The study of electroencephalography (EEG) has become a cornerstone in understanding cognitive processes, including language comprehension (Ariastuti and Wahyudin, 2022). EEG allows researchers to measure brain activity with high temporal resolution, enabling the examination of neural responses to linguistic stimuli (Wu et al., 2022). A significant body of research focuses on the temporal dynamics of event-related potentials (ERPs) during language processing tasks (Zou et al., 2021). For example, the N400 component is widely studied for its role in semantic processing, revealing insights into how the brain resolves meaning inconsistencies (Coleman, 2021). Similarly, the P600 component has been linked to syntactic processing and reanalysis during sentence comprehension (Aoyama, 2021). These findings underscore the importance of EEG in mapping the temporal stages of language comprehension (Elliott and Hodgson, 2021). Furthermore, frequency-based EEG analyses, such as alpha and theta power modulations, have been explored to understand attentional and memory mechanisms during language tasks (Zhao et al., 2020). While these studies provide a robust foundation, the impact of external factors, such as environmental stressors or altered physiological conditions like hypoxia, remains less understood (Simamora and Oktaviani, 2020). Examining how hypoxia modulates these EEG markers could reveal how adverse conditions affect language processing (Iturriaga, 2023).

Hypoxia, characterized by reduced oxygen availability, has profound effects on brain function, including cognitive and linguistic abilities (Bae and Park, 2020). Existing research highlights how hypoxia impacts attention, memory, and executive function (Yunita and Maisarah, 2020). Studies employing neuroimaging and behavioral assessments have demonstrated significant cognitive deficits under acute and chronic hypoxic conditions (Septiyanti et al., 2020). These include slower reaction times, decreased working memory capacity, and impaired decision-making. Despite these findings, there is a notable gap in the literature concerning hypoxia's effects on specific cognitive domains such as language comprehension (Seo, 2020). Investigating this relationship is essential, as language comprehension relies on the integration of multiple cognitive resources, including attention and working memory. Hypoxia-induced changes in brain physiology, such as reduced cerebral oxygenation and altered neurotransmitter dynamics, may disrupt these processes (Rusmiyanto et al., 2023). EEG studies could provide valuable insights by identifying how hypoxia modulates neural correlates of language comprehension, such as ERP components and oscillatory activity patterns (Alfallaj et al., 2021).

The interaction of hypoxia with other stressors, such as cognitive load or emotional stress, presents a complex challenge to understanding brain function (Sallam, 2023). Multimodal studies investigating combined stress effects are relatively sparse, yet crucial for understanding real-world scenarios (Zein et al., 2020). Cognitive tasks, such as language comprehension, often occur under conditions involving multiple concurrent demands. The interplay between hypoxia and additional stressors may exacerbate neural inefficiencies, leading to amplified cognitive deficits (Shaikh et al., 2023). Research utilizing EEG has shown that stressors such as mental fatigue or anxiety can modulate brainwave patterns, particularly in the alpha and beta frequency bands (Renganathan, 2021). The integration of EEG with other physiological measures, such as heart rate variability or blood oxygen saturation, could provide a holistic view of how hypoxia interacts with stress (Syakur et al., 2020). Moreover, advanced analytical techniques, such as machine learning models, could be employed to decode complex neural patterns arising from multimodal stress conditions (Sofyan, 2021). This direction not only addresses theoretical questions but also has practical implications for environments where individuals face simultaneous cognitive and physiological challenges.

In recent years, English education has become a critical area of focus due to its global significance in academic, professional, and social contexts. This subsection provides an overview of the methodology employed to enhance English learning outcomes, particularly in environments where English is taught as a second language (ESL). We aim to tackle challenges in comprehension, expression, and fluency through the integration of novel pedagogical strategies, leveraging technological advancements, and understanding linguistic nuances.

The upcoming subsections will address various facets of our approach. In Section 3.2, we formalize the problem of English education by analyzing common linguistic barriers and presenting a structured framework to model them. This foundational section establishes the key challenges in vocabulary acquisition, grammar comprehension, and cultural fluency, emphasizing their interconnected nature. In Section 3.3, we introduce a novel framework, hereafter referred to as the Dynamic Linguistic Enhancement Model (DLEM). This model builds upon insights from cognitive science and language processing to deliver adaptive and personalized learning pathways. Key components of DLEM include contextualized learning environments and multi-modal interactions, which are meticulously designed to simulate real-world communication. In Section 3.4, we propose an innovative strategy, termed the Contextual Augmented Learning Strategy (CALS), to integrate our model effectively into diverse educational settings. This strategy focuses on adaptive curriculum design, dynamic feedback systems, and the utilization of gamification to foster learner engagement and motivation. The emphasis is on scalability and flexibility, ensuring applicability across varied cultural and institutional contexts.

English education, particularly in environments where it is taught as a second language, presents unique challenges that require careful analysis and systematic formalization. To address these challenges, we introduce a mathematical and conceptual framework that captures the complexities of language acquisition, comprehension, and usage. The English learning process can be represented as a multi-stage system:

In this framework, V represents the domain of vocabulary acquisition, encompassing the process of learning and retaining new words. G denotes the grammatical structures of the language, including syntax and morphology, which govern sentence construction. C embodies the cultural and contextual understanding that is essential for meaningful and effective language use. These components interact dynamically within the cognitive capabilities of the learner and the environmental influences they encounter, creating a complex and interdependent system. This framework provides a structured approach to understanding and addressing the multifaceted nature of English language learning. The vocabulary learning process can be modeled as a stochastic process, where the probability of acquiring a word w_i at time t is dependent on exposure E(w_i, t) and reinforcement R(w_i, t). Formally:

where f is a monotonic function that combines exposure and reinforcement effects. Reinforcement often depends on the frequency and utility of w_i in specific contexts:

with α and β as tunable parameters representing learner-specific sensitivity. Grammar is structured around a set of syntactic rules R = {r₁, r₂, …, r_n}, where r_i defines transformations or associations between linguistic constructs. The learner's ability to internalize R is influenced by cognitive factors such as memory and reasoning capabilities. Define P(r_i, t) as the probability of mastering rule r_i over time:

where engage(r_i, τ) represents active interaction with r_i, and feedback(r_i, τ) captures corrective signals received during learning. The contextual use of English involves the integration of vocabulary and grammar with socio-cultural norms. We model this integration using a latent semantic space L, where each word or phrase w_i maps to a point vi∈ℝd. Contextual similarity between two phrases w_i and w_j is measured by:

Cultural fluency is then represented as the learner's capacity to form coherent trajectories in L, connecting semantic and pragmatic elements effectively. Given a target proficiency level T defined across the axes of vocabulary, grammar, and context, the objective is to design an optimal learning pathway L^* such that:

where U(L(t)) denotes the utility function capturing linguistic growth at time t.

To advance the field of English education and address multifaceted challenges in second-language learning, we propose the Dynamic Linguistic Enhancement Model (DLEM). This innovative framework synergizes cognitive science, adaptive learning methodologies, and computational advancements to deliver a personalized, structured, and engaging approach to language acquisition (as shown in Figure 1). Below, we detail its three core innovations as following.

The Contextual Augmented Learning Strategy (CALS) introduces a comprehensive framework designed to facilitate the seamless integration of advanced linguistic models into diverse educational and digital platforms (as shown in Figure 3). This section highlights the three key innovations in CALS: Adaptive Curriculum Design, Dynamic Feedback Systems, and Gamified Engagement Frameworks.

The Sleep-EDF Dataset (Wang et al., 2024) is a comprehensive collection of sleep recordings designed for research in sleep stage classification and related studies. It includes polysomnographic (PSG) data, encompassing electroencephalogram (EEG), electrooculogram (EOG), and electromyogram (EMG) signals from healthy individuals and patients with sleep disorders. The dataset spans multiple nights for some subjects, offering insights into inter-night variability. The detailed annotations and long-term recordings make it a valuable resource for sleep pattern analysis and machine learning applications in health monitoring. The EEGEyeNet Dataset (Modesitt et al., 2023) focuses on eye movement classification using EEG signals. It consists of recordings from subjects performing controlled eye movements, such as fixations and saccades, under well-defined experimental conditions. The dataset includes high-resolution EEG data and corresponding event markers, providing a robust foundation for developing models that link neural activity to ocular dynamics. Its emphasis on eye movement makes it uniquely suited for advancing research in brain-computer interfaces and cognitive neuroscience. The CHB-MIT Dataset (Duan et al., 2021) is a widely used resource for seizure detection and prediction studies, offering long-term EEG recordings from pediatric epilepsy patients. The dataset includes scalp EEG data annotated with seizure events, recorded over extended periods to capture both ictal and interictal states. The comprehensive annotations and real-world variability make it an essential benchmark for developing and evaluating algorithms in epilepsy diagnosis and management, particularly in clinical and ambulatory settings. The PhyAAt Dataset (Ahuja and Setia, 2022) is a multi-modal collection designed for physical activity analysis and assessment. It integrates accelerometer, gyroscope, and physiological data, such as heart rate, captured during various physical activities and rest states. The dataset includes diverse demographic information, ensuring its applicability across different populations. Its multi-modal nature enables the exploration of relationships between physiological and physical signals, making it a key resource for wearable technology development and health monitoring systems.

While none of the datasets used in this study were collected under natural high-altitude or hypoxic conditions, they were selected for their high signal quality, extensive annotations, and task diversity–making them well-suited for controlled evaluation of EEG-based cognitive modeling frameworks. The Sleep-EDF dataset captures physiological brain states during cognitive transitions such as sleep stage changes; CHB-MIT contains EEG recordings under clinical stress settings, including epileptic seizure episodes; and EEGEyeNet includes tasks involving attentional shifts and oculomotor coordination. Although these contexts differ from altitude-induced stress, they share critical cognitive stress features such as fluctuating attention, increased working memory demands, and altered neurophysiological baselines. To approximate real-world cognitive stressors associated with hypoxia, we designed our task stimuli and preprocessing strategy to simulate conditions of high mental load. For example, auditory comprehension inputs were structured with temporally compressed, semantically rich materials to elevate processing demands. These interventions elicit EEG dynamics (e.g., elevated theta and suppressed alpha power) that closely align with prior studies on acute hypoxic exposure. Consequently, while our current data does not originate from high-altitude populations or explicitly track participants' native language profiles, it provides a valid simulation environment for benchmarking the DLEM and CALS framework. We fully acknowledge the importance of ecological validity. Future extensions of this work will involve targeted EEG data collection from individuals residing in high-altitude regions or within hypobaric chamber conditions. This will allow for stratified model validation and domain-specific adaptation. At the current stage, however, our goal is to demonstrate the architectural generalizability of our model under controlled, stress-emulated settings, laying a foundation for field-deployable applications.

The experiments were conducted on a server equipped with an NVIDIA RTX 3090 GPU and 128 GB RAM to ensure computational efficiency. For model training, PyTorch was utilized as the primary deep learning framework. The Adam optimizer was chosen for its adaptive learning rate properties, set to an initial learning rate of 1 × 10⁻³ with a cosine annealing scheduler to gradually reduce the learning rate during training. The batch size was set to 64, balancing memory constraints and training speed. All models were trained for 100 epochs to ensure convergence while avoiding overfitting. For preprocessing, the EEG signals were band-pass filtered between 0.5 Hz and 50 Hz to remove artifacts and focus on the relevant frequency bands. We applied a band-pass filter between 0.5 Hz and 50 Hz to all EEG signals, which is a widely accepted standard in cognitive and neuropsychological studies. This frequency window was chosen to preserve the core EEG components known to reflect cognitive processes–such as theta and alpha rhythms associated with working memory and attention, and beta/gamma rhythms linked to cognitive control and perceptual integration. Frequencies below 0.5 Hz were excluded to eliminate slow baseline drifts and electrodermal artifacts, while frequencies above 50 Hz were removed to suppress powerline interference and muscle-related artifacts. The preserved bands (0.5–50 Hz) include delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and low gamma (30–50 Hz), all of which have been shown to be modulated under hypoxic conditions in existing EEG literature. While some ultra-high frequency activity (>60 Hz) has been reported in invasive or high-density EEG contexts, such ranges are more susceptible to environmental noise in scalp recordings, particularly under mobile or multi-site experimental setups. Therefore, the selected band range represents a practical and physiologically meaningful trade-off to support consistent signal processing across datasets with different recording conditions. Future work may explore dynamic filtering or high-frequency EEG analysis in closed-loop neurofeedback systems under extended hypoxic exposure. Data augmentation techniques, including random cropping and noise injection, were applied to increase model robustness. Each dataset was split into 80% training, 10% validation, and 10% testing sets, ensuring a balanced evaluation. Cross-validation was employed where applicable to ensure consistency across splits. The neural network architecture comprised a combination of convolutional and recurrent layers. The model included a convolutional feature extractor followed by bidirectional LSTMs to capture temporal dependencies. Dropout layers with a rate of 0.5 were used to mitigate overfitting, and a softmax activation function was applied at the output layer for multi-class classification tasks. Metrics used for evaluation included accuracy, precision, recall, F1-score, and area under the ROC curve (AUC). These metrics were computed for each dataset to enable a thorough assessment of the model's performance across diverse scenarios. Gradient class activation maps (Grad-CAMs) were employed to visualize model decision-making, offering interpretability for the deep learning predictions. Hyperparameter tuning was conducted using grid search, varying learning rates, batch sizes, and dropout rates. The optimal configuration was selected based on validation performance. Regularization techniques, such as L2 regularization with a weight decay factor of 1 × 10⁻⁴, were incorporated to prevent overfitting. The models were implemented with mixed precision training to accelerate computation without compromising numerical stability. For datasets with imbalanced class distributions, techniques like oversampling and class-specific weighting were applied during training to ensure fair representation. All experiments were repeated three times to account for randomness, and results were reported as mean values with standard deviations (Algorithm 1).

The performance of our proposed model was assessed against various state-of-the-art (SOTA) methods, including CLIP (Zhang et al., 2025), ViT (Touvron et al., 2022), I3D (Peng et al., 2023), BLIP (Wattasseril et al., 2023), Wav2Vec 2.0 (Chen and Rudnicky, 2023), and T5 (Grover et al., 2021), across diverse datasets such as Sleep-EDF, EEGEyeNet, CHB-MIT, and PhyAAt. Comprehensive results are summarized in Tables 1, 2, highlighting key metrics like accuracy and recall.

On the Sleep-EDF dataset, our approach demonstrated remarkable effectiveness, achieving an accuracy of 90.12% and a recall of 87.95%, showcasing its capability in discerning intricate patterns for sleep stage classification. Similarly, the EEGEyeNet dataset results underscored the method's robustness in modeling temporal and multimodal embeddings, with accuracy and recall exceeding 91% and 89%, respectively. These outcomes align with Grad-CAM visualizations, which reveal the method's ability to focus on salient temporal features.

For the CHB-MIT dataset, crucial for seizure detection, the model achieved an accuracy of 88.45%, underscoring its reliability in high-stakes clinical applications. On the PhyAAt dataset, leveraging both physical and physiological data, the model maintained high performance, with accuracy reaching 89.60%. These results collectively affirm the adaptability of our architecture across domains.

The proposed framework's integration of convolutional and recurrent components, coupled with tailored augmentations and regularization strategies, distinguishes it from existing SOTA approaches. For instance, while ViT (Touvron et al., 2022) and BLIP (Wattasseril et al., 2023) excel in certain contexts, their lack of recurrent layers limits their capacity for temporal modeling. Similarly, CLIP (Zhang et al., 2025) and Wav2Vec 2.0 (Chen and Rudnicky, 2023), relying on static embeddings, underperform in dynamic feature extraction tasks. This comparative analysis, complemented by Figures 5, 6, illustrates the consistency and superior generalization of our model across diverse applications.

The comparison across these benchmarks indicates that our proposed architecture's combination of convolutional and recurrent components, along with advanced data augmentation and regularization techniques, effectively generalizes across diverse domains. Figures 5, 6 illustrates the comparative metrics visually, affirming the consistency and robustness of the proposed model across multiple datasets. Notably, the superior performance of our model, especially in terms of AUC, emphasizes its reliability in high-stakes applications like medical diagnostics and human-computer interaction. The SOTA models, while competitive, did not incorporate domain-specific augmentations or the temporal modeling precision facilitated by our architecture. For instance, ViT (Touvron et al., 2022) and BLIP (Wattasseril et al., 2023) perform well but lack the recurrent layers necessary to fully exploit temporal dependencies in EEG and physiological data. Models like CLIP (Zhang et al., 2025) and Wav2Vec 2.0 (Chen and Rudnicky, 2023), while effective for certain tasks, rely primarily on static embeddings, which may explain their lower performance on datasets requiring dynamic feature extraction.

To evaluate the contributions of individual components in our model, an ablation study was conducted on the Sleep-EDF, EEGEyeNet, CHB-MIT, and PhyAAt datasets. The results are summarized in Tables 3, 4, which present the metrics for Accuracy, Recall, F1 Score, and AUC under different configurations. The configurations examined include the removal of specific components, denoted as “w./o. Hybrid Grammar Contextualization Engine”, “w./o. Adaptive Curriculum Design”, and “w./o. Dynamic Feedback Systems”, as well as the full model (“Ours”).

For the Sleep-EDF dataset, the removal of Hybrid Grammar Contextualization Engine resulted in a drop in accuracy from 90.12% to 86.12%, indicating the importance of this component in capturing essential sleep features. Similarly, the absence of Adaptive Curriculum Design reduced the AUC from 91.40% to 89.55%, highlighting its role in enhancing class separability. The full model outperformed all variations, achieving an F1 score of 88.65% and a recall of 87.95%, which underscores the synergistic effect of all components working in concert. A similar pattern was observed on the EEGEyeNet dataset, where the full model achieved an accuracy of 91.80% and an AUC of 93.10%, with noticeable declines when any component was excluded. These results demonstrate the importance of comprehensive feature extraction and temporal modeling strategies. For the CHB-MIT dataset, the removal of Hybrid Grammar Contextualization Engine caused a decrease in accuracy from 88.45% to 84.10% and a drop in recall from 86.90% to 81.95%. This finding suggests that Hybrid Grammar Contextualization Engine significantly contributes to identifying seizure events, likely by capturing critical temporal dynamics. Removing Dynamic Feedback Systems, which is designed to integrate multi-modal features, resulted in a reduction in AUC from 90.20% to 88.30%. This emphasizes the importance of multi-modal embeddings in achieving robust performance in seizure detection. On the PhyAAt dataset, the absence of Adaptive Curriculum Design reduced the recall from 87.85% to 85.05%, revealing its role in refining activity-specific features. The full model consistently achieved the best results, with an AUC of 91.10%, demonstrating its effectiveness in leveraging both physiological and physical signals.

The findings from the ablation study affirm that each component in our model architecture plays a critical role in optimizing performance. Hybrid Grammar Contextualization Engine likely enhances temporal feature extraction, while Adaptive Curriculum Design contributes to fine-grained feature refinement. Dynamic Feedback Systems integrates multi-modal inputs, enabling the model to learn complex relationships across signal domains. The superior performance of the full model validates the design decisions made in the architecture, highlighting its potential for applications in emotion analysis, seizure detection, and activity recognition. Figures 7, 8 visually compare the ablation study metrics, providing further insights into the impact of each component. These visualizations illustrate the consistent advantage of the full model across all datasets, reinforcing its robustness and generalizability in diverse contexts.

To evaluate the cross-linguistic generalizability of our proposed framework, we conducted additional experiments using the SEED dataset, which consists of EEG recordings from Mandarin-speaking participants performing language comprehension tasks. Table 5 presents the comparative performance of our model against six state-of-the-art baselines, consistent with those used in the main experiments. Our framework achieves the highest accuracy (88.75%), F1 score (87.20%), and AUC (89.95%) across all models, outperforming both audio-language models such as Wav2Vec 2.0 and T5, as well as vision-language models like CLIP and BLIP adapted to textual features. The robust performance observed on a Mandarin-language EEG dataset indicates that the architecture's multimodal alignment and cultural embedding mechanisms are effective beyond English, supporting its broader application to multilingual and culturally diverse populations. These findings further validate the adaptability of the DLEM and CALS components to non-English contexts, reinforcing the model's potential for global deployment in language-related cognitive modeling under environmental stressors.