Delineating the Role of Alpha Waves in Exercise-induced Neural Changes through Resting-state EEG

In contemporary society, the accelerating pace of life and the increasing burden of life pressures have led to a significant upsurge in the prevalence of emotional dysregulation. Among various manifestations of emotional dysregulation, anxiety, a prevalent subtype, poses a great challenge to individuals. Pharmacological therapies have long been regarded as a fundamental approach in the treatment of emotional dysregulation, particularly anxiety. However, research has shown that approximately 30% of individuals with emotional dysregulation do not respond to first-line medications. Moreover, long-term use of these drugs may give rise to various risks, such as metabolic disorders and cognitive impairment¹. Psychological interventions, although addressing etiological factors through evidence-based frameworks, are limited by prolonged treatment durations requiring substantial time, effort, and financial resources, alongside the delayed onset of therapeutic effects^2,3.

In recent years, exercise intervention has been demonstrating remarkable advantages in the treatment of emotional dysregulation. A multitude of studies have indicated that exercise has the potential to naturally enhance emotional states and alleviate anxiety and depression, achieved through the promotion of endogenous neurotransmitter release and the induction of synaptic changes⁴. For example, research on exercise-trained mice revealed that their hypoxic burden was reduced by 52%, and a significant enhancement in cognitive function was observed⁵. Trait anxiety, which represents an individual's relatively stable and long-lasting tendency to experience anxiety across diverse situations⁶, is a key factor in understanding the underlying mechanisms of emotional dysregulation. It serves as a core feature of chronic anxiety, and studying it can provide valuable insights into the pathophysiology of such emotional dysregulation. By understanding trait anxiety, we can better comprehend why some individuals are more prone to developing anxiety-related mood problems. In our previous work, we elaborated on the main brain regions related to emotional cognitive functions that are impaired in emotional disorders and how exercise intervention can improve these cognitive functions and relevant brain regions⁷. Additionally, we conducted two electroencephalogram (EEG) experiments to explore in detail how exercise intervention can improve the characteristics of brain activity in attention control ability among individuals with high-trait anxiety⁸.

While exercise intervention has emerged as a promising non-pharmacological approach in depression treatment, the precise neural biomarkers associated with the positive effects of exercise intervention have not yet been clearly identified^9,10. Neural oscillatory rhythms, acting as the "spatiotemporal encoders" of brain information processing, exhibit characteristic dysregulation in anxiety. For example, research has shown that prefrontal Alpha(α) desynchronization is associated with cognitive control deficits commonly observed in anxiety^11,12. This dysregulation of neural oscillatory rhythms indicates an underlying disruption in the normal neural communication processes that are crucial for emotional regulation. However, there is a dearth of studies that comprehensively explore how exercise actually reshapes emotional function by modulating cross-regional rhythmic coupling or local field potential dynamics^13,14.

Recent advances in EEG-based deep learning research have provided novel paradigms for understanding pathological mechanisms and developing precision treatments for mental disorders such as depression and anxiety¹⁵. Notably, studies using dynamic functional connectivity (DFC) of resting-state EEG combined with hidden Markov models (HMMs) have revealed significant differences in Delta ( δ), Theta (θ), Alpha (α), and Gamma (γ) band network dynamics among non-psychotic depression, psychotic depression, and schizophrenia^16,17,18. A DFC-based binary classification model achieved 73.1% accuracy in distinguishing these three conditions, outperforming traditional static analyses. Key biomarkers included θ-band DMN-SN synchronization, γ-band FPCN-limbic system synchronization, and HMM state transition probabilities, establishing a new framework for precision psychiatric classification¹⁹ employed graph theoretical analysis to demonstrate that baseline brain network features predict deep brain stimulation (DBS) efficacy in treatment-resistant depression. A random forest model using network metrics achieved 81.2% accuracy in predicting DBS response, surpassing clinical scales. Longitudinal data showed DBS reverses network dysfunction by enhancing δ-band global synchronization and reducing sgACC centrality. Additionally, left prefrontal α-wave power predicted antidepressant non-response, with a convolutional neural network (CNN) model achieving 82.3% accuracy based on α-asymmetry²⁰. Everaert et al. (2022) developed an artificial neural network model with feature selection using 460 participants to identify predictive features of emotion regulation strategies. These findings underscore the critical need to identify precise neural targets to optimize exercise prescriptions²¹.

In the realm of exercise-related neuroscience research, deep learning has emerged as a powerful tool, enabling the extraction of robust neural biomarkers from the complex, high-dimensional, and low-amplitude spatiotemporal neurological data generated by exercise interventions. Multiple studies have demonstrated that physical activity significantly modulates activation patterns in motor-related brain regions and neural oscillatory dynamics across frequency bands^22,23,24. A systematic review of 47 studies revealed consistent increases in prefrontal α/β band power following exercise, likely reflecting enhanced neuroplasticity and cortical inhibition²⁵. Both acute exercise and long-term training induced similar trends, though γ band responses showed intensity-dependent heterogeneity (e.g., moderate aerobic vs. high-intensity interval training). Four-month aerobic interventions in healthy young adults produced significant prefrontal α wave (9-12 Hz) augmentation, positively correlated with aerobic fitness gains. While behavioral improvements in reaction time or accuracy were absent, neural oscillation metrics indicated dynamic optimization of visual attention networks, suggesting α waves may serve as biomarkers for exercise efficacy²⁶. High-level sport experts exhibited elevated sensorimotor rhythm (SMR, 12-15 Hz) power during aiming tasks, concurrent with reduced prefrontal-temporal coherence, indicating automated motor skill execution and network efficiency enhancement²⁷. Notably, table tennis athletes showed reduced activation in exercise-related brain regions compared to non-athletes, suggesting long-term training builds specialized, energy-efficient neural networks²⁸.

This study focuses on trait anxiety as a specific research subject, employing electroencephalography (EEG) to collect neural data and explore its neural biomarkers, thereby providing novel insights for identifying precise neural targets. Previous research indicates that alpha waves in the prefrontal region are closely associated with emotional regulation, cognitive control, and emotional recognition (Harmon-Jones et al., 2010), playing a pivotal role in processes such as decoding external emotional cues (e.g., facial expressions, vocal tones) and modulating emotional responses. Studies suggest that alterations in prefrontal alpha activity may serve as physiological markers of emotional dysregulation, particularly in anxiety and negative emotional states^29,30,31. Resting-state electroencephalography (EEG) serves as a default experimental condition in neuroscience for investigating the dynamic properties of the brain, requiring participants to remain awake without performing any cognitive tasks³². Experimental conditions may include eyes-closed or eyes-open states.Empirical evidence indicates that changes in prefrontal alpha oscillations could function as biomarkers for impaired emotion regulation, especially in conditions characterized by anxiety and a predominance of negative affect^33,34. Its power spectral density and functional connectivity patterns can reveal the intrinsic activity characteristics of the brain, and are applicable to detecting pathological markers in neurodegenerative diseases (e.g., Alzheimer's disease), developmental disorders (e.g., developmental dyslexia)^35,36, as well as mental and emotional disorders (e.g., depression and anxiety)³⁷. Among these, the alpha rhythm under the eyes-open condition is commonly utilized in studies on emotional disorders^38,39. Consequently, this study investigates the classification performance of alpha oscillations in prefrontal regions before and after exercise interventions for trait anxiety. Building on EEG data, this research employs EEGNet to identify neural targets associated with exercise interventions for individuals with high trait anxiety. EEGNet is specifically designed for EEG signal classification and offers several key advantages over traditional and other deep learning methods, making it particularly suitable for investigating EEG patterns with limited data⁴⁰.

The resting-state EEG data were collected using a 64-channel system (Brain Products, Germany) following the 10-20 international standard, with a sampling rate of 1000 Hz and bandpass filtering (0.1-100 Hz). To ensure signal quality, electrode impedance was maintained below 5 kΩ, and ocular artifacts were removed via Independent Component Analysis (ICA). Participants were instructed to remain awake with eyes open while fixating on a cross, minimizing movement-related noise.

Key inclusion criteria for high-trait-anxiety participants were: (1) Trait Anxiety Inventory scores ≥ 55, (2) limited high-intensity exercise (< 3 days/week) to control for pre-existing fitness effects, and (3) total weekly physical activity < 600 MET-min. These criteria aimed to homogenize the sample while reflecting real-world sedentary populations. A limitation is the potential variability in resting-state EEG dynamics due to individual differences in baseline arousal or undetected subclinical conditions, which future studies could address with larger samples and multimodal assessments (e.g., fMRI or behavioral tasks).

We hypothesize that prefrontal alpha activity can effectively classify the EEG data of exercise and control. In summary, this study aims to leverage AI technologies to analyze the benefits of exercise interventions for emotional disorders, using trait anxiety as a model. Through its methodology and findings, this work seeks to enhance understanding of current developments and challenges in the field, offering guidance and insights for future research.

This study was approved by the Institutional Research Ethics Committee of Wuhan Sports University (2023016).

1. Study participants

1. Recruit 550 non-sports major students from a university on a voluntary basis. Perform pre-screening for sample selection. Use the Chinese Revised State-Trait Anxiety Inventory⁴¹ for this purpose, as shown in Table 1.
2. Classify individuals scoring 55 or higher on the Trait Anxiety subscale as having high trait anxiety and select them for the EEG experiment. Use the Trait Anxiety subscale of this inventory as the key assessment tool, following established criteria from previous ERP studies⁴². This cutoff score of 55 was grounded in prior research, ensuring a consistent and valid way to categorize participants with high trait anxiety.
3. Further assess individuals with anxiety scores higher than 55 for their physical activity levels. Participants were also required to meet three specific inclusion criteria related to their physical activity patterns in the preceding week.
   1. Ensure that participants have engaged in high-intensity physical activity for less than 3 days (with each session lasting ≥ 20 min per day).
      NOTE: High-intensity physical activity can have a significant impact on the body and brain, potentially altering neural plasticity and neurotransmitter levels. By restricting the number of high-intensity activity days, the researchers can minimize the confounding effects of such intense exercise on the relationship being studied.
   2. Ensure the total number of exercise days across all intensities is less than 5 days per week, and the total exercise volume is less than 600 MET-min/week. This will ensure that the overall exercise load of the participants is within a certain range, reducing the variability in the sample due to different exercise volumes.
   3. Confirm that participants engaged in less than 30 min of moderate-intensity activity and walked for less than 5 days per week.
      NOTE: Moderate-intensity activity and walking, although less intense than high-intensity exercise, can still influence the body's physiological and psychological state. By setting these limits, the researchers can further control the potential effects of different types of physical activity on the experimental results.

2. Task instruction

1. Instruct the participants in the exercise intervention group to engage in moderate-intensity aerobic exercise using a stationary cycle ergometer for 30 min, wearing a heart rate monitor to record physiological data. The exercise intervention was structured into three phases following established guidelines⁴³.
   1. Warm-up phase: Instruct the participants to cycle at 25 watts for 5 min to prepare musculature and cardiovascular systems for the main exercise phase.
   2. Main exercise phase: Instruct the participants to perform a 20 min continuous cycling task at a pedaling rate of 70-80 rpm at moderate intensity. Monitor heart rate every 2 s, and calculate target heart rate (THR) using the formula:
      THR (HRmax − HRrest) × desired intensity + HRrest
      where HRmax = maximal heart rate, HRrest = resting heart rate, and the desired intensity was determined according to ACSM exercise guidelines⁴³.
   3. Cool-down phase: Ask the participant to reduce cycling load to 15 W within 5 min to lower heart rate, minimize injury risk, and promote muscle relaxation.
2. Control group task: Ensure the participant sits quietly in the laboratory for 30 min and engages in a free reading task.

3. Data collection

1. Record the EEG signals of subjects using a 64-lead device, and set the electrode array according to the international standard 10-20 system with a sampling frequency of 1000 Hz.
2. Prior to data acquisition, apply medical conductive paste to all electrode-skin interfaces to ensure impedance values below 5 kΩ, thereby optimizing signal fidelity and reducing motion artifacts.
3. The sampling rate for data collection was 1000 Hz, with a bandpass filter of 0.1-100 Hz. Maintain electrode impedance below 5 kΩ.
4. Place an electrode at the left outer canthus of the left eye, and record vertical eye movements and blinks.
5. During data collection, FCz and FPz served as the original reference electrode and ground electrode, respectively.
6. Instruct the participants to keep their eyes open and fixate on a cross. Throughout the process, they are required to remain conscious and alert, avoid focusing on any specific thought, and refrain from experiencing emotional fluctuations.

4. Offline data analysis

1. Pre-processing
   1. Re-reference the raw data to the average of all electrodes.
   2. Apply a 0.1-40 Hz band-pass filter to remove low-frequency drift (e.g., respiratory artifacts) and high-frequency noise (e.g., power line interference and electromyographic artifacts).
   3. Segment all subjects' data into 120 s epochs, excluding the initial and final 10 s portions of each recording to eliminate potential edge artifacts from device initialization.
   4. Remove the electrooculogram (EOG) and electrocardiogram (ECG) artifacts using the Independent Component Analysis (ICA) technique. The Infomax algorithm decomposed the signal into statistically independent components. Artifactual components were identified through tripartite characterization: (1) spatial topography (EOG showing frontopolar dipole distributions, ECG exhibiting left-temporal dominance); (2) temporal dynamics (EOG manifesting blink-locked transient spikes, ECG displaying cardiac-rhythm periodicity); and (3) spectral signatures (EOG demonstrating low-frequency spectral concentration)⁴⁴.
   5. Remove time periods during which the EEG voltage exceeded ± 75 µV as artifacts.
2. Band-specific difference analysis
   1. First, segment the EEG signals into small segments at 120 s intervals.
   2. Analyze the time-frequency characteristics of the EEG signals using a complex Morlet wavelet.
   3. Set the parameters as follows: Select wavelet transform parameters with cycles [3 0.8], 100 frequency bins (nfreqs), and 200 time points (ntimesout) to optimize time-frequency resolution for EEG analysis⁴⁵.
   4. Perform point-by-point permutation testing (10,000 iterations) across all time-frequency points to assess statistical significance (p < 0.05, FDR-corrected).
   5. Visualize as a time-frequency plot, with time points on the x-axis and frequency bins on the y-axis.

5. Model analysis

NOTE: This convolutional neural network (CNN) achieves the learning of time-frequency features of EEG signals through a multi-scale two-dimensional convolution operation⁴⁶. The process of the CNN model is shown in Figure 1B.

1. First convolutional block
   1. Employ a 2D convolution layer (Conv2d) with in_channels = 6, out_channels = 16, kernel_size = (3, 3), and padding = 1 to extract local spatial features, where the symmetric padding ensures that the spatial dimensions remain unchanged after convolution.
   2. Apply batch normalization (BatchNorm2d) to the output of the convolution layer to stabilize training and accelerate convergence, followed by a rectified linear unit (ReLU) activation function to introduce non-linearity.
   3. Use a max pooling layer (MaxPool2d) with kernel_size = (2, 2) and stride = 2 for feature compression to reduce the spatial dimensions by half while retaining important feature information.
2. Second convolutional block
   1. Utilize another 2D convolution layer (Conv2d) with in_channels = 16, out_channels = 32, kernel_size = (3, 3), and padding = 1 to further extract high-level spatial features based on the output of the first block, with the same padding strategy to maintain spatial dimensions.
   2. Similar to the first block, apply batch normalization (BatchNorm2d) and ReLU activation sequentially to normalize features and introduce non-linearity.
   3. Again, use a max pooling layer (MaxPool2d) with kernel_size = (2, 2) and stride = 2 to compress features.
3. Fully connected layers
   1. Flatten the output features from the second convolutional block into a one-dimensional vector, which serves as the input to the fully connected layers.
   2. Ensure the first fully connected layer (Linear) maps the flattened features to a 256-dimensional feature space, followed by a dropout layer (Dropout) with a rate of 0.5 to prevent overfitting, and a ReLU activation function.
   3. The second fully connected layer (Linear) further reduces the feature dimension from 256 to 128, with another dropout layer (Dropout) with a rate of 0.5 and a ReLU activation function applied sequentially.
   4. The final fully connected layer (Linear) maps the 128-dimensional features to the number of target classes (two-class), generating the output of the model.
      NOTE: The data processing platform was configured as follows: a server with four HYGON Z100L DCUs; model building and training in Python 3.8, PyTorch-1.13. The experiment uses the torch.nn module to provide core components such as Conv2d (convolution layer), BatchNorm2d (batch normalization), ELU (activation function), AvgPool2d (average pooling), Dropout (dropout layer), Linear (fully connected layer), and CrossEntropyLoss (loss function); The torch.optim module provides the Adam optimizer for parameter updates; the torch.utils.data module provides DataLoader (data loader) and TensorDataset (data set wrapper class) for data processing; torch.device is used for device configuration (DCU); torch.save is used for model parameter saving; and the sklearn.model_selection.train_test_split function from scikit-learn is used for data set splitting; Adam was selected as the optimizer; the loss function was cross-entropy; batch_size was set to 2; and epoch was set to 100. For input data, use EEG features (based on Morlet -extracted matrices) from six prefrontal channels (Fp1, Fp2, AF3, AF4, AF7, AF8) in the alpha band with time-frequency features. Divide the experiments into training and test sets using a 7:3 ratio. Evaluate model effectiveness through classification results using accuracy and confusion matrix metrics.

EEG data processing and statistical analysis

The raw EEG data were segmented into 2 s epochs centered on the event onset, consistent with standard practices in time-frequency analysis to capture transient neural dynamics while minimizing edge artifacts. Each epoch underwent continuous wavelet transformation (CWT) using a complex Morlet wavelet with 3 cycles, which optimally balances temporal and frequency resolution for detecting oscillatory activity in the theta to gamma bands.

The left panel of Figure 2 represents the exercise group, and the right panel represents the Control Group. (1) Data processing quality: Both spectra exhibit smooth curves and the characteristic neurophysiological "1/f" decay pattern (high power at low frequencies decreasing exponentially with frequency). The highly overlapping trajectories indicate effective data preprocessing (e.g., denoising, filtering) and high baseline data quality with good signal fidelity in the frequency domain. (2) Subtle inter-group differences: Within the alpha band (8-12 Hz, shaded gray area for illustration), the control group (right) shows slightly lower power values compared to the exercise group (left), suggesting that a single bout of acute exercise may have induced a mild modulatory effect on the alpha rhythm of resting-state brain oscillations.

For statistical inference, we performed pointwise nonparametric permutation tests (5,000 iterations) across all time-frequency points. This approach controls for multiple comparisons by clustering adjacent significant points (cluster-forming threshold p < 0.05, cluster-level FDR correction), to address the non-Gaussian distribution of wavelet coefficients.

Significant differences in prefrontal electrode activity were observed within the 7-13 Hz frequency band between the exercise and reading groups, as shown in Figure 3.

Validation of CNN model classification performance

In the investigation of the impact of exercise intervention on individuals with high trait anxiety, the classification performance of the Convolutional Neural Network (CNN) model using prefrontal alpha band feature data is a crucial aspect. This analysis aims to determine whether the model can effectively distinguish between the Read group and the Exercise group, thereby providing evidence for the neural-level differences associated with exercise.

The CNN model showed high classification performance when using prefrontal alpha band feature data to discriminate between the Read and Exercise groups, with an accuracy of 83.33%, and achieved an average F1 score of 0.83 and a Kappa coefficient of 0.63. To better understand the model's performance, we turn to the binary classification confusion matrix presented in Figure 3C. In this matrix, a well-structured tool for evaluating classification models, each row represents the true category of the data, and each column represents the category predicted by the model. This layout allows for a detailed assessment of the model's ability to correctly classify different data instances. The model exhibited a relatively good classification performance for both types of data. This high recognition rate implies that the model was able to accurately identify a large proportion of the data belonging to the exercise group. In other words, the neural patterns in the prefrontal alpha band associated with exercise were distinct enough for the model to recognize them with a high degree of certainty. These results from the confusion matrix further support the overall accuracy of the CNN model.

Figure 1: Resting-state EEG acquisition and CNN-based classification workflow. (A) Left side: The recording process of resting-state electroencephalogram (EEG). Right side: The EEG waveforms and the distribution of scalp electrodes. (B) The workflow of using a Convolutional Neural Network (CNN) to classify the alpha waves of two groups. Please click here to view a larger version of this figure.

Figure 2: Power Spectral Density Comparison Between Exercise and Control Groups. Left panel: the Exercise Group; Right panel: the Control Group. Please click here to view a larger version of this figure.

Figure 3: Neural dynamics and CNN classification of exercise vs. reading groups. (A) Significant differences between groups identified by point-to-point t-tests, highlighting time-frequency clusters (p < 0.05, FDR-corrected). (B) Topographic maps of grand-averaged alpha-band (7-13 Hz) power. Maps depict the spatial distribution of neural oscillatory activity for the reading group (left) and exercise group (right). (C) Classification performance of prefrontal Alpha activity using a CNN model (accuracy: 83.3%). Please click here to view a larger version of this figure.

Table 1: Participant recruitment and screening criteria.

In the realm of mental health research, understanding the underlying neural mechanisms of interventions for individuals with high trait anxiety is of paramount importance. This study was designed with the explicit aim of exploring the neural biomarkers associated with exercise intervention in such individuals through the utilization of artificial intelligence models. The adoption of advanced deep neural networks, including models like EEGNet, has revolutionized electroencephalogram (EEG) signal processing by enabling the precise analysis and decoding of brain activity with unprecedented accuracy^47,48. Thus, this study used deep learning models to classify EEG data for exercise intervention, aiming to identify neural biomarkers of high trait anxiety and explore the intervention's impact on these markers. Participants with high trait anxiety were divided into an experimental group and a control group; their brain electrical activities were compared between exercise and reading. The results of the study were highly significant. Specifically, the prefrontal region, and particularly its α-wave activity, demonstrated remarkable accuracy in the classification tasks. This finding strongly suggests that prefrontal α waves may play a pivotal role in the regulatory effect of exercise intervention on high trait anxiety.

In the exploration of the neural mechanisms underlying emotion recognition, especially in the context of anxiety, the prefrontal α-wave activity (8-12 Hz) emerges as a key factor. This specific neural oscillatory pattern has a significant association with an individual's performance in emotion recognition tasks, and understanding its role is crucial for elucidating how emotions are processed and regulated. α Waves, reflecting brain relaxation and anxiety reduction, are crucial for emotion recognition. Their activity relates to attentional resource allocation, emotional information processing depth, and emotional response inhibition. Anxious individuals exhibit lower prefrontal α waves during emotion recognition, especially at rest. This reduction may link to their negative interpretations and hypervigilance. Studies found that the weakened prefrontal α Waves in anxious individuals are correlated with difficulties in recognizing emotional cues^49,50. Anxious individuals often overreact to threat-related information, which can interfere with the normal neural processes involved in emotion recognition. Excessive sensitivity to emotional cues (e.g., facial expressions) further impairs recognition accuracy and adaptability⁵¹.

The prefrontal α is vital for emotion regulation, particularly of negative emotions. Individuals with high trait anxiety show excessive negative emotional responses, and reduced prefrontal α waves may indicate inadequate regulation. α-wave reduction reflects overreaction to emotional stimuli and weakened inhibitory neural mechanisms, serving as a neural basis for regulation difficulties. Lee⁵² proposed that increased left prefrontal α waves correlate with better emotion regulation and fewer negative responses, while reduced right prefrontal α waves associate with poor emotion inhibition and heightened negative emotions. Chen⁵³ noted that reduced prefrontal α waves in depression relate to emotion regulation difficulties and impaired cognitive control. Depressed individuals struggle to regulate emotions, with reduced prefrontal activity exacerbating symptoms. Reduced left prefrontal α waves may reflect overreaction to negative emotions, highlighting the role of prefrontal α waves in emotion recognition and regulation for depression patients.

Exercise intervention may improve anxiety management and emotion regulation by enhancing prefrontal α waves, supporting exercise as an effective intervention for high-trait-anxiety populations⁵². Exercise intervention may enhance emotion regulation and reduce anxiety by strengthening prefrontal α waves. Reduced prefrontal α waves in anxious individuals link to overreaction to negative emotions and deficient recognition/regulation. Weakened prefrontal α waves correlate with emotion regulation difficulties and impaired cognitive control⁵⁴. This study provides new insights into exercise intervention's neurobiological mechanisms for high trait anxiety and theoretical support for its clinical application as a non-pharmacological method.This finding also aligns with and corroborates the effectiveness of single, acute aerobic exercise sessions as established in previous research. Exercise improves neurobiochemical levels, inhibits inflammation, regulates the neuroendocrine system, and enhances neural plasticity, alleviating depression⁵⁵. As A single session of acute aerobic exercise enhances activity and network connectivity within the hippocampus and parietal lobe. These findings align with Li (2018), who demonstrated that physical exercise alleviates depression and anxiety in college students, noting sports dance as most effective for depression and badminton for anxiety (Li, 2018)⁵⁶ Acute aerobic exercise shows significant benefits in treating emotional disorders.

The effectiveness of deep learning models as a testing tool is also supported by prior research. Qi & Zhang (2024)⁵⁷ highlighted that deep learning models offer new approaches to aphasia's neurocomputational mechanisms. Breedlove et al. (2020)⁵⁸ used deep generative networks to simulate the brain's top-down imagery mechanism, predicting brain responses to imagery. Other studies confirm deep learning's application in neuroimaging. For example, Jiang et al. (2022) used deep convolutional neural networks (CNNs) to identify brain activity patterns in cognitive tasks⁵⁹. Amin (2023) analyzed fMRI data via deep learning, revealing functional connections during emotion regulation and providing new biomarkers for emotional disorder treatment⁶⁰.

These studies demonstrate deep learning's power in neurocomputation, offering tools for cognitive neuroscience, especially in emotion, memory, and disorder treatment. This study used the CNN model to analyze exercise intervention data, finding that CNN distinguishes brain electrical activity changes. Prefrontal α-wave data showed higher accuracy in differentiating exercise intervention vs reading states than resting-state EEG data, suggesting prefrontal α waves play a key role in exercise-induced brain activity changes, especially in emotion regulation and cognitive control. Thus, prefrontal α waves may serve as neural markers to monitor exercise intervention's impact on brain function, providing neurobiological evidence for EEG-based personalized treatment.

The study's sample size (e.g., N = 40) may limit the statistical power to detect subtle but clinically meaningful effects, particularly in subgroup analyses (e.g., age- or pathology-specific responses). Small samples increase the risk of Type II errors (false negatives) and reduce the reliability of effect size estimates, as random variability can disproportionately influence results⁶¹. While our cross-validation approach mitigates some risks, larger cohorts are needed to validate findings across diverse populations (e.g., clinical vs. healthy, different age groups) and account for inter-subject variability in EEG dynamics⁶².

The convolutional neural network (CNN) architecture, though effective for feature extraction, is susceptible to overfitting given the limited EEG dataset size. Overfitting may lead to inflated performance metrics during training that fail to generalize to independent datasets, as CNNs can memorize noise or subject-specific artifacts rather than biologically relevant patterns⁶³. Although we employed PCA for dimensionality reduction (retaining 95% variance) and dropout regularization (p = 0.5), the model's performance on external datasets remains uncertain. Hybrid approaches (e.g., wavelet-CNN fusion with synthetic EEG augmentation) could improve robustness in future work⁶⁴.

The study's focus on eyes-open resting-state EEG may limit direct applicability to other paradigms (e.g., task-based or eyes-closed recordings). For instance, alpha-band suppression in eyes-open conditions could obscure biomarkers typically observed in eyes-closed protocols (e.g., default mode network activity)⁶⁵. Additionally, participant demographics (e.g., young adults) and controlled laboratory settings may not fully represent real-world scenarios or clinical populations with comorbidities⁶⁶. External validation in multi-center studies, using standardized acquisition protocols, is essential to confirm the method's broader utility