive progression, but evidence supporting a calibrated and independently validated individual-risk model remains limited. Overall, the certainty of evidence was rated as very low for all 4 prespecified outcomes, indicating that the estimates and interpretations may change as more rigorous evidence becomes available [54].

Emerging trends in EEG research related to MCI and SCD are being significantly influenced by advancements in deep-learning models, multimodal neuroimaging techniques, and candidate biomarkers intended to support research-stage characterization and prognostic evaluation across the AD continuum.

Recent studies have reported preliminary performance of deep-learning approaches in analyzing resting-state EEG signals to classify individuals at different stages of cognitive decline. For instance, Sibilano et al used a transformer architecture with self-attention mechanisms, which provided potentially interpretable attention patterns for the model’s decisions, enabling discrimination between SCD and MCI with notable accuracy. Additionally, their work highlighted how attention scores could facilitate a better understanding of the EEG patterns indicative of cognitive decline [34,35]. Sibilano et al [34,35] further built on this by implementing a deep-learning model that classified participants with SCD and MCI, showing encouraging but internally validated performance through multiple frequency bands of EEG and employing an innovative approach with multihead attention.

Moreover, Jiao et al [48] have highlighted the need for a holistic view, combining EEG metrics with demographic and biochemical data (such as cerebrospinal fluid and apolipoprotein E genotype) to enhance predictive capabilities regarding AD progression. Their findings indicated that EEG-based classifications showed higher performance in the reported comparisons in terms of accuracy, especially in differentiating between MCI and AD when combined with other biomarkers [35].

Furthermore, the exploration of specific EEG responses to emotional stimuli, as investigated by Lazarou et al [51], revealed that the amplitude of the N170 event-related potential could play a critical role in differentiating cognitive statuses. Their work showed significant variations in brain activation patterns among individuals with SCD, MCI, and HCs during the processing of facial stimuli, indicating the potential utility of such metrics in understanding cognitive decline stages.

Finally, Huggins et al [40] developed a deep-learning model for classifying patients with AD, MCI, and healthy aging using resting-state EEG signals. Their results indicated a high apparent accuracy, demonstrating technical feasibility, although the apparent performance requires participant-independent external validation before any diagnostic interpretation.

In summary, this body of research concerning MCI and SCD is rapidly evolving, driven by innovative neural network architectures, integration of multimodal data, and the exploration of novel biomarkers. This trend points toward a future where EEG-derived features may warrant further evaluation as candidate components of multimodal research and future clinical-assessment pathways.

This review has several strengths. First, it organized the evidence according to 4 prespecified questions rather than treating all EEG reports as a single diagnostic evidence base. Second, study selection and data extraction were independently conducted by 2 reviewers using standardized procedures. Third, risk of bias was evaluated using tools matched to the study objective and design: PROBAST+AI for prediction-model reports, QUIPS for prognostic-factor reports, and design-specific JBI checklists for other observational reports. Prediction-model reports were further categorized into development and performance-evaluation components, with signaling-question responses and applicability judgments reported at the model level. Fourth, the certainty of evidence and synthesis-level reporting bias were formally assessed for each outcome body. Reports from confirmed or potentially overlapping cohorts were also identified to reduce duplicate participant counting and overinterpretation of correlated evidence.

This systematic review has several limitations that should be acknowledged. Although 5178 records were initially screened, only 20 studies met the inclusion criteria, which may limit the generalizability of the findings, particularly for underrepresented populations or research methodologies. The search was restricted to English- and Chinese-language publications and predominantly focused on peer-reviewed literature. Studies published in other languages, dissertations, conference proceedings, and unpublished reports may therefore have been missed, increasing the possibility of publication and reporting bias. Finally, most included studies were small, methodologically heterogeneous, and conducted in selected research samples. This review was not prospectively registered, and no publicly accessible protocol was prepared; consequently, independent verification of whether all methodological decisions were prespecified is limited.

This systematic review suggests that EEG is a promising noninvasive modality for investigating neurophysiological alterations across SCD, MCI, AD, and cognitively unimpaired populations. The included reports identified potentially informative group-level differences in spectral, connectivity, event-related potential, and other quantitative EEG measures. Several EEG-based classification and prognostic models also reported encouraging preliminary performance. Together, these findings support continued investigation of EEG as a candidate biomarker and a model-development platform for early cognitive impairment.

Nevertheless, substantial heterogeneity in study populations, diagnostic definitions, EEG acquisition and preprocessing, feature selection, reference standards, and validation procedures limits the direct clinical interpretation of the available findings. Most model-performance estimates were derived from selected samples and internal validation, and none of the included reports evaluated a fixed EEG test within a representative clinical diagnostic pathway. The certainty of the evidence therefore remains very low. Rather than establishing immediate clinical applicability, the present review defines the methodological steps needed to advance the field. Addressing these methodological priorities may help determine which EEG features and models are sufficiently reproducible and transportable for clinical evaluation. With rigorous prospective validation, EEG may ultimately contribute as an adjunctive or triage tool within multimodal pathways for the assessment of early cognitive impairment.