SSDLabeler: realistic semi-synthetic data generation for multi-label artifact classification in EEG

Abstract

EEG recordings are inherently contaminated by artifacts such as ocular, muscular, and environmental noise, which obscure neural activity and complicate preprocessing. Artifact classification offers advantages in stability and transparency, providing a viable alternative to ICA-based methods that enable flexible use alongside human inspections and across various applications. However, artifact classification is limited by its training data as it requires extensive manual labeling, which cannot fully cover the diversity of real-world EEG. Semi-synthetic data (SSD) methods have been proposed to address this limitation, but prior approaches typically injected single artifact types using ICA components or required separately recorded artifact signals, reducing both the realism of the generated data and the applicability of the method. To overcome these issues, we introduce SSDLabeler, a framework that generates realistic, annotated SSDs by decomposing real EEG with ICA, epoch-level artifact verification using RMS and PSD criteria, and reinjecting multiple artifact types into clean data. When applied to train a multi-label artifact classifier, it improved accuracy on raw EEG across diverse conditions compared to prior SSD and raw EEG training, establishing a scalable foundation for artifact handling that captures the co-occurrence and complexity of real EEG.

Introduction

Electroencephalography (EEG) is a widely used, non-invasive neuroimaging technique that measures brain activity from the scalp. EEG’s high temporal resolution, portability, and relatively low cost have made it indispensable in neuroscience research, clinical diagnostics, machine learning applications, and brain–computer interface (BCI) development^1–3. EEG signals reflect small electrical potentials generated by synchronized neuronal firing, yet the sensitivity required to detect these weak signals also makes EEG highly susceptible to contamination from non-neural sources, including ocular and muscular activity, head movements, electrode fluctuations, and environmental interference. These artifacts frequently overlap with genuine brain signals in both time and frequency, degrading the signal-to-noise ratio (SNR) and reducing the reliability of downstream analyses such as neural decoding and event-related feature extraction. Despite the diversity of EEG applications, including event-related investigations in neuroscience^4,5, clinical diagnostics, and machine-learning-based classification^6–10, all users share the need to preserve meaningful neural activity while controlling for contaminating artifacts. Achieving this balance requires preprocessing methods that enhance interpretability without distorting or removing genuine neural events.

Current artifact handling strategies, however, face several limitations. Manual inspection remains common but is subjective, labor-intensive, and reliant on the experience of the reviewer¹¹. Independent Component Analysis (ICA) offers a semi-automated alternative, and ICLabel is widely adopted for classifying ICA components into brain, eye, muscle, heart, line, channel, or other categories^12–16. Yet ICA-based artifact removal eliminates entire components, which risks discarding neural signals that co-occur with artifacts. Furthermore, ICA performance is sensitive to data quality, channel count, and subject variability, limiting its reliability and necessitating expert verification¹⁷. More recent deep learning–based approaches, including end-to-end denoising models as well as convolutional attention and transformer-based architectures, have demonstrated strong performance in EEG analysis and representation learning^18–22. However, these methods often operate as black boxes and are typically developed using preprocessed EEG data, without explicitly modeling or identifying artifact contamination, which can limit robustness under varying noise conditions. In denoising models, this may additionally lead to the unintended suppression or distortion of meaningful neural features, including event-related potentials (ERPs). Their performance is further constrained by the limited availability and variability of artifact-labeled training data. Earlier artifact classification approaches relying on principle component analysis (PCA), ICA, or wavelet features were similarly limited by the scarcity of labeled data and generally underperformed compared with contemporary approaches^23–25. Nevertheless, artifact classification retains a key advantage: transparency and user control, enabling researchers to decide how to handle identified contamination rather than relying on fully automated removal.

Most existing artifact classification approaches operate under a binary “clean versus artifact” assumption, which overlooks the fact that different artifact types frequently co-occur and can have distinct effects on downstream analysis¹¹. To address this, we propose a multi-label artifact classification framework that identifies multiple artifact types (eye, muscle, channel, and other noise) within a single time-locked EEG segment (epoch). This provides transparent, per-epoch contamination information and offers users flexibility in how they handle flagged data. Supporting multi-label classification requires a diverse and realistic set of labeled examples, however, artifact-labeled datasets are scarce and inherently limited in their coverage of real-world variability.

To overcome this fundamental bottleneck of limited availability of realistic, multi-label artifacts EEG, we introduce a novel semi-synthetic data (SSD) generation method (SSDLabeler) that produces realistic multi-label contaminations reflective of those observed in raw EEG. Prior SSD approaches have primarily injected single artifact types into clean EEG, producing isolated contaminations that fail to capture the realistic co-occurrence structure found in raw EEG^18,19. In contrast, our framework, SSDLabeler, reconstructs artifact sources from ICA-derived components and uses root-mean-square (RMS) and power spectral density (PSD) criteria to verify the presence of artifacts at the epoch level before reinjection. By simultaneously mixing multiple artifact components, our SSDLabeler preserves both spatiotemporal artifact signatures and underlying brain activity, generating realistic multi-label contamination suitable for training robust classifiers. To promote robustness and external validity, all hyperparameters of the proposed SSDLabeler framework, including ICLabel confidence thresholds and RMS- and PSD-based verification criteria, were optimized using an independent external dataset prior to evaluation. This strategy minimizes dataset-specific tuning and supports the interpretation that performance improvements reflect generalizable modeling choices rather than adaptation to a particular recording setup. By providing a fully automated and transparently defined artifact-generation framework, SSDLabeler reduces subjective bias and enables reproducible benchmarking of artifact classification methods across diverse EEG applications.

Beyond improving classification, the proposed approach enables several practical applications. It supports objective, automated reporting for trial or epoch rejection, reducing subjectivity and improving reproducibility across studies. It can be applied prior to ICA to exclude highly contaminated segments, improving ICA stability and producing more reliable component labeling^26,27. It also facilitates selective cleaning in averaging-based analyses such as ERP or PSD, preventing contaminated epochs from distorting neural feature estimates. Motivated by these applications, the present study demonstrates the feasibility of combining multi-label SSD generation with multi-label artifact classification as a scalable, transparent, and researcher-centered approach to EEG preprocessing, aimed at preserving neural information while providing actionable and interpretable artifact classification.

This study makes the following contributions:

• We propose SSDLabeler, an ICA-based semi-synthetic EEG generation framework that models realistic, multi-artifact contamination with explicit multi-label annotations.

• We introduce RMS- and PSD-based artifact verification to improve label fidelity by retaining only epochs with verifiable artifact activity.

• We establish a transparent and externally validated hyperparameter optimization strategy that enhances reproducibility.

• We validate the proposed framework on both naturalistic motor execution EEG and controlled instructed-noise recordings using statistically rigorous comparisons.

• We demonstrate that our proposed SSDlabeler significantly improves multi-label artifact classification compared to raw EEG training and prior single-artifact SSD approaches.

Results

Qualitative assessment of semi-synthetic EEG signal realism

To qualitatively assess the realism of the semi-synthetic EEG, we compare representative examples of raw EEG, clean EEG (artifact-reduced), and semi-synthetic EEG generated using the proposed SSDLabeler framework. All examples are shown from a single subject and a single trial to preserve temporal characteristics and transient artifact behavior, which may be obscured by averaging across trials or subjects. Representative channels were selected to illustrate signal behavior across distinct scalp regions with different physiological and artifact-related characteristics. Fpz was selected due to its high sensitivity to ocular artifacts (Fig. 1), Cz was chosen as a central midline electrode reflecting typical EEG morphology (Fig. 2), and T8 was included for its susceptibility to muscle-related and non-neural contamination (Fig. 3). At Cz, where artifact contamination is relatively low, the raw, clean, and semi-synthetic signals exhibit highly similar temporal and spectral characteristics. In contrast, differences are more apparent at Fpz and T8. In the spectral domain, the semi-synthetic signals preserve the overall spectral profile of the raw EEG, while the clean signals show reduced amplitude and decreased power, reflecting the removal of artifact components. At T8, this is particularly evident in the attenuation of higher-frequency power following artifact removal, whereas the semi-synthetic signal retains spectral characteristics consistent with the raw signal. These observations support the realism of the generated signals and their suitability for downstream classification tasks.

Fig. 1.

Example of raw, clean, and semi-synthetic EEG signals at the frontal electrode Fpz. The semi-synthetic signal preserves the temporal structure of the raw EEG while reproducing prominent ocular artifact characteristics typical of frontal regions. Power spectral density (PSD) analysis shows similar spectral profiles between raw and semi-synthetic signals, with higher power at lower frequencies, while the clean signal exhibits reduced amplitude and low-frequency power due to artifact removal.

Fig. 2.

Example of raw, clean, and semi-synthetic EEG signals at the central electrode Cz. At Cz, where artifact contamination is relatively low, the raw, semi-synthetic, and clean signals exhibit similar temporal and spectral characteristics.

Fig. 3.

Example of raw, clean, and semi-synthetic EEG signals at the temporal electrode T8. PSD analysis shows that the semi-synthetic signal preserves spectral characteristics similar to the raw EEG, particularly in higher frequency ranges associated with muscle activity, while the clean signal exhibits reduced power and a smoother spectral profile following artifact removal.

Multi-label artifact classification performance

We evaluated multi-label artifact classification performance using data from five subjects that were held out from training and validation. Three models were trained separately using (i) raw EEG, (ii) the SSD baseline from prior work, and (iii) our proposed SSDLabeler generated SSD dataset. Raw EEG and our proposed SSDLabeler generated SSD dataset were annotated using the SSDLabeler procedure, while the SSD baseline was annotated using the prior work’s method¹⁹. The resulting models were then tested on manually annotated artifacts in the Cho et al. dataset, including both motor execution (ME) and noise sessions, to provide an unbiased assessment of generalization across artifact types²⁸. The SSD-based training datasets consisted of three components: (i) semi-synthetic EEG generated via ICA-based artifact reinjection, (ii) clean EEG epochs () reconstructed from brain ICs, and (iii) the original raw EEG epochs. In contrast, the raw EEG dataset consisted only of the original raw EEG epochs and the clean EEG epochs.

The Heart and Line artifact categories were not included in the evaluation, as no corresponding instances were present in the dataset.

Motor execution session

On the ME test dataset, the artifact classification model trained using our SSDLabeler generated data achieved the highest overall accuracy (0.826), surpassing both raw EEG (0.760) and the SSD baseline from previous work (0.702). Category-wise performance is summarized in Table 1. Model trained on our SSDLabeler achieved the highest accuracies for the Eye (0.813) and Muscle (0.845) categories, while matching the best performance in the Channel (0.833) and Other (0.803) categories. In contrast, SSDLabeler internal labeling verification achieved the highest performance for the Clean category (0.860). For reference, Table 1 reports SSDLabeler’s internal labeling verification results, which quantify the agreement between the proposed SSDLabeler annotation procedure and the manually annotated labels, rather than the artifact classification model performance.

Table 1.

Classification accuracy by artifact category in the ME session (human-labeled test set).

Method	Clean	Eye	Muscle	Channel	Other	Average
Model trained on Raw EEG (SSD annotated)	0.813	0.688	0.685	0.813	0.803	0.760
Model trained on Previous Work’s SSD19	0.733	0.558	0.785	0.833	0.603	0.702
Model trained on SSDLabeler	0.835	0.813	0.845	0.833	0.803	0.826
SSDLabeler (internal label verification only)	0.860	0.773	0.600	0.833	0.803	0.774

Model trained on Raw EEG, Previous Work’s SSD and SSDLabeler refers to the accuracies from the artifact classification model. SSDLabeler (internal label verification only) reports the agreement between SSD-generated artifact labels and manual annotations, and does not correspond to a trained classification model.

To examine whether these artifact classification model performance differences reflected statistically meaningful changes in error distributions, we conducted McNemar’s tests for each artifact label, comparing (i) raw EEG with our SSDLabeler, (ii) raw EEG with the previous work’s SSD, and (iii) the previous work’s SSD with our SSDLabeler. Holm-corrected p-values indicated significant improvements of our SSDLabeler over raw EEG in the Clean, Eye, and Muscle categories (p < 0.05) seen in Fig. 4a. When comparing raw EEG and the previous work’s SSD, significant differences were revealed in the Eye, Muscle, and Other categories (p < 0.05). Raw EEG outperformed the previous work’s SSD in the Eye and Other categories, whereas the previous work’s SSD outperformed raw EEG in the Muscle categories. The complete set of McNemar’s test results is visualized in Fig. 4b. Finally, comparisons between our SSDLabeler and the previous work’s SSD, significant differences were observed in the Clean, Eye, Muscle, and Other categories (p < 0.05) seen in Fig. 4c. Our SSDLabeler outperformed the previous work’s SSD in the mentioned categories.

Fig. 4.

McNemar’s test results for per-label classification performance across the three training datasets on the motor execution session test data. The top row (a) compares Raw EEG with SSDLabeler, the middle row (b) compares Raw EEG with the Previous Work’s SSD, and the bottom row (c) compares the Previous Work’s SSD with SSDLabeler. In all panels, columns correspond to artifact categories in the following order: Clean, Eye, Muscle, Channel, and Other.

Noise session

In the noise session containing instructed artifacts, the artifact classification model trained using our SSDLabeler generated data achieved the highest overall accuracy (0.751), outperforming both raw EEG (0.478) and the previous work’s SSD (0.659). Category-wise results are shown in Table 2. SSDLabeler achieved the highest accuracy in the Eye category (0.859). For the Channel category, SSDLabeler and the previous work’s SSD achieved identical performance (0.965), while for the Other category, SSDLabeler and raw EEG showed equivalent performance (0.482). In contrast, the previous work’s SSD achieved the highest accuracy for the Clean category (1.0), exceeding SSDLabeler (0.953), and also performed best in the Muscle category (0.624).

Table 2.

Classification accuracy by artifact category in the noise session (human-labeled test set).

Method	Clean	Eye	Muscle	Channel	Other	Average
Model trained on Raw EEG (SSD annotated)	0.435	0.141	0.459	0.871	0.482	0.478
Model trained on Previous Work’s SSD19	1.0	0.377	0.624	0.965	0.329	0.659
Model trained on SSDLabeler	0.953	0.859	0.494	0.965	0.482	0.751

To evaluate whether these differences reflected significant changes in error patterns, McNemar’s tests with Holm correction were applied to each artifact label. When comparing raw EEG with our SSDLabeler, a significant difference was observed in the Clean, Eye, and Channel categories (p < 0.05), indicating that our SSDLabeler outperformed raw EEG in these specific artifact types as seen in Fig. 5a. Comparisons between raw EEG and the previous work’s SSD revealed significant differences in the Clean, Eye, Channel, and Other categories (p < 0.05), as seen in Fig. 5b. The previous work’s SSD outperformed raw EEG in the Clean, Eye, and Channel categories, whereas raw EEG outperformed the previous work’s SSD in the Other category. Finally, comparing the previous work’s SSD with our SSDLabeler revealed significant differences in the Eye and Other categories (p < 0.05), with our SSDLabeler showing better performance in both cases, as seen in Fig. 5c.

Fig. 5.

McNemar’s test results for per-label classification performance across the three training datasets on the noise session test data. The top row (a) compares Raw EEG with SSDLabeler, the middle row (b) compares Raw EEG with the Previous Work’s SSD, and the bottom row (c) compares the Previous Work’s SSD with SSDLabeler. In all panels, columns correspond to artifact categories in the following order: Clean, Eye, Muscle, Channel, and Other.

Ablation study

The ablations were conducted for both the ME and noise session test data. For the ME test data, our SSDLabeler (A+B+C) was compared against ablation 1 (A+B) and ablation 2 (A+C). Our SSDLabeler achieved the best overall accuracy (0.826) in comparison to ablation 1 (0.793) and ablation 2 (0.694). Across the evaluated categories, SSDLabeler achieved the highest accuracy in the Eye (0.813) and Muscle (0.845) categories, while showing equivalent performance to the ablation models in the Channel (0.833) and Other (0.803) categories as seen in Table 3. To evaluate whether these ablations reflected significant changes in error patterns, McNemar’s tests with Holm correction were applied to each artifact label. When comparing our SSDLabeler with ablation 1, a significant difference was observed for the Muscle category only (p < 0.05), while no significant difference were observed in the Clean, Eye, Channel, and Other categories, as seen in Fig. 6a. Although ablation 1 achieved a slightly higher accuracy in the Clean category, this difference was not statistically significant. When comparing our SSDLabeler and ablation 2, significant differences were revealed in the Clean, Eye, and Muscle categories (p < 0.05), while no significant differences were observed in the Channel and Other categories. Our SSDLabeler outperformed ablation 2 in all the mentioned categories. The complete set of McNemar’s test results is visualized in Fig. 6b. Finally, in the comparisons between ablation 1 and ablation 2, significant differences were observed in the Clean, Eye, and Muscle categories (p < 0.05), while no significant differences were observed in the Channel and Other categories, as seen in Fig. 6c. Ablation 1 outperformed ablation 2, hinting that RMS and PSD thresholding may hold more importance than recombining all artifacts.

Table 3.

Ablation classification accuracy by artifact category in the ME session (human-labeled test set).

Method	Clean	Eye	Muscle	Channel	Other	Average
Model trained on SSDLabeler: A+B+C	0.835	0.813	0.845	0.833	0.803	0.826
Model trained on Ablation 1: A+B	0.843	0.793	0.695	0.833	0.803	0.793
Model trained on Ablation 2: A+C	0.663	0.670	0.500	0.833	0.803	0.694

Fig. 6.

McNemar’s test results for per-label classification performance across the ablation datasets on the motor execution session test data. The top row (a) compares SSDLabeler with Ablation 1, the middle row (b) compares SSDLabeler with Ablation 2, and the bottom row (c) compares Ablation 1 with Ablation 2. In all panels, columns correspond to artifact categories in the following order: Clean, Eye, Muscle, Channel, and Other.

For the instructed-noise test data, our SSDLabeler (A+B+C) was compared against ablation 1 (A+B) and ablation 2 (A+C). Ablation 1 achieved the highest overall accuracy (0.784), followed closely by ablation 2 (0.781) and our SSDLabeler (0.751). Category-wise performance across the five artifact classes is summarized in Table 4. Although SSDLabeler achieved lower accuracies than ablation 1 in the Clean, Eye, and Muscle categories, while showing equivalent performance in the Channel and Other categories, no statistically significant differences were observed between SSDLabeler and ablation 1. When comparing SSDLabeler with ablation 2, SSDLabeler achieved lower accuracy in the Clean and Muscle categories, higher accuracy in the Eye category, and equivalent performance in the Channel and Other categories. However, no statistically significant differences were observed between the models. Finally, when comparing ablation 1 and ablation 2, ablation 1 achieved higher accuracy in the Eye category, whereas ablation 2 achieved higher accuracy in the Clean and Muscle categories. Equivalent performance was observed in the Channel and Other categories, with no statistically significant differences detected overall. The complete set of McNemar’s test results is shown in Fig. 7.

Table 4.

Ablation classification accuracy by artifact category in the noise session (human-labeled test set).

Method	Clean	Eye	Muscle	Channel	Other	Average
Model trained on SSDLabeler: A+B+C	0.953	0.859	0.494	0.965	0.482	0.751
Model trained on Ablation 1: A+B	0.988	0.871	0.612	0.965	0.482	0.784
Model trained on Ablation 2: A+C	1.0	0.835	0.624	0.965	0.482	0.781

Fig. 7.

McNemar’s test results for per-label classification performance across the ablation datasets on the noise session test data. The top row (a) compares SSDLabeler with Ablation 1, the middle row (b) compares SSDLabeler with Ablation 2, and the bottom row (c) compares Ablation 1 with Ablation 2. In all panels, columns correspond to artifact categories in the following order: Clean, Eye, Muscle, Channel, and Other.

Discussion

This study introduced a multi-label artifact classification framework supported by a novel ICA-based SSD generation method designed to overcome the persistent shortage of artifact-labeled EEG data. A major challenge in EEG preprocessing is that artifacts rarely occur in isolation; instead, they frequently co-occur, overlap in time, and exhibit nonstationary dynamics that complicate both manual inspection and automated removal. Existing ICA-based SSD approaches address these challenges only partially, as they typically inject a single artifact type into clean EEG¹⁹, producing simplified examples that do not adequately represent the complex contamination patterns present in real EEG. By contrast, the SSDLabeler method reintroduces multiple ICA-isolated artifact components simultaneously, preserving their inherent spatiotemporal structure and producing a richer, more ecologically valid representation of artifact mixtures. Crucially, the proposed SSDLabeler framework also differs fundamentally from training directly on raw EEG in terms of label quality and interpretability. While raw EEG preserves all recorded signal content, it also contains heterogeneous, poorly characterized noise sources whose origins, structures, and relevance are often unknown. In contrast, our SSDLabeler explicitly reconstructs EEG using verified brain and artifact ICs, while excluding low-probability and ambiguous components that are likely to reflect unlabeled or non-interpretable noise. This controlled reconstruction preserves task-relevant neural activity and well-defined artifact structure, while substantially reducing uncertainty in the associated labels. As a result, the proposed SSDLabeler provides training data with higher label confidence and a more constrained learning signal, allowing the classifier to focus on meaningful brain and artifact features rather than adapting to uncontrolled or unlabeled noise variability. This distinction explains why SSD-based training yields more reliable and generalizable artifact classification than training directly on raw EEG, despite the latter containing the full signal mixture.

The results demonstrate that this design choice is a primary driver of the observed performance gains. Models trained on the proposed SSDLabeler consistently outperformed those trained on raw EEG and on the previous SSD baseline across both the ME and instructed noise sessions. Importantly, these improvements were not confined to a single artifact category, but instead manifested as balanced gains across most labels. This indicates that exposing the classifier to realistic multi-artifact combinations during training enables it to learn more generalizable representations of contamination, rather than overfitting to isolated artifact signatures. The ablation analysis further supports this interpretation by showing that recombining multiple artifact categories within the same epoch accounts for the majority of the performance improvement, confirming that modeling artifact co-occurrence is critical for effective multi-label classification. In addition to multi-artifact recombination, the proposed framework incorporates RMS- and PSD-based verification to ensure that only component epochs containing verifiable artifact activity are used for SSD generation. As ICA-derived artifact components do not necessarily express artifact activity uniformly across time, RMS- and PSD-based verification ensures that only epochs with confidently expressed contamination are used for reconstruction, reducing label noise introduced by ICLabel classification alone. By excluding ambiguous or low-confidence epochs that may be mislabeled by ICLabel, this verification step improves the realism and reliability of the synthetic data and reduces label noise during training. The use of RMS- and PSD-based verification is grounded in established EEG preprocessing practice. Transient, high-amplitude artifacts such as ocular activity, line noise, and channel artifacts are commonly characterized by increases in signal energy, which RMS effectively captures^11,29. In contrast, muscle artifacts are known to exhibit elevated high-frequency spectral power, making PSD-based criteria well suited for their detection^18,30. By integrating these signal-level criteria with ICLabel-based spatial classification, the proposed framework leverages complementary physiological characteristics of artifacts, improving label reliability beyond IC classification alone. Although the ablation results show that artifact verification contributed more than multi-artifact recombination, the highest and most consistent performance is achieved when both components are combined. In the ME session, multi-artifact recombination plays a crucial complementary role by exposing the classifier to realistic co-occurring artifact patterns, enabling improved classification of naturalistic EEG.

An important contributor to this robustness is the hyperparameter optimization strategy employed in this study. All SSDLabeler hyperparameters were determined using an independent external dataset prior to evaluation on the primary dataset, decoupling parameter selection from the test distribution. Unlike prior SSD-based approaches that adopt fixed parameter settings, this strategy reduces tuning bias and improves external validity. As a result, the observed gains reflect stable properties of artifact structure rather than adaptation to a specific dataset, supporting the practical applicability of the proposed framework across recording systems and experimental paradigms. In contrast, many existing SSD-based methods adopt fixed or heuristically chosen parameters optimized on the same dataset used for evaluation, which can limit robustness across datasets.

The instructed noise session provided complementary validation under controlled conditions, in which participants intentionally generated specific artifact behaviors. On this dataset, the SSDLabeler-trained classifier achieved the highest overall accuracy, particularly for the eye artifact. Notably, reduced performance in muscle artifact classification was observed only in the instructed noise session and not in the ME session. This pattern reflects a well-known challenge in EEG preprocessing. Muscle contamination exhibits high inter- and intra-subject variability, especially under deliberately exaggerated movements, and is less consistently isolated by ICA^11,12,14,30. In addition, it is important to note that all data were resampled to 125 Hz to ensure consistency across datasets. While this sampling rate is sufficient to capture the primary frequency content of most EEG artifacts, it may reduce sensitivity to very high-frequency components associated with muscle activity. However, muscle artifacts are not exclusively characterized by high-frequency content alone; they also exhibit distinct spatial distributions and temporal dynamics, which remain preserved at this sampling rate. The observed performance differences in muscle classification are therefore more likely attributable to the inherent variability of muscle artifacts and the structure of the instructed-noise paradigm, rather than the resampling procedure itself. Interestingly, the ablation analysis revealed that both ablation variants achieved higher overall classification accuracy than the proposed SSDLabeler in the noise session, with ablation 1 showing the strongest performance. Notably, SSDLabeler (A+B+C) underperformed compared to the ablation models in several categories, particularly Muscle, where the largest performance gap was observed. This suggests that combining all verification criteria may introduce stricter or interacting constraints that reduce sensitivity to related artifacts. In contrast, simpler configurations (A+B or A+C) may better preserve salient muscle features, resulting in improved classification performance under these conditions. This result can be attributed to the structural characteristics of the instructed-noise data. Unlike real-world EEG, which contains highly variable, frequently co-occurring, and unconscious artifact actions, the noise session consists of temporally isolated epochs dominated by a single, strong, well-defined, and conscious performance of the artifact type. Under these simplified conditions, SSD variants that assume single-artifact contamination per epoch, particularly when combined with RMS- and PSD-based verification, are naturally better aligned with the data distribution and can achieve higher classification accuracy. In contrast, the proposed SSDLabeler is explicitly designed to model realistic EEG contamination, where artifact strength and quality vary, and multiple artifact sources may co-occur within the same segment. While this design improves generalization to naturalistic EEG, it may introduce additional variability that is not present in the instructed-noise paradigm. Consequently, the superior performance of the ablation models in the noise session does not indicate a limitation of the proposed SSDLabeler, but rather highlights the potential mismatch between simplified evaluation conditions and the real-world structure of artifacts. These findings reinforce the importance of evaluating artifact handling methods under both controlled and naturalistic conditions and underscore the value of multi-artifact modeling for general-purpose EEG preprocessing.

Despite these strengths, several limitations remain. While this study focused on multi-label artifact classification, artifacts such as heart and line noise were not observed in the present dataset, which limits evaluation for these categories. The limited availability of labeled examples may also impact reliability for these categories. As a downstream task, future work should investigate how SSD-augmented preprocessing influences other analyses, including source localization, connectivity estimation, and neural decoding performance in BCI settings. Extending the framework to time-resolved segmentation may further improve temporal precision and enable real-time artifact-aware preprocessing.

In addition, the evaluation was conducted using a fixed split with a relatively small independent test set (n = 5 subjects). This design was motivated by the requirement for manual annotation to establish high-quality ground truth labels, which is inherently time-intensive. While this approach ensures reliable and unbiased evaluation, it limits the number of subjects that can be included and may not fully capture inter-subject variability across larger populations. Future work will focus on expanding the number of annotated subjects and exploring scalable or semi-automated annotation strategies to enable broader validation.

Finally, another potential limitation of the proposed framework is the computational cost associated with ICA, particularly in the context of real-time applications. However, in this study, ICA is used exclusively during the offline SSD generation stage to create artifact-labeled training data. This preprocessing step is performed prior to model deployment and does not impact real-time inference. Once trained, the artifact classification model can be applied directly to incoming EEG data without requiring ICA or semi-synthetic generation. Thus, the proposed approach introduces additional computational cost during training but maintains efficiency during deployment, representing a trade-off between improved classification performance and offline processing time.

Overall, the findings demonstrate that explicitly modeling artifact co-occurrence through multi-artifact SSD generation is a key factor in improving classification robustness and generalization. By combining ICA-derived artifact structure with verification-based selection and realistic recombination, the proposed framework advances beyond prior SSD approaches and establishes a scalable foundation for transparent, artifact-aware EEG preprocessing pipelines.

Methods

Dataset and preprocessing

We used the publicly available EEG dataset introduced by Cho et al., which contains recordings from ME, motor imagery (MI), rest, and an instructed noise session²⁸. EEG was acquired from 52 participants using a 64-channel BioSemi ActiveTwo system at a sampling rate of 512 Hz. For the present study, only the ME sessions were used to construct the training, validation, and test sets for artifact classification. The ME paradigm provides continuous EEG recordings with spontaneous and co-occurring artifacts, making it well-suited for evaluating multi-label artifact detection under realistic conditions. Data from 42 subjects were used for training, 5 for validation, and 5 for an independent test set. The dataset was divided into fixed training, validation, and independent test sets rather than using cross-validation. This design choice was motivated by the requirement to manually annotate EEG data to establish ground-truth labels. Manual annotation is a time-intensive process requiring expert knowledge, which constrained the number of subjects that could be feasibly annotated within this study. Therefore, a subset of subjects was selected for validation and independent testing to ensure high-quality annotations while maintaining a fully independent evaluation protocol. The validation and test sets consisted exclusively of raw and cleaned EEG data. The validation set was used for model selection without manual annotation, while the independent test set was manually annotated to provide a high-quality ground truth for unbiased evaluation. The dataset allocation, preprocessing steps, and hyperparameter optimization strategy are summarized in Fig. 8.

Fig. 8.

Overview of datasets and processing pipeline used in this study. The BCI2000 motor imagery dataset (top, blue) was used exclusively for external hyperparameter optimization of the proposed SSDLabeler framework, using both ME and MI recordings from three subjects³¹. This design avoids dataset-specific tuning and better validates generalization. The Cho et al. dataset (bottom, gray) was used for SSD generation and artifact classification²⁸. ME and instructed-noise sessions from 42 subjects were used to generate SSD-based training data, while 10 held-out subjects were reserved exclusively for validation and testing. Only the test set was manually annotated to provide ground truth for final evaluation. All datasets underwent identical preprocessing prior to SSD generation and classification, ensuring a fair and externally valid evaluation.

Each ME session consisted of 20 trials per hand. At the start of each trial, subjects viewed a black screen followed by a fixation cross presented for two seconds. This was followed by a randomly ordered instruction cue (“left hand” or “right hand”) presented for three seconds, during which participants executed the cued motor movement. The three-second movement period was extracted for all analyses in this study.

In addition to the task-related sessions, the Cho et al. dataset also includes a dedicated noise session designed to capture instructed artifact behaviors. Five types of non-task-related activities were recorded: eye blinking, vertical eye movements (up/down), horizontal eye movements (left/right), head movements, and jaw clenching. For each artifact type, participants performed the instructed movement twice for 5 seconds. These noise recordings contain clearly observable, high-amplitude artifacts, providing a complementary dataset for evaluating artifact classification under controlled conditions and benchmarking the behavior of models trained on semi-synthetic contamination.

Preprocessing followed a modified version of the pipeline described by Zhang et al.¹⁸. Raw EEG was band-pass filtered between 1 and 50 Hz using a zero-phase finite impulse response (FIR) filter of order 6600 implemented with the EEGLAB filter_data function²⁹. Power-line interference at 60 Hz was suppressed using the filter_notch function. After filtering, continuous data were detrended to remove slow drifts and segmented into three-second epochs aligned to cue onset. Finally, the data were downsampled from 512 Hz to 125 Hz, following common EEG preprocessing practice and ensuring compatibility with downstream model architectures. To ensure consistency across datasets, all recordings were resampled to a common sampling rate of 125 Hz, as the dataset used for hyperparameter optimization was originally sampled at 160 Hz. This unified sampling rate avoids introducing dataset-specific variability and maintains a consistent preprocessing pipeline across all experiments.

To optimize the hyperparameters of the proposed SSDLabeler framework, a separate and independent EEG dataset was used. This is intended to demonstrate that the proposed method has dataset-independent generalizability and external validity. The dataset used was derived from publicly available BCI2000 motor imagery recordings which contain EEG data from 109 subjects³¹. From this corpus, data from three subjects were selected for hyperparameter optimization. While sharing a similar 64-channel EEG montage and a comparable experimental paradigm, this dataset differed from the Cho et al. dataset in terms of data acquisition setup, experimental environment, and task composition, as both the ME and MI recordings were included. For each selected subject, four recording files were used, comprising two ME and two MI sessions, each containing 15 trials, resulting in a total of 12 files used for hyperparameter optimization. MI data were included to validate the generalizability of the optimized hyperparameters, ensuring that SSDLabeler parameters are not overly specialized to a single task condition. EEG preprocessing followed the same pipeline used for the Cho et al. dataset to ensure methodological consistency and avoid introducing dataset-specific preprocessing or tuning biases. Because the BCI2000 motor imagery dataset does not include artifact annotations, recordings from the public datasets were manually annotated and used as ground truth for validating hyperparameter configurations. This data were used exclusively for hyperparameter selection and were not included in the final training, validation, or test sets.

All EEG datasets used in this study were obtained from previously published and publicly available sources^28,31. In the original studies, all participants provided informed consent, and consent was obtained from legal guardians where applicable, in accordance with the respective institutional ethical guidelines. The present work used only anonymized recordings from these public datasets, and no new human data were collected. Manual annotation was performed by the authors with expertise in EEG analysis. The annotators, as contributing authors, consented to their involvement in the study.

ICA and artifact IC selection

Independent Component Analysis (ICA) was applied to the preprocessed EEG to separate the multichannel signals into temporally independent and spatially fixed components. ICA is widely used for isolating neural sources from ocular, muscular, cardiac, and environmental artifacts by exploiting statistical independence^12,13,15. Prior to decomposition, all epochs were concatenated across trials to form a continuous matrix suitable for ICA. Specifically, the epoched data of size , with N trials, C=64 channels, and samples, were reshaped into a matrix of size , ensuring that ICA operated on the full dataset rather than on individual segments. Principal component analysis (PCA) whitening was then performed. FastICA was used for source decomposition to remain consistent with the preprocessing pipeline of EEGDenoiseNet, which our method builds upon^16,18. Although ICLabel was trained on Infomax-derived components, it relies on features such as scalp topography and spectral properties, which are expected to remain qualitatively similar across ICA methods when source separation is successful¹⁴.

Artifact-related components were identified using ICLabel, an automated classifier trained to assign probabilistic labels to ICA components¹⁴. ICLabel outputs probabilities across seven categories: brain, eye, muscle, heart, line noise, channel noise, and other. Components with an artifact probability (p) \ge 0.6 for any non-brain category were retained as artifact ICs. Components with a dominant brain label were kept for reconstructing clean EEG, while components with low ICLabel confidence was excluded to avoid introducing ambiguous or unlabeled components into the reconstructed EEG. Reconstruction of EEG signals from all the ICs can be performed using the ICA mixing matrix, defined as

1

where denotes the flattened EEG signal, is the ICA mixing matrix, and is the matrix of source signals corresponding to K independent components. Let denote the index set of the subset of ICs, with . Reconstruction from a subset of ICs is performed by selecting the corresponding columns and rows of the ICA matrices,

2

where and . This formulation is used to reconstruct both clean EEG (from brain-labeled ICs) and class-specific artifact signals (from artifact-labeled ICs) throughout Algorithms 1–5. This procedure produced a structured library of artifact ICs, ocular, muscular, cardiac, line, channel noise, and other noise, that served as the basis for semi-synthetic EEG generation, while ensuring that neural activity reconstructed from the brain-labeled ICs remained as clean as possible as defined in Algorithm 1.

Algorithm 1.

ICLabel component labeling (A)

SSD labeler for artifact annotation

Although ICLabel provides probabilistic classifications of independent components, some IC-derived artifact epochs may still be ambiguous or may not contain meaningful artifact activity. To ensure that only segments exhibiting clear, artifact-verified contamination were used for SSD generation, we developed SSDLabeler, a tool that integrates ICLabel outputs with RMS- and PSD-based thresholding criteria for multi-label data annotation as shown in Fig. 9. SSDLabeler performs the secondary screening based on RMS and PSD features, followed by multi-artifact reconstruction () and multi-label output () seen in algorithm 2.

Fig. 9.

Preprocessing and annotation pipeline for semi-synthetic EEG generation using SSDLabeler. Raw EEG was first preprocessed with bandpass (1–50 Hz) and notch filtering (60 Hz), then epoched into three-second segments. ICA and ICLabel were then applied to decompose the signals and assign artifact probabilities to each IC. Artifact ICs with ICLabel scores \ge 0.6 (e.g., eye, muscle, heart, line, channel, or other) were reconstructed and segmented into time windows (A). RMS and PSD thresholding were subsequently applied to ensure the presence of true contamination within each segment (B). The region highlighted in red corresponds to the conventional SSD procedure used in prior work, including preprocessing, ICA decomposition, ICLabel-based classification, and IC reconstruction. The region highlighted in blue represents the novel components introduced in the proposed SSDLabeler framework, specifically the RMS- and PSD-based verification and the generation of multi-label SSD annotations.

RMS thresholding is applied to capture transient, high-amplitude non-neural events–including ocular activity, cardiac activity, line noise, channel noise, and other abrupt contaminations by evaluating signal energy relative to percentile-based thresholds computed within each IC. In contrast, PSD thresholding is used to identify muscle-related artifacts, which typically exhibit elevated high-frequency spectral power. By detecting increases in power within a muscle-specific frequency range, PSD verification filters out IC segments with weak or absent muscle contamination.

By combining data-driven ICLabel probabilities with adaptive RMS- and PSD-based verification, SSDLabeler ensures that only reliably contaminated IC segments are selected for reinjection into clean EEG. This additional screening step reduces false positives from ICLabel alone and yields artifact annotations that more accurately reflect the temporal and spectral characteristics of real EEG contamination, thereby supporting the generation of a higher-quality and more realistic semi-synthetic EEG dataset.

Algorithm 2.

SSDLabeler: multi-artifact semi-synthetic EEG generation (A+B+C).

RMS thresholding

The RMS value provides a measure of the overall signal energy within an epoch. In Eq. (3), denotes the epoched referenced signal, where N is the number of epochs, C is the number of channels, and is the number of time samples per epoch. For the n-th epoch, the RMS is computed by first calculating the channel-wise RMS across time and then averaging across channels:

3

where denotes the EEG amplitude of epoch n, channel c, and time sample t. To quantify artifact strength, a standardized effect score was computed (Eq. 4) that compares the RMS of artifact-labeled epochs, , with that of clean EEG, , normalized by the standard deviation of the clean RMS distribution, :

4

The mean and standard deviation were computed from clean EEG epochs within each session after reconstructing the brain ICs. All available clean epochs per session were used to estimate these statistics, with the number of epochs varying by dataset (e.g., n=20 in the ME dataset). As RMS is computed over many time samples within each epoch and normalization is performed on a per-subject basis, this provides a stable baseline for standardization. Thresholds for each artifact category were defined by taking the percentile (p) of all scores (Eq. 5), enabling the selection of the strongest artifact instances during RMS screening. The percentile thresholds used were were determined through hyperparameter optimization as follows: Eye (0.1), Heart (0.99), Line (0.99), Channel (0.99), and Other (0.99).

5

Eye artifacts produced higher RMS values due to their large-amplitude, low-frequency characteristics, resulting in a lower optimal threshold. In contrast, non-muscle artifact categories exhibited less consistent or lower-amplitude signals, leading to higher optimal thresholds to reduce false positives. Epochs whose RMS scores fell below their respective category threshold were excluded from SSD generation, ensuring that only IC segments containing clear artifact activity were reinjected into clean EEG (Fig.10; Algorithm 3). RMS thresholding was particularly effective for identifying transient, high-amplitude non-neural events; consequently, all artifact types except muscle were screened using this method.

Fig. 10.

RMS thresholding procedure used for artifact verification. For all artifact types except muscle, IC-derived signals were segmented into epochs and the RMS was computed for each segment. Epochs with RMS values exceeding the predefined threshold were labeled as artifact-present (1), whereas those below the threshold were labeled as clean (0). The example plot illustrates , where segments above the red dashed line are identified as eye-related artifacts.

Algorithm 3.

RMS-based artifact verification (B1)

PSD thresholding

PSD thresholding was used to label muscle artifacts within the epochs, as they are typically represented by large bursts in power in the high-frequency range. Following Zhang et al., we compared the PSD of each candidate muscle IC epoch against an exponential reference template representing clean EEG spectral characteristics using correlation analysis¹⁸.

The exponential reference template was introduced by Zhang et al. as a preprocessing criterion to identify segments that deviate from the characteristic spectral structure of clean EEG activity. In this study, we adapted this approach for artifact verification, where low similarity to the clean EEG template was interpreted as evidence of muscle contamination rather than as a rejection criterion. The PSD template was originally defined over the frequency range of 1–120 Hz, whereas in this study, we restrict the frequency range to Hz to match the bandwidth of our preprocessed EEG signals and to focus on the most relevant frequency components for our dataset. The clean PSD template was therefore defined as

6

For each epoch, the PSD was computed using Welch’s method and averaged across channels before correlation with the clean PSD template. Unlike Zhang et al., who rejected low-correlation segments, we retained these segments as verified muscle artifacts as defined in Algorithm 4. Epochs with correlation łe 0.8 relative to the clean template were labeled as muscle artifact (1), while those above were discarded as seen in Fig. 11.

Fig. 11.

PSD thresholding procedure for muscle artifact annotation. The PSD of muscle component segments and correlation values were computed. Segments with correlation łe 0.8 relative to the clean PSD template were classified as containing muscle artifacts (1), while those above the threshold were labeled as clean (0). A simulated muscle spectrum () was used as an example to show elevated high-frequency activity characteristic of muscle contamination.

Algorithm 4.

PSD verification for muscle artifacts (B2)

SSDLabeler hyperparameter optimization

To determine the optimal hyperparameters for SSDLabeler, we performed a structured hyperparameter search that jointly evaluated the ICLabel probability threshold and the class-specific RMS and PSD verification thresholds. All SSDLabeler hyperparameters, including ICLabel probability thresholds and RMS- and PSD-based verification criteria, were optimized exclusively using the BCI2000 motor imagery dataset and were fixed prior to evaluation on the Cho et al. dataset^28,31. The objective was to identify a set of hyperparameters that maximized artifact annotation accuracy while maintaining consistent performance across all datasets. As ICLabel assigns each independent component (IC) a probability distribution over seven classes, the confidence required for accepting an IC as an artifact can influence which components contribute to the reconstructed artifact signals¹⁴. We therefore evaluated a set of ICLabel probability thresholds,

7

such that an IC was assigned to its ICLabel-predicted artifact class only if its winning probability exceeded p. ICs whose winning probability was below the threshold were excluded. For each value of p, this procedure yielded a different IC-to-class mapping and a corresponding reconstructed artifact signal for each artifact category.

After the IC assignment, the per-epoch RMS-based standardized scores were computed for non-muscle artifacts and the PSD-correlation values for muscle artifacts (see previous subsections). To avoid evaluating the full combinatorial space of all parameters, we performed a one-parameter-at-a-time sweep for each artifact class while reusing the same cached feature matrices.

For the five artifact categories screened with RMS (eye, heart, line noise, channel noise, other), we evaluated using percentile scores to define the threshold,

8

and selected the value that maximized the per-class accuracy when comparing the predicted labels against the ground-truth annotations for that category. For muscle artifacts, which were evaluated using PSD correlation, we swept the correlation thresholds

9

and selected the value that produced the highest muscle artifact annotation accuracy. For each ICLabel probability threshold p, the sweeps above produced a set of optimal parameters,

10

To determine the final configuration, we computed the macro-accuracy (Eq. 11), defined as the mean classification accuracy across all seven artifact classes, for each tested value of p. Ground-truth labels used for this evaluation were obtained through manual annotation. The ICLabel probability threshold that produced the highest macro-accuracy across recordings was selected as the final operating point for SSDLabeler, together with its corresponding set of class-specific hyperparameters. This data-driven optimization procedure ensured a principled selection of parameters that balanced artifact sensitivity and specificity while maintaining robustness across subjects and recording sessions. Following hyperparameter optimization on the external BCI2000 dataset, the final SSDLabeler configuration used throughout this study was fixed according to Eq. 10. The ICLabel probability threshold was set to . RMS-based verification thresholds were defined as (10th percentile) and , , , and were all set to the 99th percentile. Muscle artifact verification was performed using PSD-based correlation thresholding with . These hyperparameters were held constant across all experiments and were not tuned on the primary Cho et al. evaluation dataset.

11

Semi-synthetic data generation (SSDLabeler)

We developed a SSD generation framework, termed SSDLabeler, to produce realistic multi-label artifact–contaminated EEG, defined in Algorithm 2. Similar to the previous single-artifact SSD approach (Algorithm 6), ICA and ICLabel were used to classify independent components into artifact categories based on their probabilistic labels. To maintain consistent conditions across methods, the ICLabel probability threshold was set to \ge 0.6 for both our SSDLabeler and the previous work’s SSD implementation. The previous approach originally used a 0.8 threshold, which we adjusted to match our optimized hyperparameter settings and ensure a fair comparison.

The key difference between prior SSD methods and the SSDLabeler framework lies in the handling of artifact components. Previous SSD approaches inject only a single artifact category into clean EEG, producing isolated contaminations that do not reflect the co-occurrence structure of real-world artifacts. In contrast, SSDLabeler incorporates an additional verification step: only artifact IC epochs that pass the RMS- or PSD-based thresholding criteria are selected for reinjection. This ensures that each IC segment contains verifiable ocular, muscular, cardiac, line, channel, or other artifact activity.

Beyond differences with prior SSD approaches, SSDLabeler also differs fundamentally from training directly on raw EEG. To enable a fair comparison, raw EEG is not modified but is instead annotated using the same ICLabel- and RMS/PSD-based verification criteria, as formalized in Algorithm 7. Raw EEG preserves all recorded signal content, including low-probability or ambiguous components that may reflect unknown or unlabeled noise sources. In contrast, SSDLabeler reconstructs EEG using only ICs with confident ICLabel assignments and verified artifact presence, explicitly excluding components with uncertain labels. As a result, the generated SSD preserves meaningful brain and artifact features while reducing label uncertainty introduced by uncontrolled noise sources.

The verified artifact ICs are then mixed back into clean EEG to create multi-artifact contaminated signals, while preserving both the underlying neural activity and the spatiotemporal characteristics of the artifact components. This process generates realistic multi-label contamination that more closely resembles the complexity and overlapping structure of raw EEG recordings. The final output for the SSD-based training dataset consisted of three components: (i) SSD-generated artifact-contaminated epochs, (ii) clean EEG epochs reconstructed from brain ICs, and (iii) original raw EEG epochs. In contrast, the raw EEG baseline contained only the original raw EEG epochs and clean EEG, without controlled artifact reconstruction.

Algorithm 5.

Multi-artifact SSD generation (C)

Algorithm 6.

Single-artifact SSD generation (previous work baseline¹⁹)

Algorithm 7.

Raw EEG annotation using SSDLabeler criteria

SSDLabeler ablation study

To isolate and validate the contributions of the proposed SSDLabeler framework, we conducted an ablation study based on the previous ICA-based SSD generation method. The SSDLabeler framework introduces two novel additions to this baseline: RMS- and PSD-based artifact thresholding and the recombination of multiple artifact categories within the same epoch. Together with ICLabel-based component selection, these elements define three components of the SSD generation pipeline: (A) ICLabel-based artifact component selection (Algorithm 1), (B) RMS- and PSD-based artifact thresholding (Algorithm 3, 4), and (C) multi-artifact recombination (Algorithm 5).

Using these components, two ablation conditions were constructed. The first ablation (A+B) evaluated the effect of RMS and PSD verification without multi-artifact recombination seen in Algorithm 8. In this setting, artifact IC epochs were screened using SSDLabeler’s RMS- and PSD-based thresholds, but only a single artifact category was injected into the clean EEG, consistent with prior SSD approaches. Artifact labels for this ablation were derived using the SSDLabeler annotation procedure, reflecting the use of explicit artifact verification. The second ablation (A+C) evaluated the effect of multi-artifact recombination without RMS and PSD verification seen in Algorithm 9.

Both ablation SSDs were used as training datasets for multi-label artifact classification in the same manner as the proposed SSDLabeler and the previous work’s SSD. This design enabled a controlled assessment of how artifact verification and multi-artifact recombination independently contribute to classification performance.

Algorithm 8.

Ablation 1: (A+B)

Algorithm 9.

Ablation 2: (A+C)

Model

To evaluate the utility of the SSD, we trained a deep neural network for multi-label artifact classification. The classifier architecture was adapted from the EEG encoder branch of the PredANN model, originally developed for music classification using EEG and audio through contrastive learning³². For the present study, the original classification module was used and modified with a multi-label prediction head, enabling the model to assign multiple artifact labels to each EEG segment. In this streamlined configuration, the model focuses exclusively on identifying artifact-contaminated EEG based on its spatiotemporal structure.

Model architecture and losses

The adapted model consists of a two-dimensional convolutional neural network (2D CNN) designed to extract hierarchical representations from multichannel EEG. Inputs were formatted as matrices of size C × T, where C = 64 EEG channels and T = 375 time samples correspond to three-second epochs sampled at 125 Hz. The encoder is composed of multiple convolutional blocks with kernel sizes selected to capture spatial dependencies across channels as well as temporal dynamics within each trial. Each block contains convolutional layers followed by batch normalization, nonlinear activation functions, and dropout for regularization. The resulting feature maps are flattened and passed to a fully connected classification head with sigmoid activation, allowing independent probability estimates for each artifact category (eye, muscle, heart, line, channel, and clean), although heart and line categories were not evaluated due to the absence of corresponding instances in the dataset.

The model was trained using the binary cross-entropy loss function, consistent with multi-label classification. Labels for each input segment were generated using the artifact validation procedure associated with the corresponding dataset, such that each training set was annotated according to its own artifact-contamination method. Optimization was performed using the Adam optimizer with an initial learning rate of and a batch size of 48. Training proceeded for 5000 epochs to ensure full convergence and stable multi-label predictions.

Conclusion

This study introduced a novel SSD generation framework to address a long-standing limitation in EEG research: the lack of high-quality, artifact-labeled data that captures the complexity, diversity, and co-occurrence structure of real-world contamination. By leveraging ICA-derived artifact components together with RMS- and PSD-based thresholding, the proposed method ensures that only verified artifact-containing epochs are incorporated into the SSD. Crucially, unlike prior approaches that inject a single artifact type into clean EEG, our method reintroduces multiple independent artifact components concurrently, preserving their spatial projections, temporal morphology, and natural overlap. This multi-artifact design produces a realistic and ecologically valid representation of contamination, allowing the synthetic data to more closely reflect the realistic properties of raw EEG.

The effectiveness of this SSD framework was demonstrated through a multi-label artifact classification task evaluated on both real-world ME recordings and a controlled, instructed noise session with intentionally produced artifacts. Across both settings, the classifier trained on our SSDLabeler achieved superior overall accuracy compared with models trained on raw EEG or on the previous SSD baseline. These improvements were observed across nearly all artifact categories, indicating that two complementary factors contribute to improved generalization. First, the realistic variability and co-occurrence structure introduced by multi-artifact SSD generation expose the classifier to contamination patterns that closely resemble those encountered in raw EEG. Second, the RMS- and PSD-based verification steps improve label fidelity by ensuring that only epochs containing verifiable artifact activity are included in training.

Beyond achieving high classification accuracy, the value of the proposed SSDLabeler framework extends to broader methodological implications for EEG preprocessing. Realistic multi-artifact synthetic data enables automated and reproducible artifact identification, reducing dependence on subjective manual inspection and enhancing consistency across studies. Improved artifact classification further benefits preprocessing pipelines by supporting pre-ICA quality screening, preserving ICA stability, and facilitating selective cleaning in analyses such as ERP or PSD. These capabilities reflect a shift toward principled, scalable, and automated preprocessing strategies suited for large datasets, diverse populations, and real-time applications where manual annotation is impractical. More broadly, this work demonstrates that SSD constructed from ICA-isolated artifact components, combined with RMS- and PSD-based verification and multi-artifact recombination, provides a stronger and more reliable training signal for artifact-aware machine learning models than raw EEG alone. By accurately modeling the mixed and overlapping structure of EEG contamination, the proposed SSDLabeler provides a foundation for training classifiers that generalize across subjects, experimental paradigms, and recording systems. As EEG datasets continue to expand in scale and complexity, such automated augmentation strategies will play an increasingly important role in ensuring robust and reproducible downstream analyses. Looking ahead, extending the current architecture with a temporal segmentation head capable of generating continuous, pointwise predictions represents an important direction for future development. Such an approach would allow finer temporal tracking of artifact dynamics, improve real-time applicability in BCI systems, and enable integration with more advanced neural decoding.

In summary, the proposed SSDLabeler framework offers a robust and scalable solution for generating multi-artifact EEG by leveraging ICA-derived artifact structure and signal-level verification, enabling high-performance multi-label artifact classification and improving the reliability of automated preprocessing. By more accurately reproducing the complexity of real-world EEG contamination, this work contributes a significant step toward building fully automated, artifact-aware pipelines suitable for neuroscience, clinical diagnostics, and brain–computer interface applications.

Author contributions

T.A. conceptualized and designed the framework, method, and experiments. A.C. also designed the method and experiments. A.C., S.M., and T.A. implemented the code and conducted the experiment. A.C., T.A., S.M., and N.P. analyzed and interpreted the results. A.C. and T.A. wrote the main manuscript and prepared tables and figures. A.C. and S.M. organized the code. T.A., A.C., and N.P. drafted and edited the manuscript. All authors reviewed and approved the manuscript. N.P. supervised and coordinated the research project.

Funding

The authors received no specific funding for this work.

Code availability

Upon acceptance, the source code used to conduct the experiments will be publicly available at https://github.com/AkimaCon/SSDLabeler.

Data availability

The EEG data used in this study were obtained from publicly available datasets. All participants in the original studies provided informed consent, as reported in the respective original publications. The motor execution dataset described by Cho et al.²⁸ is available through the GigaDB repository at https://gigadb.org/dataset/100295. The BCI2000-based EEG Motor Movement/Imagery dataset³¹, used for external hyperparameter validation, was obtained from the PhysioNet repository (https://physionet.org/content/eegmmidb/1.0.0/). No new human or animal data were collected in this study.

Declarations

Competing interests

The authors declare no competing interests.