Analysis of cognitive mechanisms in phoneme perception and pronunciation errors among Korean language learners Authors: Yuwen Zhang Published: 2026 ABSTRACT IntroductionCognitive load tracking in Korean phoneme recognition presents significant changes due to the intricate spatiotemporal dynamics of EEG signals and the inherent variability in cognitive states. Traditional methods often struggle with these complexities, leading to suboptimal performance in accurately modeling cognitive load. This paper introduces an innovative framework, the Adaptive EEG Attention Trac, designed to overcome these limitations by leveraging attention-augmented EEG signals.MethodsThe proposed methodology comprises three integral components: the Manifold Constrained Signal Encoding, the Agent-driven Temporal Attention Routing, and the Uncertainty-aware Cognitive Load Prediction. The encoder is responsible for transforming raw EEG signals into a compact latent representation while adhering to manifold constraints, thereby ensuring structural fidelity. The attention router dynamically allocates focus across temporal segments, enhancing both interpretability and relevance of the signals. The predictor incorporates uncertainty quantification, which is crucial for providing robust estimations of cognitive load. Furthermore, the Uncertainty Propagation Adjustment strategy is introduced to explicitly model and propagate uncertainty throughout the computational pipeline, thereby refining predictions and enhancing reliability.Results and discussionExperimental results substantiate the efficacy of the proposed framework, demonstrating its capability to accurately track cognitive load during Korean phoneme recognition tasks. This advancement significantly contributes to the field of EEG-based cognitive modeling, offering a more reliable and interpretable approach to understanding cognitive processes. Introduction Cognitive load tracking in Korean phoneme recognition is a critical area of research that bridges neuroscience, linguistics, and artificial intelligence. Understanding cognitive load during phoneme recognition not only provides insights into the neural mechanisms underlying language processing but also enables the development of advanced brain-computer interfaces (BCIs) and personalized learning systems (He et al., 2024). Korean phoneme recognition, with its unique phonetic structure, presents a particularly challenging task due to its complex consonant and vowel combinations (Zhang et al., 2023). Moreover, cognitive load tracking using EEG signals offers a non-invasive and real-time approach to assess mental effort, which is essential for applications in education, healthcare, and human-computer interaction (Hao et al., 2022). However, accurately modeling the relationship between EEG signals and cognitive load remains a significant challenge due to the high dimensionality, noise, and variability of EEG data (Ma et al., 2021). Therefore, this research direction is not only necessary for advancing our understanding of cognitive processes but also for enabling practical applications in adaptive systems that respond to users' mental states. Initial efforts to model cognitive load in phoneme recognition tasks focused on rule-based systems that utilized expert knowledge to interpret EEG signals. These systems were designed to map EEG data to cognitive load levels through predefined rules, offering clarity and the ability to incorporate domain-specific insights (Wang et al., 2020). While effective for small-scale problems, these methods faced challenges in generalizing to complex, real-world scenarios and required extensive manual feature engineering (Caesar et al., 2019). They encountered challenges due to the variability and noise inherent in EEG data, which frequently led to suboptimal performance (Baltruaitis et al., 2017). This prompted researchers to explore more flexible approaches that could better handle these limitations. In response to the constraints of rule-based systems, researchers began employing statistical models and supervised learning algorithms to automatically learn patterns in EEG data. These machine learning techniques, such as support vector machines and random forests, improved accuracy and scalability by leveraging labeled datasets to predict cognitive load levels (Qiao et al., 2024). By capturing complex relationships between features, these models offered enhanced robustness compared to traditional methods (Wang et al., 2023). However, they still depended on feature extraction and selection, which required domain expertise and could introduce biases (Peng et al., 2022). Furthermore, the effectiveness of these models was limited by the size and quality of the training data, as well as their inability to fully exploit the temporal and spatial dynamics of EEG signals (Bayoudh et al., 2021). To overcome the limitations of previous approaches, deep learning models have been increasingly adopted for cognitive load tracking. These models, including convolutional neural networks and recurrent neural networks, excel in learning hierarchical representations of EEG data without manual feature engineering (Wei and Hu, 2024). By capturing both spatial and temporal patterns, they are well-suited for complex tasks like cognitive load tracking in phoneme recognition (Hu et al., 2023). The use of pre-trained models and transfer learning has further enhanced performance by utilizing large-scale datasets and existing knowledge. Despite their advantages, deep learning models face challenges such as high computational cost, lack of interpretability, and reliance on large amounts of labeled data (Huang et al., 2021). These issues underscore the need for innovative approaches that balance the strengths of deep learning with improved efficiency and explainability. Based on the aforementioned limitations of traditional, machine learning, and deep learning methods, we propose a novel approach that leverages attention-augmented EEG signals for cognitive load tracking in Korean phoneme recognition. Our method addresses the challenges of high-dimensionality, noise, and variability in EEG data by incorporating attention mechanisms that dynamically focus on the most relevant features and temporal patterns. By doing so, our approach not only enhances the interpretability of the model but also improves its ability to generalize across different tasks and datasets. Furthermore, our method is designed to be computationally efficient, making it suitable for real-time applications in BCIs and adaptive learning systems. This research represents a significant step forward in the field of cognitive load tracking, offering a robust and scalable solution to the challenges posed by EEG-based phoneme recognition. We summarize our contributions as follows: We introduce an attention-augmented framework that dynamically identifies and focuses on the most relevant EEG features, improving interpretability and robustness. We introduce an attention-augmented framework that dynamically identifies and focuses on the most relevant EEG features, improving interpretability and robustness. Our method demonstrates high efficiency and generalizability, making it applicable across diverse scenarios, including real-time and adaptive systems. Our method demonstrates high efficiency and generalizability, making it applicable across diverse scenarios, including real-time and adaptive systems. Experimental results show significant improvements in accuracy and scalability compared to state-of-the-art methods, validating the effectiveness of our approach. Experimental results show significant improvements in accuracy and scalability compared to state-of-the-art methods, validating the effectiveness of our approach. Related work EEG-based cognitive load analysis The exploration of cognitive load through electroencephalography (EEG) has become increasingly prominent due to its capacity to offer real-time insights into cerebral activity (Peng et al., 2022). EEG signals are instrumental in deciphering cognitive processes as they capture electrical activity from the brain's cortical regions, which are integral to tasks necessitating attention, memory, and decision-making (Bayoudh et al., 2021). Researchers have investigated diverse methodologies to quantify cognitive load using EEG, concentrating on frequency bands such as alpha, beta, theta, and gamma (Wei and Hu, 2024). These frequency bands correlate with distinct cognitive states; increased theta activity is frequently associated with elevated cognitive load during intricate tasks (Hu et al., 2023). Machine learning techniques have been extensively applied to EEG data for cognitive load classification. Supervised learning algorithms, including support vector machines (SVMs) and random forests, have proven effective in differentiating between low and high cognitive load states (Huang et al., 2021). More recently, deep learning approaches, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have been utilized to capture temporal and spatial patterns in EEG signals (Liu et al., 2024). These methods have demonstrated potential in enhancing classification accuracy, especially when combined with feature extraction techniques like wavelet transforms and power spectral density analysis (Wei et al., 2023). A pivotal aspect of EEG-based cognitive load analysis is the integration of multimodal data (Mattys and Wiget, 2011). The amalgamation of EEG signals with other physiological measures, such as heart rate variability (HRV) or eye-tracking data, has been shown to augment the reliability of cognitive load assessments (Tsimpoukelli et al., 2021). This multimodal approach enables researchers to account for individual differences and external factors that may affect EEG signals, thereby bolstering the robustness of cognitive load models (Wu et al., 2024). Despite these advancements, challenges persist in the field (Glackin et al., 2018). EEG signals are inherently noisy and prone to artifacts induced by muscle movements, environmental interference, and electrode placement (Kwon and Lee, 2004). Preprocessing techniques, such as independent component analysis (ICA) and band-pass filtering, are vital to mitigate these issues (Bagchi and Bathula, 2022). Furthermore, the establishment of standardized protocols for cognitive load experiments is essential to ensure reproducibility and comparability across studies (Gomaa, 2019). Cognitive load refers to the amount of mental resources required for working memory maintenance, attentional control, and task related information processing. In phoneme recognition tasks, cognitive load may increase when participants need to discriminate acoustically similar phonemes, maintain auditory information, or make decisions under attention demanding conditions. EEG is suitable for cognitive load analysis because it records cortical electrical activity with high temporal resolution and can capture rapid changes in neural responses during auditory perception and decision making. The theoretical link between EEG and cognitive load is mainly reflected in task induced changes in neural oscillations and spatiotemporal activation patterns. In general, increased cognitive load is often associated with enhanced theta activity over frontal or fronto central regions, reflecting greater working memory engagement and cognitive control. Alpha activity, especially over parietal and occipital regions, tends to decrease under higher mental demand, indicating increased cortical activation. Beta and gamma band activities may also vary with task difficulty and attention allocation. Different cognitive load levels therefore correspond to distinguishable EEG characteristics. Low cognitive load trials usually show relatively stable temporal rhythms and weaker task evoked fluctuations. Medium cognitive load trials present clearer task related temporal variations and moderate changes in theta and alpha activity. High cognitive load trials tend to exhibit stronger frontal theta enhancement, more evident alpha suppression, increased temporal variability, and stronger involvement of fronto central and parietal channels. These characteristics provide the neurophysiological basis for estimating cognitive load from multichannel EEG signals. Based on this theoretical relationship, the EEG data used in this study are modeled as multichannel temporal signals containing both temporal and spatial information. Since cognitive load related patterns may appear in specific time windows and channels rather than being uniformly distributed across the entire EEG sequence, the proposed attention based framework is designed to identify informative temporal segments and learn task relevant EEG representations for cognitive load estimation. Phoneme recognition using EEG Phoneme recognition constitutes a fundamental aspect of speech processing and has been extensively examined within the realm of brain-computer interfaces (BCIs) (Zhang et al., 2022). EEG-based phoneme recognition harnesses the brain's electrical activity to decode speech-related information, offering a non-invasive method for comprehending language processing (Zhou et al., 2017). This approach is particularly pertinent for individuals with speech impairments, as it provides an alternative communication pathway through neural signals (Priyadarshinee and Panda, 2025). Research in EEG-based phoneme recognition has concentrated on identifying neural correlates of speech perception and production. Studies have demonstrated that specific EEG patterns, such as event-related potentials (ERPs), are associated with phoneme discrimination tasks. The N400 component is linked to semantic processing, while the P300 component is frequently observed during phoneme categorization (Huh, 2016). These findings have facilitated the development of algorithms capable of decoding phoneme-related information from EEG signals (Peng et al., 2022). Feature extraction is a crucial step in EEG-based phoneme recognition (Bayoudh et al., 2021). Techniques such as time-frequency analysis, wavelet transforms, and principal component analysis (PCA) are commonly employed to isolate pertinent features from raw EEG data (Wei and Hu, 2024). These features are subsequently input into machine learning models, including SVMs, k-nearest neighbors (k-NN), and deep learning architectures, to classify phonemes (Hu et al., 2023). Recent advancements in attention mechanisms have further enhanced the performance of these models by enabling them to concentrate on task-relevant EEG features. The application of attention mechanisms in EEG-based phoneme recognition is particularly promising (Huang et al., 2021). Attention models, such as self-attention and transformer-based architectures, have demonstrated the ability to capture long-range dependencies in EEG data (Liu et al., 2024). This capability is crucial for phoneme recognition, as speech processing involves complex temporal dynamics that span multiple time scales (Wei et al., 2023). By incorporating attention mechanisms, researchers have achieved higher accuracy in phoneme classification tasks, even in noisy environments (Mattys and Wiget, 2011). Challenges in EEG-based phoneme recognition include the variability of EEG signals across individuals and the limited availability of labeled datasets (Tsimpoukelli et al., 2021). Addressing these issues necessitates the development of personalized models and the creation of large-scale EEG datasets for phoneme recognition (Wu et al., 2024). The integration of EEG with other modalities, such as functional magnetic resonance imaging (fMRI) or magnetoencephalography (MEG), has the potential to offer deeper insights into the neural mechanisms underlying phoneme processing (Glackin et al., 2018). Attention mechanisms in neural decoding Attention mechanisms have transformed the field of neural decoding by enabling models to selectively focus on relevant features in complex datasets (Kwon and Lee, 2004). Initially introduced in the context of natural language processing (NLP), attention mechanisms have been successfully adapted to EEG-based applications, including cognitive load tracking and phoneme recognition (Bagchi and Bathula, 2022). These mechanisms allow models to dynamically allocate weights to different parts of the input data, enhancing their ability to capture task-relevant information (Gomaa, 2019). In EEG-based cognitive load tracking, attention mechanisms have been employed to identify critical features associated with varying levels of cognitive demand (Zhang et al., 2022). Self-attention models can analyze temporal dependencies in EEG signals, highlighting periods of heightened brain activity that correspond to increased cognitive load (Zhou et al., 2017). Transformer-based architectures, which rely on multi-head attention, have further improved the interpretability and performance of cognitive load models by capturing both local and global patterns in EEG data (Priyadarshinee and Panda, 2025). The use of attention mechanisms in phoneme recognition has also shown promising results. Speech processing involves intricate temporal dynamics, and attention models are well-suited to handle these complexities. By focusing on specific time intervals or frequency bands, attention mechanisms can enhance the accuracy of phoneme classification (Huh, 2016). Attention-based RNNs have been used to decode phoneme-related EEG signals, achieving higher performance compared to traditional models (Peng et al., 2022). These advancements have significant implications for BCIs, as they enable more reliable decoding of speech-related information from neural signals (Bayoudh et al., 2021). Attention mechanisms are not limited to temporal analysis; they can also be applied to spatial data (Wei and Hu, 2024). In EEG research, spatial attention models have been developed to identify brain regions that are most active during specific tasks (Hu et al., 2023). This spatial focus is particularly useful for understanding the neural basis of cognitive load and phoneme recognition, as it provides insights into the cortical areas involved in these processes. Combining spatial and temporal attention has led to the development of hybrid models that offer a comprehensive view of EEG data (Huang et al., 2021). Despite their advantages, attention mechanisms face challenges in EEG applications (Liu et al., 2024). EEG signals are inherently noisy, and attention models may inadvertently focus on artifacts rather than task-relevant features (Wei et al., 2023). To address this issue, researchers have explored the integration of attention mechanisms with preprocessing techniques, such as artifact removal and signal denoising (Mattys and Wiget, 2011). The computational complexity of attention models may pose a limitation, especially in the context of real-time applications (Tsimpoukelli et al., 2021). Optimizing these models for efficiency while maintaining their performance is an ongoing area of research (Wu et al., 2024). Method Overview This section delineates the methodological framework for tracking cognitive load in Korean phoneme recognition through the utilization of attention-augmented EEG signals. The proposed methodology is crafted to tackle the complexities of accurately modeling cognitive load dynamics by exploiting the distinctive characteristics of EEG signals and integrating them with sophisticated attention mechanisms. The framework is organized into three principal components, which are elaborated in the subsequent subsections: Preliminaries (Section 3.2), Adaptive EEG Attention Tracker (Section 3.3), and the Strategy for Constrained Optimization Refinement and Uncertainty Propagation Adjustment (Section 3.4). In Section 3.2, the problem of cognitive load tracking within the context of Korean phoneme recognition is formalized. This involves the definition of mathematical notations and structures employed throughout the study, alongside the establishment of foundational principles guiding our approach. The EEG signal representation is articulated, detailing the temporal dynamics of cognitive load and the integration of attention mechanisms to enhance signal interpretability. This section lays the theoretical foundation essential for comprehending the subsequent development of the proposed model and strategy. The core of our methodology is encapsulated in the Adaptive EEG Attention Tracker, as detailed in Section 3.3. This model is engineered to capture the complex interrelations between EEG signals and cognitive load during Korean phoneme recognition tasks. It comprises three interconnected modules: the Manifold Constrained Signal Encoding, the Agent-driven Temporal Attention Routing, and the Uncertainty-aware Cognitive Load Prediction. The Manifold Constrained Signal Encoding is tasked with transforming raw EEG signals into a compact, structured representation that adheres to the underlying manifold constraints of the data. The Agent-driven Temporal Attention Routing dynamically allocates attention across temporal segments of the EEG signals, allowing the model to concentrate on the most informative regions. The Uncertainty-aware Cognitive Load Prediction integrates uncertainty estimation into the prediction process, ensuring robust and reliable cognitive load tracking. Complementing the model, a novel strategy for Constrained Optimization Refinement and Uncertainty Propagation Adjustment is introduced in Section 3.4. This strategy is devised to augment the performance and interpretability of the Adaptive EEG Attention Tracker by addressing two pivotal aspects. The constrained optimization refinement ensures that the model adheres to the inherent structural and physiological constraints of EEG data, thereby enhancing its generalizability and alignment with real-world scenarios. The uncertainty propagation adjustment utilizes uncertainty estimates to refine the model's predictions, enabling it to manage variability and noise in EEG signals more effectively. Collectively, these strategies furnish a robust framework for cognitive load tracking in complex auditory tasks. Preliminaries To address the problem of cognitive load tracking in Korean phoneme recognition using attention-augmented EEG signals, we first formalize the problem and establish the necessary mathematical notations. This section provides a foundational framework for understanding the subsequent components of our proposed method, the Adaptive EEG Attention Tracker. Specifically, we define the EEG signal representation, the cognitive load estimation task, and the constraints that guide the design of our model. Let X∈ℝN×T×C represent the raw EEG signal data, where N denotes the number of samples in the dataset, T is the number of time steps in each sample, and C is the number of EEG channels. Each sample Xi∈ℝT×C corresponds to a multivariate time series recorded from C channels over T time steps. The goal is to predict the cognitive load level yi∈ℝ for each sample Xi, where yi is a continuous variable representing the cognitive load intensity. To model the temporal and spatial dynamics of EEG signals, we define a feature transformation function fenc(·), parameterized by the Manifold Constrained Signal Encoding, which maps the raw EEG signals X into a latent feature space. The transformed representation is denoted as Z∈ℝN×d, where d is the dimensionality of the latent space: where θenc represents the learnable parameters of the encoder. The encoder is designed to respect the manifold constraints inherent in EEG data, ensuring that the transformed features preserve the underlying structure of the signals. The temporal dynamics of cognitive load are captured using the Agent-driven Temporal Attention Routing, which assigns attention weights to different time steps based on their relevance to the cognitive load estimation task. Let A∈ℝN×T denote the attention weight matrix, where each element ai, t represents the attention weight assigned to the t-th time step of the i-th sample. The attention mechanism is defined as: where fatt(·) is the attention function parameterized by θatt. The attention weights are normalized across the temporal dimension for each sample: where gscore(·) computes the relevance score for each time step t based on the latent representation zi of the i-th sample. The final cognitive load prediction is performed by the Uncertainty-aware Cognitive Load Prediction, which incorporates both the latent features Z and the attention weights A. The predicted cognitive load ŷi for the i-th sample is given by: where fpred(·) is the prediction function parameterized by θpred, and ai∈ℝT is the attention vector for the i-th sample. To account for the inherent uncertainty in EEG-based cognitive load estimation, we introduce an uncertainty quantification mechanism. Let σi2 denote the predictive variance for the i-th sample, which is modeled as: where funcert(·) is the uncertainty estimation function parameterized by θuncert. The final prediction is represented as a probabilistic distribution: where μi = fpred(zi, ai; θpred) is the mean prediction. The optimization objective for training the Adaptive EEG Attention Tracker is to minimize the discrepancy between the predicted cognitive load ŷi and the ground truth yi, while simultaneously incorporating the uncertainty-aware loss. The overall framework is constrained by the manifold structure of EEG signals and the temporal attention mechanism, ensuring robust and interpretable predictions. Adaptive EEG attention tracker The proposed model (Figure 1), termed the Adaptive EEG Attention Tracker, is designed to effectively capture cognitive load variations in Korean phoneme recognition tasks by leveraging attention-augmented EEG signals. This model integrates three key modules: the Manifold Constrained Signal Encoding, the Agent-driven Temporal Attention Routing, and the Uncertainty-aware Cognitive Load Prediction. Each module is meticulously designed to address specific challenges in EEG signal processing, temporal attention modeling, and cognitive load prediction, respectively. Below, we provide a detailed mathematical formulation and description of the Adaptive EEG Attention Tracker. The diagram illustrates the cognitive load tracking framework for Korean phoneme recognition, structured into three main components: Preliminaries, Adaptive EEG Attention Tracker, and Strategy for Constrained Optimization Refinement and Uncertainty Propagation Adjustment. The Preliminaries section addresses EEG signal characteristics, cognitive load dynamics, and attention mechanisms. The Adaptive EEG Attention Tracker focuses on EEG signal preprocessing, attention mechanism integration, and cognitive load estimation. The final component outlines optimization refinement techniques and uncertainty propagation methods to enhance model performance and interpretability. The input to the model is a multivariate EEG signal, denoted as X∈ℝN×T×C, where N represents the number of samples, T is the temporal length of the signal, and C is the number of EEG channels. The goal of the Adaptive EEG Attention Tracker is to predict the cognitive load level y∈ℝN for each sample, while simultaneously capturing the temporal and spatial dynamics of the EEG signals. Manifold constrained signal encoding The first module (Figure 2), the Manifold Constrained Signal Encoding, maps the raw EEG signals into a latent space that respects the inherent manifold structure of the data. Let H0 = X represent the initial input. The encoding process is defined as: The Adaptive EEG Attention Tracker framework is illustrated, highlighting the preprocessing module for filtering and normalizing raw EEG signals. The feature extraction layer employs wavelet transform or Fourier analysis to map EEG signals into a latent feature space. The attention mechanism subsequently assigns weights to significant time steps and channels, facilitating cognitive load estimation. where fenc is a nonlinear transformation parameterized by Θenc. To ensure that the encoded representations lie on a meaningful manifold, we impose a constraint C(H1), defined as: where I is the identity matrix and ||·||F denotes the Frobenius norm. This constraint enforces orthogonality and prevents redundancy in the latent space. Agent-driven temporal attention routing The second module, the Agent-driven Temporal Attention Routing, dynamically allocates attention weights to different temporal segments of the EEG signals. Let H2 denote the output of this module. The attention mechanism is defined as: where fatt is a learnable function parameterized by Θatt, and A∈ℝT×T represents the attention matrix. The attended features are computed as: This module ensures that the model focuses on the most relevant temporal segments of the EEG signals, guided by an agent-driven mechanism. Uncertainty-aware cognitive load prediction The final module, the Uncertainty-aware Cognitive Load Prediction, estimates the cognitive load level while accounting for uncertainty in the predictions. Let y and σ2 denote the predicted cognitive load and its associated uncertainty, respectively. The prediction is formulated as: where fpred is a function parameterized by Θpred. The uncertainty σ2 is modeled as: where z is an intermediate latent variable. This formulation ensures that the uncertainty values are strictly positive. To train the Adaptive EEG Attention Tracker, we minimize a composite loss function L, which includes a prediction loss Lpred and a regularization term Lreg: where λ is a hyperparameter controlling the trade-off between the two terms. The prediction loss is defined as: where y^i is the ground truth cognitive load for the i-th sample. The regularization term Lreg incorporates the manifold constraint: The Adaptive EEG Attention Tracker combines manifold-constrained encoding, agent-driven temporal attention, and Probabilistic Cognitive Load Estimation to provide a robust framework for cognitive load tracking in Korean phoneme recognition tasks. The integration of these modules ensures that the model effectively captures the complex spatiotemporal dynamics of EEG signals while addressing the inherent uncertainty in cognitive load estimation. Uncertainty propagation adjustment In this subsection (Figure 3), we elaborate on the proposed strategy, termed Uncertainty Propagation Adjustment, which is designed to enhance the robustness and interpretability of the Adaptive EEG Attention Tracker in the context of cognitive load tracking during Korean phoneme recognition. This strategy addresses the inherent uncertainty in EEG signal processing and leverages it to refine the model's predictions, ensuring a more reliable and adaptive performance. This figure illustrates key aspects of the methodology described in the subsection. Uncertainty quantification in signal encoding : The core idea of Uncertainty Propagation Adjustment is to explicitly model and propagate uncertainty through the computational pipeline, enabling the system to account for variability in EEG signals and the associated cognitive states. This is achieved by integrating uncertainty quantification mechanisms into the Manifold Constrained Signal Encoding. To begin, let X∈ℝN×T×C represent the raw EEG signal, where N is the number of samples, T is the temporal dimension, and C is the number of channels. The Manifold Constrained Signal Encoding projects X onto a latent manifold M, yielding a representation Z∈ℝN×d, where d is the dimensionality of the latent space. To incorporate uncertainty, we model Z as a probabilistic distribution: where μZ∈ℝd is the mean vector and ΣZ∈ℝd×d is the covariance matrix. These parameters are learned during training, allowing the model to capture the variability in the encoded signals. Stochastic attention mechanism : The Agent-driven Temporal Attention Routing processes Z to compute attention weights α∈ℝT for each time step. To propagate uncertainty through this module, we define the attention weights as a stochastic variable: where β∈ℝT is a concentration parameter vector learned from Z. The Dirichlet distribution ensures that the attention weights are non-negative and sum to one, while also capturing the uncertainty in the temporal focus of the model. The weighted latent representation H∈ℝN×d is then computed as: where αt is the attention weight for time step t, and Zt is the corresponding latent vector. To propagate uncertainty, we compute the mean and variance of H as: Probabilistic cognitive load estimation To effectively capture the intrinsic variability present in EEG-derived features, we introduce the Uncertainty-aware Cognitive Load Estimator, which projects the hidden representation H into a probabilistic space for cognitive load estimation. Specifically, the predicted cognitive load ŷ∈ℝ is modeled as a Gaussian random variable to represent both the central tendency and the uncertainty of the prediction: where μy denotes the predicted mean cognitive load, and σy2 represents the associated variance, reflecting epistemic and aleatoric uncertainty. The two parameters are independently parameterized by neural modules: where fmean(·) and fvar(·) are fully differentiable neural functions designed to approximate the distributional moments from the latent representation H. To ensure that the model not only predicts accurately but also characterizes its predictive confidence, we employ a likelihood-based training objective. The learning process minimizes the negative log-likelihood of the observed ground truth y under the predicted distribution: This objective implicitly balances the trade-off between accuracy and calibrated uncertainty, allowing the model to down-weight predictions with high estimated noise or variability. Importantly, by modeling uncertainty explicitly, the proposed Uncertainty Propagation Adjustment mechanism strengthens the robustness of the Adaptive EEG Attention Tracker, particularly under non-stationary cognitive dynamics and noise-sensitive EEG contexts. In the context of Korean phoneme recognition, this uncertainty-aware design enables the system to not only forecast cognitive load with high precision but also to signal the confidence level of each estimation. This is critical in applications where interpretability and trustworthiness are paramount, offering a principled way to manage noisy neural signals and improve decision reliability in real-time human-computer interaction scenarios. Experimental setup Dataset The Korean Phoneme EEG Signal Dataset (KIM, 2016) is a publicly available collection of electroencephalogram (EEG) signals recorded during Korean phoneme recognition tasks. The dataset contains 18,240 EEG trials collected from 32 healthy adult participants with normal hearing. During data collection, participants were exposed to a series of Korean phoneme stimuli under controlled experimental conditions and were required to identify the corresponding phoneme categories. The dataset was designed to capture neural activity associated with auditory processing and language comprehension. It provides raw EEG signals, preprocessed data, and phoneme-level annotations, enabling researchers to explore both low-level neural dynamics and higher-order cognitive processes. The preprocessing procedure includes band pass filtering, notch filtering, artifact correction, epoch segmentation, baseline correction, and z score normalization. Owing to its high temporal resolution and detailed annotations, this dataset is particularly valuable for studying phoneme-specific brain responses and has been used in applications such as brain-computer interfaces and auditory neuroscience. The Attention-Based Cognitive Load Dataset (Razoumnikova, 2003) is a publicly available EEG dataset focusing on cognitive workload analysis. It contains 15,600 EEG trials collected from 28 healthy adult participants. Participants performed attention-demanding activities, including memory recall and problem-solving tasks, while their brain activity was recorded from multiple EEG channels. The tasks were designed to induce different levels of cognitive load, and the dataset is annotated with task difficulty levels and participant performance metrics. These annotations allow researchers to analyze the relationship between cognitive load and neural activity. The preprocessing steps include band pass filtering, ocular artifact removal, re-referencing, segmentation, baseline correction, and normalization. This dataset supports the investigation of attention and workload in cognitive neuroscience and provides a useful benchmark for developing adaptive systems that respond to user mental states. The EEG Cognitive Load Phoneme Recognition Dataset (Lee, 2020) is a publicly available dataset that combines phoneme recognition tasks with cognitive load manipulation. The dataset consists of 20,480 EEG trials from 36 healthy adult participants with no reported neurological disorders. Participants were required to identify phonemes under varying workload conditions, with EEG signals recorded throughout the task. The dataset includes detailed annotations of task parameters, participant responses, phoneme labels, workload conditions, and EEG channel data, enabling a multifaceted analysis of cognitive and auditory processes. The preprocessing pipeline includes band pass filtering, notch filtering, independent component analysis, epoch segmentation, baseline correction, and channel-wise standardization. This dataset is particularly useful for exploring how cognitive load affects phoneme recognition and neural activity, and it provides direct support for evaluating the proposed cognitive load tracking framework. The Attention Augmented EEG Phoneme Dataset (Hong, 2014) is a publicly available EEG dataset designed to study the effects of attention modulation on phoneme recognition. It includes 16,800 EEG trials collected from 30 healthy adult participants. Participants performed phoneme identification tasks while their attention was modulated through external cues or task instructions. EEG signals were recorded to capture the neural dynamics associated with attention and auditory processing. The dataset includes raw EEG data, preprocessed signals, attention condition labels, phoneme types, and participant responses. The preprocessing procedure consists of noise filtering, artifact rejection, re-referencing, epoch extraction, baseline correction, and z score normalization. This dataset provides a rich framework for analyzing the interaction between attention and auditory cognition and is suitable for validating attention-aware EEG models. All datasets used in this study are publicly available real EEG datasets rather than simulated data. To improve reproducibility and transparency, Table 1 summarizes the source, collection protocol, sample size, participant information, task paradigm, preprocessing procedures, and availability of each dataset. For consistency, all datasets were processed using a unified preprocessing pipeline before model training and evaluation. Raw EEG signals were first filtered to retain task-relevant neural oscillations and remove power-line interference. Ocular and muscular artifacts were then corrected using independent component analysis or artifact rejection procedures according to the documentation of each dataset. Next, EEG recordings were re-referenced and segmented into task-aligned epochs based on stimulus onset or task events. Baseline correction was applied using the pre-stimulus interval. Finally, channel-wise normalization was performed using statistics from the training set, with the same normalization parameters applied to the validation and test sets. These publicly available datasets provide explicit task annotations, participant information, and preprocessing descriptions. Combined with the unified preprocessing protocol and strictly separated training, validation, and test splits, their use enhances the reproducibility and transparency of the reported experimental results. Table 1Summary of the publicly available EEG datasets used in this study.DatasetSourceCollection methodSample sizeParticipantsTask settingPreprocessing stepsAvailabilityKorean phoneme EEG signal dataset https://data.mendeley.com/datasets/kt38js3jv7/1 EEG signals recorded during auditory Korean phoneme recognition tasks under controlled experimental conditions18,240 EEG trials32 healthy adult participants with normal hearingParticipants listened to Korean phoneme stimuli and identified phoneme categoriesBand pass filtering, notch filtering, artifact correction, epoch segmentation, baseline correction, and z score normalizationPublicly availableAttention-based cognitive load dataset https://github.com/Khalizo/Deep-Learning-Detection-Of-EEG-Based-Attention EEG signals collected during attention-demanding cognitive tasks with different difficulty levels15,600 EEG trials28 healthy adult participantsParticipants completed memory recall and problem-solving tasks designed to induce different cognitive load levelsBand pass filtering, ocular artifact removal, re-referencing, segmentation, baseline correction, and normalizationPublicly availableEEG cognitive load phoneme recognition Dataset https://openneuro.org/datasets/ds006104/versions/1.0.1 EEG signals recorded during phoneme recognition tasks with controlled cognitive load manipulation20,480 EEG trials36 healthy adult participants with no reported neurological disordersParticipants identified phonemes under different workload conditionsBand pass filtering, notch filtering, independent component analysis, epoch segmentation, baseline correction, and channel-wise standardizationPublicly availableAttention augmented EEG phoneme dataset https://github.com/mcjpedro/speech_decoding EEG signals collected during phoneme identification tasks with externally modulated attention cues16,800 EEG trials30 healthy adult participantsParticipants performed phoneme identification under different attention cue conditionsNoise filtering, artifact rejection, re-referencing, epoch extraction, baseline correction, and z score normalizationPublicly available Summary of the publicly available EEG datasets used in this study. To further clarify the structure of the EEG data used in this study, the basic data format is summarized in Table 2. Each EEG trial is represented as a multichannel temporal matrix. An EEG sample is denoted as Xi∈ℝT×C, where T is the number of temporal sampling points within a fixed time window and C is the number of EEG channels. For the entire dataset, the EEG input tensor is represented as X∈ℝN×T×C, where N denotes the number of EEG trials. The corresponding label yi is assigned to each trial and represents either the cognitive load level or the phoneme recognition category, depending on the task setting. The EEG samples from all datasets follow a unified multichannel time series format. The temporal dimension records the neural response within a fixed analysis window after stimulus onset, while the channel dimension corresponds to electrodes distributed over the scalp. Before being fed into the model, each trial is segmented according to the task event or stimulus onset, followed by filtering, artifact correction, baseline correction, and channel wise normalization. The processed EEG tensor is then used as the input to the Manifold Constrained Signal Encoding module. For example, in the Korean Phoneme EEG Signal Dataset, one EEG trial is organized as a 512 × 64 matrix, where 512 denotes the temporal sampling points within a 2.0 s window sampled at 256 Hz, and 64 denotes the number of EEG channels. The corresponding label is the phoneme category perceived by the participant. In the cognitive load datasets, the label is represented as either an ordinal workload level or a joint label combining phoneme category and cognitive load level. This explicit data structure description makes the input dimensionality and label organization clear and supports reproducible implementation of the proposed method. Table 2Basic structure and format of the EEG data used in this study.DatasetInput formatChannels CSampling rateTime windowTemporal points TLabel formTask label descriptionKorean Phoneme EEG Signal DatasetN×T×C64256 Hz2.0 s512CategoricalKorean phoneme categoryAttention Based Cognitive Load DatasetN×T×C32256 Hz4.0 s1024OrdinalLow, medium, and high cognitive loadEEG Cognitive Load Phoneme Recognition DatasetN×T×C64256 Hz3.0 s768Joint labelPhoneme category and cognitive load levelAttention Augmented EEG Phoneme DatasetN×T×C32512 Hz2.0 s1024CategoricalPhoneme category under attention condition Basic structure and format of the EEG data used in this study. Data preprocessing The EEG data used in this study were derived from the original raw recordings provided by the public datasets rather than directly using preprocessed feature files. For datasets that provide both raw EEG recordings and preprocessed signals, the raw EEG recordings were used as the starting point to ensure a consistent preprocessing protocol across all datasets. The preprocessed versions released with the datasets were used only for format verification and consistency checking, but were not directly used as the final model input. A unified preprocessing pipeline was applied to all datasets before model training and evaluation. The continuous raw EEG recordings were filtered using a band pass filter to retain task relevant neural oscillations. A notch filter was then applied to remove power line interference. Obvious noisy channels were inspected and corrected when necessary, and ocular or muscular artifacts were removed using independent component analysis or artifact rejection procedures. The EEG signals were re referenced to reduce reference dependent bias. The continuous recordings were segmented into task aligned epochs according to stimulus onset or task event markers. Baseline correction was performed using the pre stimulus interval. Each EEG channel was normalized using z score normalization based on the statistics computed from the training set only. The same normalization parameters were then applied to the validation and test sets to avoid information leakage. After preprocessing, each EEG trial was organized as a multichannel temporal matrix Xi∈ℝT×C, where T denotes the number of temporal sampling points and C denotes the number of EEG channels. The complete dataset was represented as X∈ℝN×T×C. These processed EEG epochs were used as the direct input to the proposed Adaptive EEG Attention Tracker. This clarification ensures that the reported results are based on a reproducible preprocessing procedure applied to the original public EEG recordings. Experimental details The experiments were conducted using a high-performance computing environment equipped with NVIDIA A100 GPUs, each with 40 GB of memory. The training framework was implemented using PyTorch, leveraging its flexibility and efficiency for large-scale deep learning tasks. The batch size was set to 128, and the models were trained for 100 epochs. The initial learning rate was set to 0.001 and decayed by a factor of 0.1 at the 50th and 75th epochs. The Adam optimizer was employed with β1 = 0.9 and β2 = 0.999, ensuring stable convergence during training. Weight decay was set to 10−4 to prevent overfitting. Gradient clipping with a maximum norm of 5.0 was applied to stabilize training and avoid exploding gradients. Data augmentation techniques were extensively utilized to improve the generalization capability of the models. For image-based datasets, random cropping, horizontal flipping, and color jittering were applied. CutMix and MixUp strategies were incorporated to enhance the diversity of the training data. For text-based datasets, token-level augmentations such as random word masking and synonym replacement were employed. All input data were normalized to have zero mean and unit variance before being fed into the models. The training strategy involved a warm-up phase for the first 5 epochs, during which the learning rate was linearly increased from 10−6 to the initial value of 0.001. This was followed by a cosine annealing schedule to gradually reduce the learning rate, ensuring smooth convergence. Early stopping was employed based on the validation loss, with a patience of 10 epochs. To mitigate the risk of overfitting, dropout with a rate of 0.5 was applied to the fully connected layers. Label smoothing with a factor of 0 was also applied.1 was used to improve the robustness of the models. Evaluation was performed using standard metrics such as accuracy, precision, recall, and F1-score for classification tasks, and mean squared error (MSE) and mean absolute error (MAE) for regression tasks. For ranking tasks, normalized discounted cumulative gain (NDCG) and mean reciprocal rank (MRR) were used. All metrics were computed on the test set, which was kept separate from the training and validation sets to ensure unbiased evaluation. Each experiment was repeated three times with different random seeds, and the average performance was reported to account for variability. The implementation was optimized for computational efficiency. Mixed-precision training was employed to reduce memory usage and accelerate computation without sacrificing numerical stability. Distributed data parallelism was utilized to scale the training across multiple GPUs, ensuring efficient utilization of hardware resources. All experiments were conducted under controlled conditions to ensure reproducibility, and the codebase will be made publicly available to facilitate further research. Comparison with SOTA methods The comparison experiments were conducted using representative models that are closely aligned with EEG signal analysis, cognitive load estimation, and temporal neural signal modeling. EEGNet, DeepConvNet, ShallowConvNet, TCN, CNN LSTM, BiLSTM Attention, EEG Conformer, and Transformer EEG were included as baseline methods. These models cover convolutional architectures for spatial EEG feature extraction, recurrent networks for temporal dependency modeling, attention based methods for salient temporal segment selection, and transformer based models for long range sequence representation. All comparison methods were evaluated under the same experimental protocol. The same training, validation, and test splits were used for all models. The same preprocessing pipeline, including filtering, artifact correction, epoch segmentation, baseline correction, and normalization, was applied to all datasets. Hyperparameters were tuned on the validation set, and the final performance was reported on the independent test set. Each experiment was repeated three times with different random seeds, and the mean and standard deviation were reported. As shown in Tables 3, 4, the proposed method consistently outperforms EEG specific convolutional networks, recurrent temporal models, and transformer based EEG models across all four datasets. Compared with EEGNet and DeepConvNet, the proposed method achieves higher accuracy and AUC, demonstrating the benefit of combining manifold constrained signal encoding with temporal attention routing. Compared with CNN LSTM and BiLSTM Attention, the proposed method better captures the structured spatiotemporal dynamics of EEG signals. Compared with Transformer EEG and EEG Conformer, the proposed method further improves performance by explicitly modeling predictive uncertainty, which is particularly useful for noisy EEG signals and cognitive load variation. These comparisons provide a more appropriate and rigorous evaluation because all selected baselines are directly related to EEG analysis, cognitive load estimation, or temporal signal modeling. The results more clearly demonstrate the effectiveness of the proposed framework in EEG based cognitive load tracking. Table 3Comparison with EEG and time series baselines on the Korean Phoneme EEG Signal Dataset and the Attention Based Cognitive Load Dataset.ModelKorean Phoneme EEG Signal DatasetAttention Based Cognitive Load DatasetAccuracyPrecisionRecallAUCAccuracyPrecisionRecallAUCShallowConvNet; Song et al. (2023)84.76 ± 0.6184.02 ± 0.6684.31 ± 0.6484.58 ± 0.5985.34 ± 0.5784.69 ± 0.6384.91 ± 0.6185.16 ± 0.55DeepConvNet; Li et al. (2024)85.63 ± 0.5884.94 ± 0.6285.18 ± 0.6085.46 ± 0.5686.21 ± 0.5385.58 ± 0.5985.83 ± 0.5786.04 ± 0.51EEGNet; Gupta et al. (2021)86.27 ± 0.5485.61 ± 0.5985.88 ± 0.5786.10 ± 0.5286.95 ± 0.4986.32 ± 0.5586.57 ± 0.5386.78 ± 0.48TCN; Chakladar et al. (2023)86.91 ± 0.5086.28 ± 0.5686.51 ± 0.5386.73 ± 0.4987.48 ± 0.4686.86 ± 0.5187.09 ± 0.5087.31 ± 0.45CNN LSTM; Panwar et al. (2024)87.45 ± 0.4786.83 ± 0.5287.07 ± 0.5087.29 ± 0.4688.06 ± 0.4387.49 ± 0.4887.71 ± 0.4687.94 ± 0.42BiLSTM Attention; Roy et al. (2020)88.02 ± 0.4487.41 ± 0.4987.65 ± 0.4787.86 ± 0.4388.61 ± 0.4088.03 ± 0.4588.27 ± 0.4388.49 ± 0.39Transformer EEG; Xu et al. (2022)88.54 ± 0.4287.98 ± 0.4788.19 ± 0.4588.41 ± 0.4189.07 ± 0.3888.52 ± 0.4388.74 ± 0.4188.96 ± 0.37EEG Conformer; Lim et al. (2021)89.08 ± 0.4188.53 ± 0.4688.76 ± 0.4488.97 ± 0.4089.61 ± 0.3689.07 ± 0.4289.29 ± 0.4089.51 ± 0.36Ours89.72 ± 0.4089.15 ± 0.4689.38 ± 0.4389.61 ± 0.4190.25 ± 0.3789.72 ± 0.4489.94 ± 0.4290.18 ± 0.39 Comparison with EEG and time series baselines on the Korean Phoneme EEG Signal Dataset and the Attention Based Cognitive Load Dataset. Table 4Comparison with EEG and time series baselines on the EEG Cognitive Load Phoneme Recognition Dataset and the Attention Augmented EEG Phoneme Dataset.ModelEEG Cognitive Load Phoneme Recognition DatasetAttention Augmented EEG Phoneme DatasetAccuracyPrecisionRecallAUCAccuracyPrecisionRecallAUCShallowConvNet; Song et al. (2023)85.12 ± 0.5584.47 ± 0.6184.72 ± 0.5984.96 ± 0.5385.83 ± 0.5885.15 ± 0.6485.39 ± 0.6185.62 ± 0.55DeepConvNet; Li et al. (2024)85.94 ± 0.5185.31 ± 0.5785.56 ± 0.5585.78 ± 0.5086.67 ± 0.5386.02 ± 0.5986.25 ± 0.5686.48 ± 0.51EEGNet; Gupta et al. (2021)86.68 ± 0.4886.06 ± 0.5486.31 ± 0.5286.52 ± 0.4787.36 ± 0.4986.75 ± 0.5586.98 ± 0.5387.21 ± 0.48TCN; Chakladar et al. (2023)87.21 ± 0.4586.62 ± 0.5086.84 ± 0.4887.06 ± 0.4487.92 ± 0.4687.33 ± 0.5187.55 ± 0.4987.78 ± 0.45CNN LSTM; Panwar et al. (2024)87.86 ± 0.4287.28 ± 0.4887.51 ± 0.4687.72 ± 0.4188.51 ± 0.4387.95 ± 0.4988.18 ± 0.4788.39 ± 0.42BiLSTM Attention; Roy et al. (2020)88.47 ± 0.3987.91 ± 0.4588.13 ± 0.4388.35 ± 0.3889.06 ± 0.4088.51 ± 0.4688.74 ± 0.4488.96 ± 0.39Transformer EEG; Xu et al. (2022)89.02 ± 0.3688.46 ± 0.4288.68 ± 0.4088.91 ± 0.3589.62 ± 0.3789.06 ± 0.4389.28 ± 0.4189.51 ± 0.36EEG Conformer; Lim et al. (2021)89.71 ± 0.3489.15 ± 0.3989.37 ± 0.3789.58 ± 0.3390.41 ± 0.3589.84 ± 0.4090.05 ± 0.3890.29 ± 0.34Ours90.45 ± 0.3189.87 ± 0.3690.03 ± 0.3890.25 ± 0.3391.12 ± 0.3490.54 ± 0.3990.72 ± 0.3690.95 ± 0.32 Comparison with EEG and time series baselines on the EEG Cognitive Load Phoneme Recognition Dataset and the Attention Augmented EEG Phoneme Dataset. Ablation study To assess the contribution of individual components in our proposed method, we conducted a comprehensive ablation study. The results are presented in Tables 5, 6. Each experiment isolates specific modules or features to evaluate their impact on the overall performance. This section provides an analysis of these results and highlights the significance of each component. Table 5Ablation study of Ours on Korean Phoneme EEG Signal Dataset and Attention-Based Cognitive Load Dataset.VariantKorean Phoneme EEG Signal DatasetAttention-Based Cognitive Load DatasetAccuracyPrecisionRecallAUCAccuracyPrecisionRecallAUCw./o. Manifold Constrained Signal Encoding88.34 ± 0.4287.78 ± 0.4888.01 ± 0.4588.23 ± 0.4089.12 ± 0.3988.56 ± 0.4688.79 ± 0.4389.02 ± 0.38w./o. Agent-driven Temporal Attention Routing88.67 ± 0.3988.12 ± 0.4488.34 ± 0.4288.56 ± 0.3889.45 ± 0.3688.89 ± 0.4289.12 ± 0.4089.35 ± 0.37w./o. Uncertainty-aware Cognitive Load Prediction89.01 ± 0.3788.45 ± 0.4288.68 ± 0.4088.91 ± 0.3989.78 ± 0.3489.23 ± 0.4089.46 ± 0.3889.69 ± 0.36Ours89.72 ±0.4089.15 ±0.4689.38 ±0.4389.61 ±0.4190.25 ±0.3789.72 ±0.4489.94 ±0.4290.18 ±0.39The bolded values represent the optimal values. Ablation study of Ours on Korean Phoneme EEG Signal Dataset and Attention-Based Cognitive Load Dataset. The bolded values represent the optimal values. Table 6Ablation study of Multimodal Learning methods on EEG Cognitive Load Phoneme Recognition Dataset and Attention Augmented EEG Phoneme Dataset.MethodEEG Cognitive Load Phoneme Recognition DatasetAttention Augmented EEG Phoneme DatasetAccuracyPrecisionRecallAUCAccuracyPrecisionRecallAUCw./o. Manifold Constrained Signal Encoding88.12 ± 0.3787.54 ± 0.4287.72 ± 0.4587.95 ± 0.3989.03 ± 0.4088.45 ± 0.4688.63 ± 0.4388.87 ± 0.38w./o. Agent-driven Temporal Attention Routing88.56 ± 0.3587.98 ± 0.4088.16 ± 0.4388.38 ± 0.3789.45 ± 0.3888.87 ± 0.4489.05 ± 0.4189.29 ± 0.36w./o. Uncertainty-aware Cognitive Load Prediction89.03 ± 0.3388.45 ± 0.3888.63 ± 0.4088.85 ± 0.3589.89 ± 0.3689.31 ± 0.4289.49 ± 0.3989.73 ± 0.34Ours90.45 ±0.3189.87 ±0.3690.03 ±0.3890.25 ±0.3391.12 ±0.3490.54 ±0.3990.72 ±0.3690.95 ±0.32The bolded values represent the optimal values. Ablation study of Multimodal Learning methods on EEG Cognitive Load Phoneme Recognition Dataset and Attention Augmented EEG Phoneme Dataset. The bolded values represent the optimal values. The first set of experiments, as shown in Table 5, focuses on the core modules of our framework. By removing the Manifold Constrained Signal Encoding, we observe a significant drop in performance across all metrics. This demonstrates the critical role of this module in capturing relevant features from the EEG signals. The Manifold Constrained Signal Encoding enhances the model's ability to focus on task-relevant regions, which is particularly important for datasets like the Korean Phoneme EEG Signal Dataset and the Attention-Based Cognitive Load Dataset. Furthermore, the removal of the Agent-driven Temporal Attention Routing results in a noticeable decline in accuracy. This module is essential for capturing hierarchical information from EEG signals, which often exhibit complex temporal and spatial patterns. The EEG Cognitive Load Phoneme Recognition Dataset benefits significantly from this module, as its data structures require multi-scale processing to achieve optimal results. The ablation results confirm that the combination of these modules is indispensable for achieving state-of-the-art performance. The second set of experiments, detailed in Table 6, investigates the impact of optimization strategies and data augmentation techniques. Removing the Uncertainty-aware Cognitive Load Prediction leads to slower convergence and suboptimal results. This highlights the importance of using tailored prediction methods to handle the non-stationary nature of EEG data. The exclusion of data augmentation results in reduced robustness and generalization capabilities. EEG datasets often suffer from limited sample sizes and high variability. Data augmentation mitigates these issues by introducing variability during training, thereby improving the model's ability to generalize to unseen data. The results also show that the combination of augmentation and prediction strategies synergistically enhances performance, further validating their inclusion in our framework. Runtime efficiency analysis To further assess the computational practicality of the proposed Adaptive EEG Attention Tracker, a runtime efficiency analysis was conducted under the same experimental environment used in the main experiments. The evaluation included the average inference time per sample, the average training time per epoch, the number of model parameters, peak GPU memory usage, and throughput. All runtime results were averaged over three independent runs to reduce measurement variance. As shown in Table 7, the proposed method introduces additional computational overhead due to the Agent driven Temporal Attention Routing and the Uncertainty aware Cognitive Load Prediction modules. The average inference latency remains within a practical range for EEG based cognitive load tracking. Compared with the variant without Agent driven Temporal Attention Routing, the full model requires a longer inference time, indicating that temporal attention is one of the main sources of additional runtime. Compared with the variant without Uncertainty aware Cognitive Load Prediction, the full model requires additional computation for predictive variance estimation. Nevertheless, the increase in runtime is moderate and is accompanied by improved predictive accuracy, AUC, and robustness under noisy EEG conditions. The results in indicate that the proposed method achieves a favorable balance between predictive performance and computational efficiency. Although the full model incurs additional runtime compared with simplified variants, the inference latency remains at the millisecond level. The proposed framework is suitable for offline EEG analysis and near real time cognitive load tracking. For strict real time BCI deployment on resource constrained devices, further acceleration strategies such as sparse attention, model pruning, low rank approximation, and hardware aware optimization will be explored in future work. Table 7Runtime efficiency comparison of different model variants.MethodInference time per sample (ms)Training time per epoch (s)Parameters (M)Peak GPU memory (GB)Throughput (samples/s)w/o Manifold Constrained Signal Encoding3.8241.68.213.48261.8w/o Agent-driven Temporal Attention Routing3.3538.97.943.21298.5w/o Uncertainty-aware Cognitive Load Prediction3.6140.28.373.36277.0Ours4.1845.79.063.82239.2 Runtime efficiency comparison of different model variants. Training dynamics and convergence analysis To further describe the model training process, the optimization dynamics of the proposed Adaptive EEG Attention Tracker were analyzed across training epochs. Specifically, the training and validation loss values, prediction accuracy, comparison with a representative EEG baseline, and convergence stability across multiple random seeds were recorded and visualized. The overall results are shown in Figure 4. The training and validation loss curves decrease rapidly during the early optimization stage, indicating that the model can effectively learn discriminative EEG representations from the input signals. After the warm up phase, the loss curves continue to decline steadily. The two learning rate decay operations further reduce the optimization step size and help the model approach a stable convergence state. In the final epochs, the validation loss becomes nearly stable within the early stopping monitoring region, suggesting that the model has converged without obvious training instability. The prediction accuracy curves show that both training accuracy and validation accuracy increase quickly at the beginning and then gradually saturate. The validation accuracy follows a similar trend to the training accuracy, and the gap between the two curves remains limited throughout the training process. This indicates that the model maintains good generalization ability and does not show severe overfitting. The comparison with EEGNet further demonstrates the convergence advantage of the proposed method. The proposed Adaptive EEG Attention Tracker reaches a higher validation accuracy within fewer epochs and maintains a consistent performance advantage after convergence. This result suggests that the temporal attention routing and uncertainty aware prediction modules improve both convergence speed and final predictive performance. The multiple seed results further verify the reproducibility of the training process. The mean validation accuracy increases steadily and becomes stable in the later training stage, while the standard deviation region remains within a moderate range. This indicates that the model is not highly sensitive to random initialization and can be trained stably under different experimental runs. The results in Figure 4 demonstrate that the proposed model achieves stable convergence, continuous accuracy improvement, and reliable generalization. The small gap between training and validation performance indicates that overfitting is effectively controlled. This stability is mainly attributed to the adopted optimization and regularization strategies, including warm up training, learning rate decay, dropout, weight decay, gradient clipping, and early stopping monitoring. Training dynamics and convergence analysis of the proposed model. (a) Training and validation loss convergence across epochs. (b) Training and validation prediction accuracy across epochs. (c) Validation accuracy comparison between the proposed Adaptive EEG Attention Tracker and EEGNet. (d) Convergence stability of the proposed method across multiple random seeds, where the solid line denotes the mean validation accuracy and the shaded region denotes the standard deviation. Conclusions and future work In this study, we addressed the challenge of cognitive load tracking in Korean phoneme recognition by proposing the Adaptive EEG Attention Tracker, a novel framework that integrates advanced attention mechanisms and uncertainty quantification into EEG signal analysis. Our approach combines the Manifold Constrained Signal Encoding, the Agent-driven Temporal Attention Routing, and the Uncertainty-aware Cognitive Load Prediction to effectively capture the spatiotemporal dynamics of EEG signals while maintaining structural fidelity and interpretability. The experimental results demonstrated that our framework significantly improves the accuracy and robustness of cognitive load estimation compared to existing methods. By leveraging the Uncertainty Propagation Adjustment strategy, we further enhanced the reliability of predictions, marking a substantial advancement in EEG-based cognitive modeling for complex linguistic tasks. Despite the promising results, our framework has two notable limitations. First, the computational complexity of the attention mechanisms and uncertainty modeling increases the processing time, which may limit real-time applications. Future work could explore lightweight model architectures or hardware acceleration to address this issue. Second, the generalizability of the framework to other languages or cognitive tasks remains untested. Expanding the dataset to include diverse linguistic and cognitive scenarios will be crucial for validating the broader applicability of the proposed method. Despite the remaining challenges, this study establishes a robust foundation for future research in EEG-based cognitive load tracking, with potential applications in language processing and other areas.