Advanced Trans-EEGNet Deep Learning Model for Hypoxic-Ischemic Encephalopathy Severity Grading

Abstract

Hypoxic-ischemic encephalopathy (HIE) is a brain injury condition that poses a significant risk to newborns, potentially causing varying degrees of damage to the central nervous system. Its clinical manifestations include respiratory distress, cardiac dysfunction, hypotension, muscle weakness, seizures, and coma. As HIE represents a progressive brain injury, early identification of the extent of the damage and the implementation of appropriate treatment are crucial for reducing mortality and improving outcomes. HIE patients may face long-term complications such as cerebral palsy, epilepsy, vision loss, and developmental delays. Therefore, prompt identification and treatment of hypoxic-ischemic symptoms can help reduce the risk of severe sequelae in patients. Currently, hypothermia therapy is one of the most effective treatments for HIE patients. However, not all newborns with HIE are suitable for this therapy, making rapid and accurate assessment of the extent of brain injury critical for treatment. Among HIE patients, hypothermia therapy has shown better efficacy in those diagnosed with moderate to severe HIE within 6 h of birth, establishing this time frame as the golden period for treatment. During this golden period, an accurate assessment of HIE severity is essential for formulating appropriate treatment strategies and predicting long-term outcomes for the affected infants. This study proposes a method for addressing data imbalance and noise interference through data preprocessing techniques, including filtering and SMOTE. It then employs EEGNet, a deep learning model specifically designed for EEG classification, combined with a Transformer model featuring an attention mechanism that excels at capturing long-term sequential features to construct the Trans-EEGNet model. This model outperforms previous methods in computation time and feature extraction, enabling rapid classification and assessment of HIE severity in newborns.

1. Introduction

Hypoxic-ischemic encephalopathy (HIE) is one of the most common causes of neonatal brain disease. In high-income countries, three to five out of every 1000 newborns are affected by HIE [1]. Newborns with HIE face risks of mortality and cerebral palsy; even if they survive, they may suffer from long-term neurological conditions such as epilepsy, sensorineural hearing loss, vision loss, and developmental delays [2]. The initial symptoms of HIE include brain hypoxia and impaired blood supply, which subsequently lead to other neurological injuries. HIE is a progressive condition that develops over time [1].

Clinical and imaging studies have shown that hypothermia therapy has a neuroprotective effect on the brain. Currently, therapeutic hypothermia (TH) has become the standard treatment for newborns with moderate to severe HIE [3]. However, the window for hypothermia therapy is time limited, with the optimal treatment initiation within 6 h after birth [2]. Current diagnostic methods primarily rely on clinical diagnosis and other clinical markers to assess the severity of brain injury post-birth [1]. However, related studies indicate that diagnosis based on the neonatal clinical status and the degree of metabolic acidosis is insufficiently accurate in predicting the long-term prognosis of newborns [4].

In research on the grading of HIE using electroencephalograms (EEGs), Stevenson et al. [5] proposed a method that extracts non-stationary features based on time-frequency distribution (TFD). This method generates TFDs over short epochs (64 s, 50% overlap) and extracts instantaneous amplitude and frequency features. Combined with more standard quantitative EEG features, these features are used to classify HIE severity using a multi-class linear discriminant classifier.

Regarding the classification of HIE, Matić et al. [6] proposed a tensor-based method that adaptively segments continuous EEG data and extracts three features from each segment to construct a tensor. This tensor is then dimensionally reduced using a multi-dimensional decomposition method and classified using a least-squares support vector machine (SVM). Subsequently, Matić et al. [7] introduced a detection method based on dynamic inter-burst intervals (dIBIs), combined with an SVM classifier, to quantify the EEG of neonates with HIE. Ahmed et al. [8] proposed a supervised learning approach using a Gaussian Mixture Model (GMM) and SVM to assess the severity of HIE. Raurale et al. [9] employed convolutional neural networks (CNNs) and introduced a classification framework that combines quadratic time-frequency distributions with CNNs.

Given the complexity and time-consuming nature of continuous EEG monitoring and interpretation, which can only be assessed by specialized neurophysiologists, the use of computational algorithms for automatic EEG grading can overcome the challenges of conducting extensive and continuous EEG monitoring for all suspected HIE neonates in the neonatal intensive care unit (NICU). This allows neonatologists to assess brain function in a simple and timely manner [9]. Convolutional neural networks (CNNs) have succeeded significantly in various medical imaging tasks [10]. Thus, this study aims to classify the severity of HIE using novel deep learning models. The main objectives of this study include improving the accuracy of EEG-based grading, supporting clinical decision-making, advancing EEG-based analysis, and further promoting broader applications in the field of medical imaging.

Previous studies on HIE classification have primarily utilized basic neural network architectures [9]. However, as a time-series data type, EEG signals contain crucial information for accurate classification in both temporal and spatial dependencies between time points and different electrodes [11]. This study plans to use a Transformer model known for capturing long-range dependencies with an inbuilt attention mechanism that effectively extracts temporal features [12]. On the other hand, EEGNet is a neural network designed explicitly for EEG analysis, featuring depthwise separable convolutional structures, which have demonstrated superior classification performance over other models in previous studies. This model offers advantages in EEG feature extraction while maintaining high computational efficiency [13]. Therefore, this study will integrate the Transformer and EEGNet models to propose an advanced Trans-EEGNet deep learning model, aiming to develop a framework for grading the severity of HIE that surpasses existing methods in computation time and feature extraction.

2. Literature Review

2.1. Neonatal Hypoxic-Ischemic Encephalopathy (HIE)

Neonatal hypoxic-ischemic encephalopathy (HIE) is clinically defined as a condition affecting fetuses at a gestational age greater than 36 weeks, caused by hypoxia and asphyxia due to abnormal umbilical blood flow [14]. HIE typically occurs during labor, making it an unpredictable emergency [15]. The progression of HIE can be divided into two stages: the first stage involves brain injury caused by hypoxia and ischemia, known as primary energy failure [2], characterized by reduced cerebral blood flow and oxygen supply [3]. As secondary energy failure and the accumulation of reactive metabolites occur, the condition worsens, leading to the second stage, which results in delayed brain injury [2]. Pathophysiologically, HIE encompasses a series of processes caused by excitotoxicity and oxidative injury, ultimately leading to cell death [3]. These pathological features indicate the extent of brain damage in HIE, and the severity of the damage is reflected in the grading system for HIE.

For grading, HIE is typically classified into mild, moderate, and severe categories based on the Modified Sarnat Encephalopathy Grade (MSEG) [16]. According to the study by Horn et al. [17], infants with mild HIE may exhibit abnormal alertness, average activity levels, and exaggerated Moro reflexes while maintaining normal autonomic function. Symptoms of moderate HIE include lethargy, reduced activity, hypotonia, diminished primitive reflexes, constricted pupils, bradycardia, or hyperventilation. When HIE progresses to the severe stage, symptoms may include coma, lack of spontaneous activity, severe muscle weakness, absence of primitive reflexes, unresponsive pupils, or episodes of apnea.

Neonatal hypoxic-ischemic encephalopathy (HIE) is one of the leading causes of neonatal mortality and disability worldwide, accounting for 23% of global infant mortality [14]. In a small study involving 28 full-term infants born between 1980 and 1982, those with severe perinatal asphyxia and moderate or severe encephalopathy exhibited multi-organ dysfunction involving the lungs, central nervous system, kidneys, heart, as well as metabolic and hematologic systems [18]. Over the past two decades, the incidence of neonatal brain injury has not significantly decreased in developed countries [14]. In high-income countries, the incidence of HIE is approximately two cases per 1000 live births, while in low- and middle-income countries, the incidence is higher [19].

Before the widespread adoption of therapeutic hypothermia, studies on the outcomes of infants with perinatal asphyxia and HIE indicated that the disability rate for infants with moderate encephalopathy ranged from 6% to 21%. In contrast, for those with severe encephalopathy, the disability rate was as high as 42% to 100% [18]. Infants with severe HIE have a very high mortality rate, and even those who survive without functional deficits often face challenges with cognitive deficits. Even in the absence of severe disabilities, these children may exhibit mild motor dysfunction and tend to enroll in school later than their peers. Moreover, compared to children who were healthy at birth, infants with moderate and severe HIE show a significantly higher need for educational support [18].

The most effective emerging therapy for infants with HIE is moderate therapeutic hypothermia. Hypothermia therapy can reduce free radical and glutamate levels, decrease oxygen demand, and inhibit apoptosis, making it an effective treatment method [20]. Shankaran et al. [21] demonstrated that reducing brain temperature by 2 °C to 5 °C can provide neuroprotection for neonates. Brain cooling has improved various pathological processes that cause brain injury, including excitatory amino acids, cerebral energy state, cerebral blood flow, metabolism, nitric oxide production, and apoptosis. For full-term infants with severe acidosis at birth or requiring continuous ventilation for resuscitation, therapeutic hypothermia should be administered if moderate or severe encephalopathy is detected within the first 6 h after birth [3]. Hypothermia therapy has transformed the clinical management of HIE and is now considered the standard treatment for moderate to severe HIE in high-income countries [15].

2.2. Research on Convolutional Neural Networks in Electroencephalography

Ranjani and Supraja [22] proposed a method based on Deep Convolutional Neural Networks (DCNNs) for classifying electroencephalogram (EEG) signals related to autism and epilepsy. This method first converts EEG signals into power spectral density energy diagrams (PSDEDs) and then automatically applies DCNNs to extract features from the PSDED. Using a fully connected layer, the signals are classified into autism, epilepsy, and regular groups. Experimental results indicate that the proposed model performs well in terms of classification.

Wang et al. [23] introduced an EEG emotion recognition method combining CNNs with a spatiotemporal attention mechanism. The method first uses CNNs to extract local features from EEG signals and then applies the spatiotemporal attention mechanism to focus on emotional regions. The experimental results demonstrate that this approach effectively enhances the accuracy of EEG-based emotion recognition.

Li et al. [24] proposed a motor imagery EEG classification algorithm based on a CNN-LSTM feature fusion network. This method utilizes a parallel connection of CNNs and LSTM to capture the spatiotemporal features of EEG signals simultaneously. Subsequently, intermediate features are extracted through convolutional layers, and a flattening layer is added to merge all features into a fully connected layer, thereby improving classification accuracy.

Li et al. [25] proposed a multi-scale one-dimensional convolution model (MS-FTSCNN) designed to capture and integrate EEG signal frequency and temporal and spatial domain features. This method effectively fuses features from the three domains to enhance emotion recognition accuracy. Additionally, through one-dimensional convolution and reduced network parameters, the model can more precisely process the frequency domain features of emotion-related EEG signals. A summary of related research is presented in Table 1.

2.3. Related Research on Attention Mechanisms

2.3.1. Multi-Head Attention

Xie et al. [11] proposed a Transformer-based framework that combines deep learning networks with spatiotemporal information for raw EEG classification. This framework first uses convolutional neural networks (CNNs) to extract spatiotemporal features and then encodes these features into continuous representations through multi-head attention. Finally, a classifier categorizes the encoded representations into different classes. Experimental results indicate that this framework outperforms traditional deep learning methods in raw EEG classification.

Xin et al. [26] introduced an Attention-based Waveform Convolutional Neural Network (AMWCNN) for epilepsy EEG classification. This method decomposes EEG signals into different frequency components using a waveform convolution layer to extract features and then employs a multi-head attention mechanism to emphasize essential features in the EEG, thereby improving classification performance.

Wei et al. [27] proposed a Transformer Capsule Network (TC-Net) for EEG-based emotion recognition. TC-Net incorporates a novel patch-merging strategy within the EEG Transformer architecture to enhance the capture of local features. It also integrates an Emotion Capsule module to refine features for emotion state classification. This method effectively extracts global information from EEG signals and achieves state-of-the-art performance on emotion recognition datasets DEAP and DREAMER.

Mulkey et al. [28] presented a deep learning approach based on Vision Transformers for predicting delirium in critically ill elderly patients. This method converts EEG signals into two-dimensional images and trains a Vision Transformer model on these images to learn delirium-related patterns from EEG data. Table 2 summarizes the related research.

2.3.2. Single-Head Attention

Merity et al. [29] explored the application of single-head attention in RNN models. The study found that single-head attention can significantly reduce the consumption of computational resources and memory when capturing different parts of the input compared to multi-head attention. While Transformer models typically use multi-head attention to enhance feature extraction capabilities, incorporating single-head attention in specific model layers can maintain model performance while improving inference speed and memory efficiency. The results indicate that single-head attention performs well even in resource-constrained environments.

Michel et al. [30] validated the practical effectiveness of multi-head and single-head attention in Transformer models. The study pointed out that although Transformer models use multi-head attention to capture different parts of the input, removing many attention heads during training does not significantly impact model performance. The research found that single-head attention in specific layers can maintain model performance while improving computational speed and memory efficiency. Therefore, the study concluded that the advantages of multi-head attention are primarily in training stability, while single-head attention remains effective in feature extraction.

Liu et al. [31] empirically compared multi-head and single-head attention during Transformer training. The study revealed that incorporating appropriate initialization, regularization, and multi-layer single-head attention in the model effectively captures dependencies across multiple positions, offering efficiency advantages over multi-head attention. After improving stability, deep single-head Transformers achieved better performance. The research demonstrated that using multi-layer scaled dot-product attention helps extract richer features, thereby enhancing the model’s overall performance. Table 3 summarizes the related research.

3. Materials and Methods

3.1. Dataset

The dataset used in this study consists of 77 multi-channel electroencephalogram (EEG) recordings, each lasting for 1 h. Each recording is graded based on the severity of background abnormalities. Two neonatal EEG experts independently scored each segment, and in cases of disagreement, they jointly reviewed the EEG to reach a consensus. The grading system evaluates various attributes of the EEG, including amplitude and frequency, continuity, sleep–wake cycle, symmetry and synchrony, and characteristics of abnormal waveforms. According to the HIE classification criteria proposed by Murray et al. [32], the EEG recordings are categorized into four levels: normal or mildly abnormal (Grade 1), moderately abnormal (Grade 2), severely abnormal (Grade 3), and inactive (Grade 4), as shown in Table 4.

3.1.1. Research Subject

The data for this study was collected from 53 term neonates diagnosed with hypoxic-ischemic encephalopathy (HIE) in the neonatal intensive care unit of Cork University Maternity Hospital, Ireland. A total of 53 EEG recordings were collected from these infants. The recording of these EEG data received approval from the Cork Research Ethics Committee at the Cork University Teaching Hospitals. After obtaining written informed consent from guardians or parents, the Cork Research Ethics Committee approved the release of the anonymized dataset, which is publicly available at https://zenodo.org/records/7477575#.Y--W-nZBy5d (accessed on 10 January 2023).

3.1.2. Data Acquisition

This study focused on patients clinically determined to be at risk of seizures, at a gestational age between 36 and 44 weeks, and admitted to the neonatal intensive care unit (NICU). After obtaining written consent from guardians or parents, EEG recordings were conducted between January 2011 and February 2017. The data were collected from eight neonatal centers across Ireland, the Netherlands, Sweden, and the United Kingdom. From this cohort, 284 neonates with a clinical diagnosis of hypoxic-ischemic encephalopathy (HIE) and at least 6 h of valid EEG recordings were selected, while 18 were excluded due to comorbid diagnoses.

472 neonates were initially recruited for the study, with 284 ultimately selected. Their inclusion was primarily due to the presence of clinically confirmed HIE symptoms and valid EEG recordings. The remaining 188 neonates were excluded from the study. Of these, 18 were excluded due to comorbid diagnoses, 68 infants were reserved for a validation set for future EEG algorithm development, and 17 were excluded because EEG recordings were not conducted within 48 h after birth. Among the remaining 181 neonates, 54 with consent for data sharing were included, with their EEGs recorded at Cork University Maternity Hospital in Ireland. After thorough inspection of the EEG quality, one neonate with a low-quality recording was excluded. The final sample consisted of 53 neonates, with a median gestational age of 40 weeks and a male proportion of 62%, of which 58% received therapeutic hypothermia. Figure 1 shows the EEG diagram of HIE.

3.2. Pre-Processing

This study first addressed the issues of outliers and missing values in the dataset through data cleaning, resulting in a refined dataset comprising 76 hourly records. Among these, 46 records were classified as HIE Grade 1 (mild abnormalities), 12 as HIE Grade 2 (moderate abnormalities), and 9 each as HIE Grade 3 (significant abnormalities) and Grade 4 (inactive) (Figure 2). The following sections will provide a detailed explanation of the preprocessing methods used in this study. Section 3.2.1 will describe the methods for handling the imbalanced dataset, while Section 3.2.2 will introduce the filtering techniques.

3.2.1. Handling Imbalanced Datasets

Handling imbalanced datasets is a common challenge in machine learning, as the efficiency of predictive models can significantly decrease when data are imbalanced [33]. When dealing with medical data, imbalanced datasets are a common challenge. Traditional oversampling methods may lead to overfitting, as repeated data cannot provide additional information to the model. SMOTE (Synthetic Minority Over-sampling Technique) addresses this issue by generating new synthetic samples, enriching the distribution of the minority class, and reducing the risk of overfitting. In medical datasets, minority classes often contain critical features. By creating new samples based on the distribution of the original minority data, SMOTE preserves data diversity and enhances the model’s generalization capability. Thabtah et al. [34] highlighted that SMOTE (Synthetic Minority Over-sampling Technique) performs better than under-sampling methods, especially in larger datasets with unclear decision boundaries.

The Safe-level SMOTE method assigns a safety level to each positive sample before generating synthetic instances. The generated synthetic instances are closer to samples with the highest safety levels, ensuring that all synthetic samples are created only within safe regions. This approach can improve the predictive performance of classifiers for the minority class [35]. Given that previous studies have successfully used Safe-level SMOTE to address the class imbalance in EEG datasets related to adult anxiety states, this study intends to use Safe-level SMOTE to resolve the imbalance issue in the dataset.

The main steps of Safe-level SMOTE involve first calculating the Safe-level value for each minority class sample, which is defined by the number of its k-nearest neighbors that belong to the minority class. Next, for each minority class sample, one of its k-nearest neighbors that belongs to the minority class is selected as a reference point to generate new synthetic samples. Finally, these steps are repeated for all minority class samples in the dataset until the minority class reaches the desired proportion. Safe-level SMOTE is particularly suitable for applications such as medical data analysis or financial fraud detection, where misclassification costs are high. It focuses on generating synthetic samples in safe regions of the feature space, enabling more accurate classification of minority classes.

After applying the Safe-level SMOTE method and processing the data, the study balanced the number of samples across the four levels, resulting in 184 hourly records. HIE Grade 1 remained at 46 records, while SMOTE-balanced HIE Grades 2, 3, and 4 each contained 46 records, as illustrated in the bar chart below (Figure 3).

3.2.2. Filtering

Due to electroencephalogram (EEG) data characteristics, multi-electrode recordings and high-frequency sampling generate a large number of data, leading to substantial noise and outliers. Many of these data points may be irrelevant to the task, causing the model to learn patterns from this noise. Focusing on essential features during training is challenging, leading to overfitting in EEG signal analysis [36]. As a result, the trained model may need more generalization capability, resulting in suboptimal performance when testing on new data. Reducing the sampling frequency is a standard method for avoiding overfitting [37].

Pan et al. [38] noted that the computational complexity of neural networks increases as the input data dimension grows. Downsampling filtering methods based on reducing the sampling frequency can significantly decrease the subsequent computational complexity of the model. Therefore, this study downsampled the data to 40 Hz. Since each record has a total duration of one hour, with an original sampling frequency of 200 Hz, the raw data comprised 720,000-time steps. After downsampling to 40 Hz, each time step is 0.03 s, reducing the number of time steps to 120,000 per record.

3.2.3. Data Segmentation

Each one-hour record was segmented into five-minute segments with a 50% overlap in time steps, resulting in 23 five-minute segments per hour of data, as illustrated in the figure below. After oversampling, 184 one-hour records were obtained, which, after segmentation, yielded 4232 five-minute data segments (Figure 4).

In the Trans-EEGNet model, 10-fold cross-validation will be used to evaluate the model. The training-to-validation ratio is set at 8:2. The training set is divided into ten subsets. For each training fold, a portion of the subsets is randomly selected as the validation set, while the remaining subsets are used as the training set for model training. Upon completion of training, the average value from the 10-fold cross-validation will be used as the result for each performance evaluation, which will then be compared to the baseline system (Figure 5).

3.3. Research Methods

3.3.1. EEGNet

EEGNet is one of the most popular and successful deep learning architectures in BCI (brain–computer interface) applications. It utilizes separable and depthwise convolution to build EEG classification models and incorporates well-known EEG feature extraction concepts from the BCI field [39]. EEGNet is a CNN-based compact network inspired by the Xception architecture. It replaces traditional square convolutions with depthwise separable convolution structures, encapsulating temporal and spatial filters and significantly reducing the number of parameters. The depthwise separable convolution layers learn the temporal patterns for each channel and the correlations between different channels [40].

The effectiveness of EEGNet has been demonstrated in various BCI tasks [13]. Compared to other deep learning architectures commonly used for BCI classification, EEGNet requires fewer trainable parameters while achieving higher decoding accuracy and shorter training times [41]. EEGNet’s strength lies in its ability to learn diverse and interpretable features across various BCI tasks [13].

The basic architecture of EEGNet is primarily divided into two blocks. In Block 1, there are two sequentially arranged convolutional layers. The first layer captures feature maps of different frequency bands from the input EEG signals with a kernel size of (1, Fs/2), where Fs represents the sampling frequency. The next layer performs depthwise convolution with a kernel size of (C, 1), where C is the number of channels. This layer is not fully connected to the previous feature maps, reducing the number of trainable parameters and explicitly enabling the network to learn spatial filters for each temporal filter. The output of Block 1 is frequency-specific spatial feature maps [40].

In Block 1, two convolution steps are performed: first, a 2D convolution layer (Conv2D), followed by a depthwise convolution layer (DepthwiseConv2D). Batch normalization is applied after both the Conv2D and DepthwiseConv2D layers. The Conv2D layer focuses on extracting feature maps of different frequency bands from the input EEG signals, with a kernel size of (1, Fs/2), where Fs represents the sampling frequency. The subsequent depthwise convolution layer uses the same kernel size (C, 1), where C is the number of channels. This layer is not fully connected to the previous feature maps, which reduces the number of trainable parameters and enables precise learning of the spatial features for each temporal filter. One advantage of depthwise convolution is reducing the parameters required for model training [40]. Not all feature maps are fully connected to the depthwise convolution layer, allowing spatial features to be learned from combining Conv2D and depthwise convolution layers, directly corresponding to each temporal filter. This approach is inspired by the Filter Bank Common Spatial Pattern (FBCSP) technique and controls the number of spatial filters by adjusting depth parameters, which need to be learned during training.

Block 2 involves separable convolution. In this block, pointwise convolution is added after applying depthwise convolution to the output of Block 1. This approach reduces the number of parameters and decouples the relationships within and between feature maps [40]. By using separable convolution, the model’s parameter requirements can be further reduced, allowing for better summarization of each feature map. The combination of pointwise and depthwise convolution aids in optimally integrating these feature maps and helps to represent feature maps based on time scales.

In Block 3, a Softmax function processes the features from the previous block for classification tasks, allowing either binary or multi-class classification depending on the nature of the dataset [40] (Figure 6).

3.3.2. Transformer

The attention mechanism was initially proposed in natural language processing (NLP), with the self-attention mechanism becoming a core component of the Transformer model. The Transformer architecture has achieved outstanding performance in translation tasks within NLP [42]. This model follows an encoder–decoder structure, modeling through the stacking of self-attention and pointwise fully connected layers. The model multiplies the input vectors with three different weight matrices to generate query vectors (Q), key vectors (K), and value vectors (V).

The Transformer is the first sequence transduction model that relies entirely on the attention mechanism to capture long-range dependencies without the need for complex recurrent neural networks (RNNs) or convolutional neural networks (CNNs). Therefore, in EEG research, the Transformer has an advantage in effectively extracting both spatial and temporal dependencies from EEG signals, thereby enhancing performance in classification tasks [11]. Using the Transformer to model long-range dependencies can compensate for the shortcomings of CNNs in extracting global features [42].

3.3.3. The Proposed Advanced Trans-EEGNet Deep Learning Model

The EEGNet model used in this study is a CNN-based EEG classification model with excellent local feature extraction capabilities [13]. Due to its structure, which is specifically designed for the characteristics of EEG signals, EEGNet can effectively extract features such as amplitude, phase, and phase differences across different frequency bands of EEG data, as well as relationships between adjacent time points. Additionally, EEGNet requires fewer training parameters, resulting in a faster convergence speed.

This study also employs the attention mechanism from the Transformer to extract global features. This mechanism captures long-range dependencies between different time points and dynamically weights the input features, allowing the model to focus on the critical features required for classification tasks. Its advantage lies in its ability to extract the average amplitude, average phase, and average phase difference of EEG signals over different time intervals. It can reveal long-range dependencies, effectively handling the complex time-series information in EEG signals.

Moreover, the single-head attention mechanism, when handling sustained states and memory layers, can avoid excessive matrix multiplication operations due to its simplified computation method, significantly reducing memory usage and computational resource consumption. These layers reduce computational and memory requirements without adversely affecting the model’s overall performance (Figure 7).

This study combines EEGNet with the attention mechanism of the Transformer, integrating the advantages of EEGNet in local feature extraction [13] with the strengths of the attention mechanism in global feature extraction [43]. This enables more effective capture of both local and global features in EEG data, further improving classification accuracy. The attention mechanism’s ability to capture long-range dependencies within input sequences is beneficial for analyzing the time-series information in EEG signals. Therefore, combining these two models can enhance the capability to process EEG signals.

Although EEGNet has high computational efficiency, the Transformer model is less efficient computationally, and applying self-attention mechanisms directly to frameworks like convolutional networks may only sometimes achieve the expected performance improvements [44]. Kuang and Michoski [45] noted that multi-head attention mechanisms can encounter issues of redundancy and over-parameterization when capturing information from different subspaces, with some attention heads learning similar features. This may negate the lightweight advantage of the original EEGNet model. EEGNet is a lightweight and compact model, with its primary benefit being the reduction in parameter count and increased computational efficiency. Introducing an overly complex multi-head attention mechanism could have adverse effects.

Based on this, this study proposes a novel hybrid model that combines EEGNet with a single-head attention mechanism (scaled dot-product attention). This approach retains the computational efficiency of EEGNet while leveraging the attention mechanism to capture global features, thereby improving the classification performance of EEG data. Specifically, the contribution of this study lies in integrating EEGNet’s advantage in local feature extraction with the attention mechanism’s ability to capture global features and using single-head attention to avoid the redundancy and over-parameterization issues associated with multi-head attention mechanisms. This reduces model training time and memory usage, maintaining the model’s lightweight nature and efficiency.

In summary, this study proposes a novel EEG classification model that combines EEGNet with a scaled dot-product attention mechanism. The model effectively captures both local and global features while maintaining high computational efficiency. The detailed structure of the model is shown in Figure 8.

Trans-EEGNet adopts the three-layer convolution architecture proposed by EEGNet and adds an attention layer composed of Scaled Dot-product Attention. Its formula is as follows:

(1)$\mathrm{A}\mathrm{t}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\left(\mathrm{Q},\mathrm{K},\mathrm{V}\right)=softmax\left({\displaystyle \genfrac{}{}{0pt}{}{{\mathrm{Q}\mathrm{K}}^{\mathrm{T}}}{\sqrt{{\mathrm{d}}_{\mathrm{k}}}}}\right)\mathrm{V}$

Q, K, and V represent the query vector (Q), key vector (K), and value vector (V), respectively. The softmax function is applied to compute the attention weights, converting the raw attention scores into attention weights, thereby creating a valid probability distribution that sums to 1. The overall model is divided into two blocks, with a detailed explanation of each layer’s architecture provided below.

First, the initial block of the model performs preliminary feature extraction through a 2D convolution. The number of convolutional kernels is 32, with a size of (1, 64), and the activation function used is the Exponential Linear Unit (ELU). The padding mode is set to “valid”, meaning no padding is applied to the input data during the convolution operation. As a result, the convolutional kernels only compute over the valid portions of the input data without extending beyond the data boundaries. The advantage of using the ELU activation function lies in its ability to produce small negative values instead of zeros for negative inputs, thus helping to avoid the vanishing gradient problem. This improves the model’s learning capability and accelerates convergence. Its definition is as follows:

(2)$\left\{\begin{array}{c}x,ifx>0\\ \alpha \left(e^x-1\right),ifx\le 0\end{array}\right.$

The Exponential Linear Unit (ELU) function features negative outputs with a smooth transition around zero, aiding faster convergence and improved model performance. In contrast, the ReLU function outputs zero for negative inputs, leading to the “dead neuron” problem. ELU’s exponential form in the negative region retains non-zero gradients, effectively mitigating this issue. By combining the efficiency of ReLU with the smoothness of Sigmoid/Tanh, ELU demonstrates clear advantages in addressing dead zones and accelerating model convergence.

On the other hand, the output feature map is smaller than the input feature map, which reduces the computational load. This layer mainly focuses on extracting features in the time domain. After the convolution operation, batch normalization is applied to the output feature maps to accelerate model convergence and improve stability. Subsequently, a MaxPooling2D layer with a size of (2, 2) is used to further reduce the feature maps’ dimensions and computational complexity while retaining essential features.

After the initial feature extraction, this study introduces a scaled dot-product attention mechanism in the second layer. By incorporating the attention mechanism at an earlier stage, its ability to extract global features is fully leveraged. It avoids an excessive increase in the computational complexity of subsequent layers, thus achieving a balance between model performance and complexity. When applying the attention mechanism, it is necessary to reshape the feature maps into a one-dimensional form to calculate the relationships between different feature positions. Specifically, the feature maps are first padded to a four-dimensional shape with ((0, 0), (16, 17)) padding and then reshaped into a one-dimensional sequence through a Reshape layer before being processed by the attention mechanism. The attention mechanism typically requires input in a sequential format to compute the relationships between different positions.

In this layer, layer normalization is used, which stabilizes the feature distribution without relying on batch size. Since Trans-EEGNet is a deep learning model that integrates an attention mechanism with a three-layer convolutional architecture, it has a higher complexity and many learnable parameters. To prevent the model from overfitting to the training data during the training process and to maintain generalization ability, dropout regularization is employed. During training, the Dropout layer randomly drops some neurons and connections to reduce the model’s reliance on specific weights [45].

EEGNet is composed of three convolutional network types: 2D convolutional networks (Conv2D), depthwise convolutional networks (DepthwiseConv2D), and separable convolutional networks (SeparableConv2D) [13]. Therefore, the DepthwiseConv2D and SeparableConv2D convolutional layers are constructed in the second block of the model.

The depthwise convolutional network focuses on capturing frequency feature maps across different channels of EEG signals and ultimately outputs spatial feature maps characterized by frequency (time). Zero padding fills the blank areas in the attention mechanism’s output with zeros, adjusting the dimensions for subsequent convolution operations. The depthwise convolution layer then processes the resulting output. In this depthwise convolution, 16 convolutional kernels of size (3, 3) are used, with the activation function being the Exponential Linear Unit (ELU). This convolution method effectively extracts local features while maintaining computational efficiency. Depthwise convolution reduces the computational load by performing convolutions separately for each input channel, extracting features from each channel.

In the second layer of the second block, SeparableConv2D is used for computation. Separable convolution includes a depthwise convolution followed by a pointwise convolution layer. The output of the depthwise convolution is first zero-padded to align with the subsequent separable convolution operation, and then the result is processed through the separable convolution. This layer uses 8 convolutional kernels of size (8, 4), with ELU as the activation function. Due to the characteristics of the EEGNet architecture, which separates the spatial and temporal feature extraction layers, this layer requires fewer parameters. It can capture relationships within and between feature maps, outputting feature maps that emphasize the relationships between channels (spatial). On the other hand, separable convolution effectively extracts both local and cross-channel features while maintaining computational efficiency [46].

Finally, the outputs from all convolutional and pooling layers are flattened and passed through a fully connected layer (Dense) with 64 neurons, using the ELU activation function. The output layer (Dense) uses the Softmax activation function, providing the final classification result with 4 neurons. The fully connected layer efficiently classifies the extracted features, while the Softmax activation function converts the outputs into a probability distribution, making it suitable for multi-class classification. The Softmax function transforms input values into probabilities such that their sum equals 1, and its formula is as follows:

(3)$\sigma {\left(x\right)}_j={\displaystyle \frac{e^{x_j}}{{\sum }_{i=1}^C{e}^{x_i}}}whereJ=1,\dots ..c$

$\sigma \left(x_i\right)$ represents the output of the Softmax function, which is the probability value of the input $x_i$ after normalization, and c denotes the number of classes. With this design, our model can efficiently extract and combine time-frequency features from EEG signals and enhance its performance through the attention mechanism. Combining deep convolution, separable convolution, and the attention mechanism, this network architecture exhibits muscular flexibility and adaptability when processing EEG data.

3.3.4. Majority Voting

She et al. [47] pointed out that recording a full hour of EEG data provides more stable and reliable observations in monitoring HIE. In contrast, shorter recordings may be influenced by transient physiological fluctuations or interference signals, leading to less accurate data. However, directly handling an entire hour of data in model training consumes significant computational resources, especially for deep learning models. Splitting the data into smaller segments can reduce the amount of data input into the model each time, thereby decreasing the computational load and improving the efficiency of model training and inference [48]. Therefore, this study trains the model by segmenting one hour of EEG data into five-minute segments and, after training, integrates the output using a voting mechanism.

This study adopts a majority voting mechanism to ensure the final classification result’s reliability. Specifically, the labels for all 23 five-minute segments are collected, and the frequency of each label is counted. The label with the highest occurrence is selected as the final output for the one-hour data. This voting mechanism combines multiple prediction results, offsetting potential errors from individual predictions, thereby enhancing the robustness of the classification outcome. Through this voting mechanism, the model’s classification accuracy is significantly improved, and it can produce a comprehensive classification result for the one-hour EEG data (Figure 9).

4. Experimental Design

4.1. Experimental Process

The detailed research process is as follows:

1. Apply filtering to the data, reducing the original frequency from 200 Hz to 40 Hz.

2. Use Safe-level SMOTE to handle the imbalanced data classes, balancing the number of samples across all classes to reduce the potential impact of class imbalance on the classification task.

3. Segment the one-hour EEG data (5 min per segment with a 50% overlap, resulting in 23 segments per hour).

4. Randomly split the data into training and validation sets and perform 10-fold cross-validation.

5. Adjust model parameters.

6. Each model outputs 23 classification results for every hourly segment, and the final classification result is determined through a majority voting mechanism.

7. Evaluate the model using standard metrics such as accuracy, specificity, sensitivity, precision, F-score, and the Kappa value.

8. Compare the proposed model with other existing methods to validate its effectiveness (Figure 10).

4.2. Baseline System

This study aims to evaluate the performance of the Trans-EEGNet model by comparing it with other deep learning and machine learning methods. The following introduces the two baseline systems, CNNs and EEGNet, which will be used for performance comparison.

First, the CNN (Convolutional Neural Network) is a well-established neural network algorithm that has been proven to be a solid foundation for various state-of-the-art models in EEG classification research [49]. Therefore, this study includes a comparison with the CNN to demonstrate the performance improvements of Trans-EEGNet. The other baseline system is EEGNet, a compact convolutional neural network architecture designed explicitly for EEG data. EEGNet uses depthwise separable convolution operations to extract EEG features, creating a model optimized for EEG characteristics [50]. By comparing with this foundational EEGNet architecture, we can assess whether Trans-EEGNet performs better in classification tasks.

4.3. Experimental Setup

4.3.1. Hyperparameters and Environment Setup

The same hyperparameters are used for the baseline systems and the proposed model for consistency in experiments. The hyperparameters include 10-fold cross-validation (k = 10). Through a systematic grid search process, the optimal hyperparameters for the model were determined as follows: a batch size of 64, 20 epochs, a dropout rate of 0.25, and a learning rate of 0.0001. These parameters were selected to balance training efficiency and generalization capability, minimizing overfitting while ensuring adequate model convergence. The chosen dropout rate aids in regularization, while the finely tuned learning rate facilitates gradual optimization, contributing to the model’s overall performance and stability during training. Table 5 displays the hardware environment settings. In the experiment, the training time for the CNN model ranged from 10 to 20 min, while the EEGNet model required 5 to 15 min, and the Trans-EEGNet model took 30 to 60 min. After training, all three models completed inference tasks within approximately one to three seconds.

4.3.2. Evaluation Metrics

This study uses accuracy, specificity, sensitivity/recall, precision, the F-score, Kappa value, and Matthews Correlation Coefficient (MCC) to evaluate the performance of the classifier model.

Accuracy is used to measure the proportion of correctly predicted samples out of the total number of samples. It is defined as follows:

(4)$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}\left(\mathrm{A}\mathrm{c}\mathrm{c}\right)={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$

TP (True Positive) means the number of “positive classes” correctly predicted by the model, and TN (True Negative) means the number of “negative classes” correctly predicted by the model.

Specificity is the true negative part of the correct identification. The equation is

(5)$\mathrm{S}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{t}\mathrm{y}\left(\mathrm{S}\mathrm{P}\mathrm{E}\right)={\displaystyle \frac{TP+TN}{TP+FP}}\times 100$

Sensitivity measures the proportion of correctly predicted positive samples to the total number of positive samples. It is defined as

(6)$\mathrm{S}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{i}\mathrm{t}\mathrm{y}\left(\mathrm{S}\mathrm{E}\mathrm{N}\right)={\displaystyle \frac{TP}{TP+TN}}\times 100$

Precision is used to measure the proportion of actual positive instances classified as positive instances, defined as

(7)$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\left(\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\right)={\displaystyle \frac{TP}{TP+FP}}$

The F-score is an extension of classification accuracy, which combines precision and recall. The calculation of F-score is defined as

(8)$\mathrm{F}\text{-}\mathrm{score}=2\times {\displaystyle \frac{Precision\times Sensitivity}{Precision+Sensitivity}}$

Kappa is used to evaluate the consistency of the classification model. It is defined as the consistency ratio between the actual classification results and the random classification results. Cohen’s Kappa provides a more reliable performance measure in machine learning, where using accuracy alone can be biased when dealing with imbalanced data. Its formula is:

(9)$k={\displaystyle \frac{p_o-p_e}{1-p_e}}$

$k$ is Cohen′s Kappa coefficient, ranging from 1 to −1. $p_o$ is the observed agreement, expressed as the proportion of the annotator or model predictions that are consistent with the real labels in the actual situation, $p_e$ stands for Expected Agreement, which represents the probability that the prediction result of the model or annotator is consistent with the real label under completely random conditions.

The Matthews Correlation Coefficient (MCC) is used to evaluate the consistency between the actual classification results and random classification results, especially when dealing with imbalanced datasets. By considering true positives, true negatives, false positives, and false negatives, it provides a comprehensive and balanced measure of classification performance. The formula is as follows:

(10)$MCC={\displaystyle \frac{TP\times TN-FP\times FN}{\sqrt{(TP+FP)(TP+FN)(TN+FP)(TN+FN)}}}$

5. Experimental Results and Discussion

5.1. Experimental Performance

5.1.1. Comparison with Baseline Model

First, the performance results of each model, including the mean and standard deviation, will be compared. The stability of deep learning models can be assessed by evaluating their performance’s standard deviation (STD) and mean. Achieving high consistency in performance evaluation indicates that the model’s performance is less affected by variations in the data.

The standard deviation reflects a model’s degree of performance variation across different cross-validation folds. A minor standard deviation indicates more consistent performance across different datasets, demonstrating model stability and ensuring accurate predictions on new data in real-world applications. If the standard deviation is large, it suggests significant performance variation across datasets, which may indicate overfitting issues. Therefore, standard deviation plays a crucial role in evaluating model performance.

The mean represents the average performance of the model across all cross-validation folds, providing an overview of the model’s overall performance. In most cases, models with a higher mean performance score perform better, making it an essential criterion for model selection and tuning. This study uses standard deviation to assess the model’s stability and utilizes the mean to comprehensively evaluate various performance metrics, quantifying the model’s actual performance across different datasets.

As shown in Table 6, the performance metrics of Trans-EEGNet are as follows: accuracy 0.9549 ± 0.0160, sensitivity 0.9587 ± 0.0140, specificity 0.9841 ± 0.0050, F-score 0.9623 ± 0.0145, precision 0.9607 ± 0.0149, and Kappa value 0.9841 ± 0.0050. The performance of EEGNet is accuracy 0.8819 ± 0.0278, sensitivity 0.8871 ± 0.0276, specificity 0.9588 ± 0.0097, F-score 0.8883 ± 0.0273, precision 0.8975 ± 0.0252, and Kappa value 0.8383 ± 0.0352. The performance of CNN is accuracy 0.6937 ± 0.1973, sensitivity 0.6836 ± 0.2221, specificity 0.9048 ± 0.0692, F-score 0.6609 ± 0.2108, precision 0.7253 ± 0.2108, and Kappa value 0.5821 ± 0.2667.

Based on the average performance results from 10-fold cross-validation, Trans-EEGNet achieved the best performance across all metrics. Accuracy represents the proportion of all data correctly classified and is a crucial indicator of overall model performance. Sensitivity reflects the model’s ability to capture true positive samples; a high sensitivity indicates that the model can identify most positive samples, making it helpful in determining if the model mistakenly classifies actual positive samples as negative. Specificity measures the proportion of negative samples correctly identified by the model; a high specificity indicates that the model seldom misclassifies actual negative samples as positive.

Precision refers to the proportion of true positive samples among those predicted as positive by the model. High precision means that the model’s positive predictions are more reliable, meaning that most samples predicted as positive are indeed positive. F-score combines precision and sensitivity to evaluate the model’s balanced ability to correctly classify positive samples and avoid misclassifying negative samples as positive. Cohen’s Kappa measures the consistency of a classification model, considering not only accuracy but also the possibility of random agreement. A high Kappa value indicates high consistency in the model’s classification results. It is significantly better than random guessing, making it a good indicator for assessing the model’s prediction accuracy, consistency, and reliability.

Among all the evaluation metrics mentioned above, Trans-EEGNet achieved the best performance, indicating more excellent reliability and accuracy in the classification task. Additionally, according to Anguita et al. [51], when performing cross-validation, using a lower loss as the model’s performance evaluation criterion can help achieve better generalization performance and minimize the error range between the model and the predicted data. The results in the table show that Trans-EEGNet’s generalization and adaptability are superior to those of the other two baseline models.

In summary, the Trans-EEGNet model offers a more comprehensive approach to classification and performs better across all metrics compared to other baseline models. Therefore, this study selects the proposed Trans-EEGNet model for HIE classification and serves as the basis for the subsequent voting mechanism.

To further verify the model’s statistical significance, an analysis of variance (ANOVA) was conducted, as shown in Table 7. The table shows that the model’s p-value (6.67 × 10⁻⁵) is far smaller than 0.05, indicating that the model is statistically significant. Additionally, the F-value (14.02) is relatively large, suggesting that the variance between models is significantly greater than within groups. This further confirms the effectiveness of Trans-EEGNet in extracting EEG features.

Since the ANOVA analysis results indicated that the variance between models is significantly greater than within groups, this study used a multiple comparison test (Tukey’s HSD) to compare the performance differences between different models. Table 8 presents the results of the multiple comparison test. The results show that Trans-EEGNet is significantly superior to the other models in all comparisons, indicating that this model has a significant advantage in classifying EEG data. Additionally, significant differences were observed between all three models.

Table 8 presents the results of Tukey’s HSD test for each model. This test evaluates the performance of two machine learning models in each comparison and examines the significance of the mean differences between the models. In the table, Group 1 and Group 2 represent the two model combinations being compared. The Mean Difference indicates the difference in mean absolute error (MAE) between the two models. P-adj is the adjusted p-value used to test whether the mean difference is significant. Lower and Upper represent the lower and upper bounds of the 95% confidence interval. Significant indicates whether to reject the null hypothesis; if it is “Yes”, it means that there is a significant difference between the means of the two groups.

Tukey’s HSD test results show that the mean differences between CNN and EEGNet, CNN and Trans-EEGNet, and EEGNet and Trans-EEGNet all reach a significant level (p-value smaller than 0.05). This indicates that there are statistically significant differences in performance among these models.

Mean Absolute Error (MAE) directly indicates the model’s prediction accuracy, with smaller values indicating higher accuracy. The Mean Difference compares the relative performance difference between the two models. When comparing model performance, the conclusions should be based on the actual performance metric of the models, namely MAE. According to the test results, the Trans-EEGNet model exhibits significantly lower MAE in these comparisons, indicating that its prediction performance is superior to the other models.

5.1.2. Majority Voting Result

This study uses the output of Trans-EEGNet as the input for the majority voting mechanism, with the classification results of Trans-EEGNet shown in Figure 11. A total of 437 five-minute data samples were classified. As observed in Figure 11, the classification performance for Grade 3 is excellent, with the vast majority of samples correctly classified and only one sample misclassified as Grade 4. The classification performance for Grade 4 is also good, with most samples correctly classified and only three samples misclassified as Grade 3. However, the classification errors for Grade 1 and Grade 2 are relatively higher, with most misclassifications categorized as Grade 3. The final classification accuracy reached only 0.7849, as shown in Table 9.

To improve the classification accuracy, this study combines the 23 five-minute segments within each hour using a majority voting mechanism for classification integration, aiming for better classification performance.

In the 437 five-minute data samples test set, the 23 consecutive segments for each subject were combined into one-hour data. The model classified these 23 segments, and the results were processed through a majority voting mechanism to obtain the outcome. As shown in Table 10, the performance results show significant improvement compared to the original output, with accuracy increasing from 0.7849 to 0.8421. As illustrated in Figure 12, the classification accuracy for the one-hour data of 19 subjects is better than the direct output from Trans-EEGNet.

5.2. Discussion

5.2.1. Cross-Validation Visualization Results

In terms of performance evaluation, we compared the performance metrics of three models, Trans-EEGNet, EEGNet, and CNN, and presented the results for each fold using line charts. The results show that Trans-EEGNet outperforms the other two models in all metrics, including accuracy, sensitivity, specificity, F-score, precision, and Kappa value. Trans-EEGNet has a more minor standard deviation, indicating less performance variation across different datasets and demonstrating higher stability. Moreover, the loss values for each fold were significantly lower than those of the other models, further proving its stability. As a result, Trans-EEGNet also achieved good results when classifying the new test set (Figure 13, Figure 14, Figure 15, Figure 16, Figure 17 and Figure 18).

5.2.2. Voting Mechanism

According to the study by Padfield et al. [52], data segmentation and overlapping sampling can effectively increase the diversity of data samples, thereby improving the training performance of models. Therefore, in the post-processing phase, this study adopted a majority voting mechanism to integrate the predictions of different time segments. Specifically, each hour of EEG data was divided into five-minute segments, generating 23 segments with a 50% overlap. This overlapping approach helps increase the diversity of data samples, allowing the model to capture feature variations across different periods better while reducing the bias that may arise from a single segment.

After classifying each segment, the classification results of all segments were counted, and the most frequently occurring class was selected as the final prediction. This majority voting mechanism combines multiple prediction results to offset potential errors from individual predictions, thus enhancing the robustness and accuracy of the classification results. In the test set of 437 five-minute samples, the final classification results obtained by applying majority voting to the 23 consecutive samples of each subject showed a significant improvement in accuracy, from 0.7849 to 0.8421. This demonstrates that the majority voting method is highly effective in enhancing the accuracy and robustness of the final classification.

5.2.3. Comparison of Research Results on HIE Topics

In the research on HIE grading, previous scholars have proposed various methods to address this issue. Stevenson et al. [5] proposed the TFDfeat method, which first converts EEG signals into a quadratic time-frequency distribution to extract features related to time and frequency, making it particularly suitable for handling dynamic changes in non-stationary data. This method combines these features with traditional spectral power features and uses a multi-class linear discriminant classifier (MLDC) for classification. The final decision for more extended periods is made by combining the classification results of shorter segments through a majority voting approach.

Ahmed et al. [8] proposed the GSVfeat method, which combines generative and discriminative models using Gaussian Mixture Models (GMMs) to describe the statistical characteristics of the data. This method first extracts features from short segments of EEG and then generates supervisors using GMM to represent the features of the entire segment. A support vector machine (SVM) is used for grading. This approach effectively handles data with complex feature distributions but has a higher computational burden and a more complex model structure.

Raurale et al. [53] proposed the IBIfeat method, which focuses on identifying EEG inter-beat interval (IBI) features. The duration and stability of these intervals play a critical role in HIE grading. This method uses specific algorithms to detect and quantify the intervals within the signal and then uses a multilayer perceptron (MLP) for classification. Raurale et al. [9] later proposed the CNN1d method, which uses a one-dimensional convolutional neural network (CNN) to process EEG data. The neural network automatically learns features from the data and captures signal characteristics through multiple convolutional layers. The final classification is determined through a fully connected layer and a Softmax layer, with results from each time segment combined using a voting mechanism.

Performance comparisons from previous studies indicate that neural networks outperform traditional machine learning methods in EEG classification. Neural networks can automatically extract features from EEG data without relying on manually labeled features to find the optimal classification boundaries. Raurale et al. [1] further proposed a method that combines EEG signals transformed through quadratic time-frequency distribution with CNNs for feature extraction. The preprocessing through quadratic time-frequency distribution enhances the prominence of features in EEG, and combined with the feature learning capability of CNN, this approach achieved better classification accuracy.

Compared to previous studies, the Trans-EEGNet proposed in this research combines EEGNet with a single-head attention mechanism. EEGNet is a CNN architecture specifically designed for EEG data, capable of automatically extracting features. Its lightweight design helps to reduce computational costs and improve training efficiency. The single-head attention mechanism addresses the limitation of CNNs being overly focused on local features. By contrast, earlier studies did not use attention mechanisms, making them relatively weaker in capturing long-range dependencies or multi-dimensional data relationships. This makes Trans-EEGNet more comprehensive in its feature extraction capabilities.

Due to privacy concerns, most EEG medical data are not publicly available. In their study, Raurale et al. [1] used the ANSeR dataset, a neonatal seizure recognition algorithm dataset, to compare previous methods. This dataset was collected with written consent from the parents of the newborns, but the consent did not include permission for data sharing or public release. Therefore, this study cannot directly compare its results with those of Raurale et al.’s research. All the results are shown in Table 11.

5.2.4. Limitations and Future Research

This study successfully improved the classification capability of EEG data, but it also has some limitations that need further refinement in future research. For EEG data, factors such as brain anatomy, head size, electrode placement, and inherent differences between subjects greatly limit the generalizability of EEG analysis. In other words, even if a model achieves good training results on a particular dataset, its generalization ability to other datasets may need to be improved [54].

As machine learning models become increasingly complex, their interpretability has become crucial in clinical settings. Integrating explainable artificial intelligence (XAI) techniques with digital health data is advancing steadily in precision medicine, addressing the need for transparency and comprehensibility in clinical applications. Moreover, explainable machine learning algorithms demonstrate how to analyze complex patterns, providing critical insights for predictions tailored to individual patients. Another model interpretability technique, Grad-CAM (Gradient-weighted Class Activation Mapping), visualizes the reasoning behind deep learning model predictions. Generating heat maps through weighted feature maps highlights the areas of interest in the model, aiding in understanding its decision-making process. Grad-CAM can be integrated with any convolution-based architecture, such as Xception and ResNet. In medical imaging, Grad-CAM can pinpoint abnormal regions, assisting doctors in comprehending the basis of the model’s judgments [55,56]. Explainable AI can articulate the rationale behind its diagnoses or predictions, making it easier for doctors to understand and accept AI system recommendations rather than relying solely on “accuracy” metrics. In medical data, explainable AI is crucial in enhancing system transparency and credibility, supporting medical decision-making, uncovering data biases, and advancing clinical research [56]. In the future, these technologies could be further combined to identify and interpret feature patterns associated with hypoxic-ischemic encephalopathy (HIE), facilitating more precise diagnosis and treatment decisions.

The first limitation concerns the dataset. The dataset used in this study comes from a specific environment and subjects, which may limit the model’s generalization ability in other settings and among different subjects. Therefore, future research should consider incorporating more data from diverse backgrounds and conditions to enhance the model’s applicability and stability.

Furthermore, although Trans-EEGNet combines EEGNet with a single-head attention mechanism, improving classification performance, it also increases the model’s complexity and computational cost, which could be a limitation in resource-constrained applications. Future research could explore more efficient model architectures, such as lightweight deep learning models like CapsNet, to reduce computational requirements.

The convolutional layers in EEGNet have been optimized, but the introduction of Transformer has rendered the model architecture more hybridized. This design significantly enhances the ability to extract local and global features but also increases computation complexity, further extending processing time. In Trans-EEGNet, additional fusion layers are required to integrate the features of the Transformer and EEGNet and the parameters and computations of these layers also impact computational efficiency. Transformers typically have more parameters than traditional convolutional neural networks (such as the CNN and EEGNet). This is particularly evident when handling attention mechanisms and large feature matrices, where the additional weight matrices demand more memory and computational resources.

Additionally, this study only used EEG data for classification without considering the integration of other biological signals, such as the electrocardiogram (ECG) and electromyography (EMG). Multimodal data fusion can provide richer information, enhancing classification accuracy and robustness. Future research could explore combining EEG data with other biological signals for analysis to obtain more comprehensive medical classification results.

Lastly, this study primarily focused on analyzing and classifying offline data without addressing the challenges of real-time applications. Real-time applications require handling issues such as data latency, computational resource limitations, and constraints on computation time. Future research could focus on effectively deploying and operating the Trans-EEGNet model in real-time systems to achieve more efficient real-time classification.

6. Conclusions

The Trans-EEGNet model proposed in this study achieves efficient feature extraction and classification of EEG signals by combining CNN and Transformer layers. The results of 10-fold cross-validation confirm the model’s effectiveness and stability. Comparisons with baseline models further demonstrate that Trans-EEGNet exhibits superior performance in grading hypoxic-ischemic encephalopathy (HIE), providing strong support for future clinical applications. Inspired by deep learning and attention mechanisms, this study aims to use EEGNet for local feature extraction and a Transformer for global feature extraction. The proposed Trans-EEGNet integrates the strengths of both models, enhancing the ability to model and predict temporal features of EEG signals.

In summary, this study’s main contribution is the proposal of a hybrid model that combines EEGNet with a Transformer. This model achieves better grading results through the synergistic effect of feature extraction and long-term dependency learning. The designed experiments validate the model’s effectiveness, showing that Trans-EEGNet significantly outperforms other models.

Additionally, through the majority voting mechanism, this study effectively enhances the classification performance of the Trans-EEGNet model, especially in handling EEG time-series data, allowing better management of challenges posed by transient noise and variations. This approach improves the model’s robustness and provides more reliable technical support for clinical applications. Future work will focus on interpreting the obtained long-term dependency representations and exploring the application of this method to other biological time-series signals, such as the electrocardiogram (ECG).

Author Contributions

Conceptualization, D.-H.S. and M.-H.S.; Formal analysis, S.-Y.H.; Investigation, F.-I.C. and T.-W.W.; Methodology, D.-H.S., F.-I.C. and S.-Y.H.; Resources, D.-H.S. and M.-H.S.; Software, F.-I.C. and S.-Y.H.; Supervision, D.-H.S.; Validation, D.-H.S. and T.-W.W.; Visualization, M.-H.S.; Writing—original draft, S.-Y.H.; Writing—review and editing, F.-I.C., T.-W.W. and M.-H.S. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

The Cork Research Ethics Committee approved the release of the anonymized dataset, which is publicly available at https://zenodo.org/records/7477575#.Y--W-nZBy5d (accessed on 10 January 2023).

Conflicts of Interest

The authors declare no conflicts of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. EEG diagram of HIE.

Figure 2. Bar chart of HIE graded distribution of original data.

Figure 3. Bar chart of HIE grade distribution after SMOTE.

Figure 4. Schematic diagram of one-hour data segmentation.

Figure 5. Ten-fold cross validation (The blue part is the training set, and the orange part is the testing set).

Figure 6. EEGNet architecture.

Figure 7. Scaled dot product architecture.

Figure 8. Proposed Trans-EEGNet model.

Figure 9. Post-processing voting diagram.

Figure 10. Research process flow chart.

Figure 11. Trans-EEGNet classification confusion matrix.

Figure 12. Confusion matrix of majority voting results.

Figure 13. Line chart of Accuracy in cross-validation.

Figure 14. Line chart of F1 score in cross-validation.

Figure 15. Line chart of Kappa in cross-validation.

Figure 16. Line chart of Precision in cross-validation.

Figure 17. Line chart of Sensitivity in cross-validation.

Figure 18. Line chart of Specificity in cross-validation.

Table 1

Related research on EEG and CNN.

Subject	Method	Author
Classifying autism and epilepsy using EEG signals	Proposed a method for classifying autism and epilepsy using deep convolutional neural networks (DCNNs).	[22]
Emotion recognition and classification through the EEG signals	A method for EEG emotion recognition using the hybrid spatiotemporal feature fusion neural network is proposed.	[23]
Classification of motor imagery tasks through electroencephalography	A motion imagination EEG classification algorithm based on the CNN-LSTM feature fusion network is proposed.	[24]
Emotion recognition and classification through electroencephalography	A multi-scale one-dimensional convolutional model, MS-FTSCNN, was proposed to capture and integrate the frequency, temporal, and spatial domain features of EEG signals. This model effectively combines the three types of features, thereby improving the accuracy of emotion recognition.	[25]

Table 2

Related research on multi-head attention mechanism.

Subject	Method	Author
Classification judgment through electroencephalogram	Proposed a method of combining the Transformer and CNN to classify motor imagination tasks	[11]
	An attention mechanism-based waveform convolutional neural network (AMWCNN) is proposed for epilepsy EEG classification.	[26]
	The TC-Net model was proposed, which combines a Transformer for extracting EEG features with an emotion capsule module for refining features, enabling emotion recognition classification.	[27]
	A Vision Transformer was used to analyze EEG data and predict whether elderly individuals have delirium.	[28]

Table 3

Related research on single-head attention mechanism.

Subject	Result	Author
The role of single-head attention in RNN models was explored	Compared with multi-head attention, single-head attention still has good performance in resource-constrained environments	[29]
The actual performance of multi-head attention and single-head attention in Transformer models was validated	Many attention heads often focus on similar features, leading to redundancy. Study Shows Comparable or Better Performance with Fewer Attention Heads	[30]
The effectiveness of multi-head attention and single-head attention in Transformer models was validated.	Through adaptive model initialization techniques, the single-head attention Transformer model can surpass traditional multi-head attention models in terms of stability and performance.	[31]

Table 4

HIE classification.

EEG Grade (Description)	Background	IBI(s)	Features of the EEG	Sleep-Wake Cycle
0 (normal)	Continuous	—	Normal physiologic features (e.g., anterior slow waves)	—
1 (mild abnormalities)	Continuous	—	Slightly abnormal activity (e.g., mild asymmetry, mild voltage depression)	Or poorly defined
2 (moderate abnormalities)	Discontinuous	<10	Or clear asymmetry or asynchrony	Not clearly defined
3 (major abnormalities)	Discontinuous	10–60	Severe attenuation of background patterns	Or absent
4 (inactive)	Severe discontinuous	>60	Or background activity of <10 μV	Absent

Table 5

Specifications of Experimental Equipment.

Name	Model/Version
Operating System	Windows 10 22H2
Processor	Intel Core i7-13700
Memory	ADATA DDR4-3200 64 G
SSD	Kingston KC3000 1 TB
Graphics Card	Nvidia GeForce RTX 3060

Table 6

Model performance comparison.

	Accuracy	Sensitivity	Specificity	Loss
Trans-EEGNet	0.9549 ± 0.0160	0.9587 ± 0.0140	0.9841 ± 0.0050	0.1423 ± 0.0531
EEGNet	0.8819 ± 0.0278	0.8871 ± 0.0276	0.9588 ± 0.0097	0.8855 ± 0.4580
CNN	0.6937 ± 0.1973	0.6836 ± 0.2221	0.9048 ± 0.0692	0.5829 ± 0.2371
	F-Score	Precision	Kappa
Trans-EEGNet	0.9623 ± 0.0145	0.9607 ± 0.0149	0.9366 ± 0.0185
EEGNet	0.8883 ± 0.0273	0.8975 ± 0.0252	0.8383 ± 0.0352
CNN	0.6609 ± 0.2108	0.7253 ± 0.2108	0.5821 ± 0.2667

Table 7

Model ANOVA analysis result.

Source	Sum of Squares	Degrees of Freedom	F-Value	p-Value
Model	14.02	2	14.02	6.67 × 10−5
Residual		27

Table 8

Tukey’s HSD test.

Group 1	Group 2	Mean Difference	p-adj	Lower	Upper	Significant
CNN	EEGNet	0.1872	0.0028	−0.2463	0.0724	Yes
CNN	Trans-EEGNet	0.2602	0.0001	−0.3683	0.1944	Yes
EEGNet	Trans-EEGNet	0.122	0.0051	−0.2089	0.035	Yes

Table 9

Trans-EEGNet classification performance.

	Accuracy	Recall	F1 Score	MCC
Trans-EEGNet	0.7849	0.7917	0.8007	0.8300

Table 10

Trans-EEGNet majority voting results.

	Accuracy	Recall	F1 Score	MCC
Trans-EEGNet (Majority)	0.8421	0.8500	0.8597	0.8300

Table 11

Related research on HIE classification.

Method	Accuracy	Kappa	Author
TFDfeat	77.8%	0.68	[5]
GSVfeat	77.8%	0.81	[8]
IBIfeat	77.8	0.68	[53]
CNN1d	81.5%	0.73	[9]
QTFD+CNN	88.9%	0.84	[1]
Trans-EEGNet	95%	0.93	This Study

References

1. Raurale, S.A.; Boylan, G.B.; Mathieson, S.R.; Marnane, W.P.; Lightbody, G.; O’Toole, J.M. Grading hypoxic-ischemic encephalopathy in neonatal EEG with convolutional neural networks and quadratic time–frequency distributions. J. Neural Eng.; 2021; 18, 046007. [DOI: https://dx.doi.org/10.1088/1741-2552/abe8ae] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33618337]

2. Walsh, B.H.; Murray, D.M.; Boylan, G.B. The use of conventional EEG for the assessment of hypoxic ischaemic encephalopathy in the newborn: A review. Clin. Neurophysiol.; 2011; 122, pp. 1284-1294. [DOI: https://dx.doi.org/10.1016/j.clinph.2011.03.032] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/21550844]

3. Shankaran, S. Therapeutic hypothermia for neonatal encephalopathy. Curr. Treat. Options Neurol.; 2012; 14, pp. 608-619. [DOI: https://dx.doi.org/10.1007/s11940-012-0200-y] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23007949]

4. Murray, D.M.; Ryan, C.A.; Boylan, G.B.; Fitzgerald, A.P.; Connolly, S. Prediction of seizures in asphyxiated neonates: Correlation with continuous video-electroencephalographic monitoring. Pediatrics; 2006; 118, pp. 41-46. [DOI: https://dx.doi.org/10.1542/peds.2005-1524]

5. Stevenson, N.J.; Korotchikova, I.; Temko, A.; Lightbody, G.; Marnane, W.P.; Boylan, G.B. An automated system for grading EEG abnormality in term neonates with hypoxic-ischaemic encephalopathy. Ann. Biomed. Eng.; 2013; 41, pp. 775-785. [DOI: https://dx.doi.org/10.1007/s10439-012-0710-5]

6. Matic, V.; Cherian, P.J.; Koolen, N.; Naulaers, G.; Swarte, R.M.; Govaert, P.; Van Huffel, S.; De Vos, M. Holistic approach for automated background EEG assessment in asphyxiated full-term infants. J. Neural Eng.; 2014; 11, 066007. [DOI: https://dx.doi.org/10.1088/1741-2560/11/6/066007]

7. Matić, V.; Cherian, P.J.; Jansen, K.; Koolen, N.; Naulaers, G.; Swarte, R.M.; Govaert, P.; Van Huffel, S.; De Vos, M. Improving reliability of monitoring background EEG dynamics in asphyxiated infants. IEEE Trans. Biomed. Eng.; 2015; 63, pp. 973-983. [DOI: https://dx.doi.org/10.1109/TBME.2015.2477946]

8. Ahmed, R.; Temko, A.; Marnane, W.; Lightbody, G.; Boylan, G. Grading hypoxic–ischemic encephalopathy severity in neonatal EEG using GMM supervectors and the support vector machine. Clin. Neurophysiol.; 2016; 127, pp. 297-309. [DOI: https://dx.doi.org/10.1016/j.clinph.2015.05.024]

9. Raurale, S.A.; Boylan, G.B.; Lightbody, G.; O’Toole, J.M. Grading the severity of hypoxic-ischemic encephalopathy in newborn EEG using a convolutional neural network. Proceedings of the 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC); Montreal, QC, Canada, 20–24 July 2020; IEEE: New York, NY, USA, 2020; pp. 6103-6106.

10. Greenspan, H.; Van Ginneken, B.; Summers, R.M. Guest editorial deep learning in medical imaging: Overview and future promise of an exciting new technique. IEEE Trans. Med. Imaging; 2016; 35, pp. 1153-1159. [DOI: https://dx.doi.org/10.1109/TMI.2016.2553401]

11. Xie, J.; Zhang, J.; Sun, J.; Ma, Z.; Qin, L.; Li, G.; Zhou, H.; Zhan, Y. A transformer-based approach combining deep learning network and spatial-temporal information for raw EEG classification. IEEE Trans. Neural Syst. Rehabil. Eng.; 2022; 30, pp. 2126-2136. [DOI: https://dx.doi.org/10.1109/TNSRE.2022.3194600]

12. Sun, J.; Xie, J.; Zhou, H. EEG classification with transformer-based models. Proceedings of the 2021 IEEE 3rd Global Conference on Life Sciences and Technologies (LifeTech); Nara, Japan, 9–11 March 2021.

13. Lawhern, V.J.; Solon, A.J.; Waytowich, N.R.; Gordon, S.M.; Hung, C.P.; Lance, B.J. EEGNet: A compact convolutional neural network for EEG-based brain–computer interfaces. J. Neural Eng.; 2018; 15, 056013. [DOI: https://dx.doi.org/10.1088/1741-2552/aace8c] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29932424]

14. Millar, L.J.; Shi, L.; Hoerder-Suabedissen, A.; Molnár, Z. Neonatal hypoxia ischaemia: Mechanisms, models, and therapeutic challenges. Front. Cell. Neurosci.; 2017; 11, 78. [DOI: https://dx.doi.org/10.3389/fncel.2017.00078] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28533743]

15. Marlow, N.; Shankaran, S.; Rogers, E.E.; Maitre, N.L.; Smyser, C.D.; Guidelines, N.B.S. Publications Committee. Neurological and developmental outcomes following neonatal encephalopathy treated with therapeutic hypothermia. Seminars in Fetal and Neonatal Medicine; WB Saunders: Philadelphia, PA, USA, 2021; Volume 26, 101274.

16. Thompson, C.M.; Puterman, A.S.; Linley, L.L.; Hann, F.M.; van Molteno, C.D.; Malan, A.F. The value of a scoring system for hypoxic ischaemic encephalopathy in predicting neurodevelopmental outcome. Acta Paediatr.; 1997; 86, pp. 757-761. [DOI: https://dx.doi.org/10.1111/j.1651-2227.1997.tb08581.x]

17. Horn, A.R.; Swingler, G.H.; Myer, L.; Linley, L.L.; Raban, M.S.; Joolay, Y.; Harrison, M.C.; Chandrasekaran, M.; Rhoda, N.R.; Robertson, N.J. Early clinical signs in neonates with hypoxic ischemic encephalopathy predict an abnormal amplitude-integrated electroencephalogram at age 6 hours. BMC Pediatr.; 2013; 13, 52. [DOI: https://dx.doi.org/10.1186/1471-2431-13-52] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23574923]

18. Natarajan, G.; Pappas, A.; Shankaran, S. Outcomes in childhood following therapeutic hypothermia for neonatal hypoxic-ischemic encephalopathy (HIE). Seminars in Perinatology; WB Saunders: Philadelphia, PA, USA, 2016; Volume 40, pp. 549-555.

19. O’Toole, J.M.; Mathieson, S.R.; Raurale, S.A.; Magarelli, F.; Marnane, W.P.; Lightbody, G.; Boylan, G.B. Neonatal EEG graded for severity of background abnormalities in hypoxic-ischaemic encephalopathy. Sci. Data; 2023; 10, 129. [DOI: https://dx.doi.org/10.1038/s41597-023-02002-8] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36899033]

20. Allen, K.A.; Brandon, D.H. Hypoxic Ischemic Encephalopathy: Pathophysiology and Experimental Treatments. Newborn Infant Nurs. Rev.; 2011; 11, pp. 125-133. [DOI: https://dx.doi.org/10.1053/j.nainr.2011.07.004]

21. Shankaran, S.; Laptook, A.R.; Ehrenkranz, R.A.; Tyson, J.E.; McDonald, S.A.; Donovan, E.F.; Fanaroff, A.A.; Poole, W.K.; Wright, L.L.; Jobe, A.H. et al. Whole-body hypothermia for neonates with hypoxic–ischemic encephalopathy. N. Engl. J. Med.; 2005; 353, pp. 1574-1584. [DOI: https://dx.doi.org/10.1056/NEJMcps050929]

22. Ranjani, M.; Supraja, P. Classifying the autism and epilepsy disorder based on EEG signal using deep convolutional neural network (DCNN). Proceedings of the 2021 International Conference on Advance Computing and Innovative Technologies in Engineering (ICACITE); Greater Noida, India, 4–5 March 2021; IEEE: New York, NY, USA, 2021; pp. 880-886.

23. Wang, X.; Wang, Z. CNN with self-attention in EEG classification. Proceedings of the International Conference on Human-Computer Interaction; Virtual, 26 June–1 July 2022; Springer Nature: Cham, Switzerland, 2022; pp. 512-526.

24. Li, H.; Ding, M.; Zhang, R.; Xiu, C. Motor imagery EEG classification algorithm based on CNN-LSTM feature fusion network. Biomed. Signal Process. Control; 2022; 72, 103342. [DOI: https://dx.doi.org/10.1016/j.bspc.2021.103342]

25. Li, F.; Hao, K.; Wei, B.; Hao, L.; Ren, L. MS-FTSCNN: An EEG emotion recognition method from the combination of multi-domain features. Biomed. Signal Process. Control; 2024; 88, 105690. [DOI: https://dx.doi.org/10.1016/j.bspc.2023.105690]

26. Xin, Q.; Hu, S.; Liu, S.; Zhao, L.; Zhang, Y.D. An attention-based wavelet convolution neural network for epilepsy EEG classification. IEEE Trans. Neural Syst. Rehabil. Eng.; 2022; 30, pp. 957-966. [DOI: https://dx.doi.org/10.1109/TNSRE.2022.3166181]

27. Wei, Y.; Liu, Y.; Li, C.; Cheng, J.; Song, R.; Chen, X. TC-Net: A transformer capsule network for EEG-based emotion recognition. Comput. Biol. Med.; 2023; 152, 106463. [DOI: https://dx.doi.org/10.1016/j.compbiomed.2022.106463] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36571938]

28. Mulkey, M.A.; Huang, H.; Albanese, T.; Kim, S.; Yang, B. Supervised deep learning with vision transformer predicts delirium using limited lead EEG. Sci. Rep.; 2023; 13, 7890. [DOI: https://dx.doi.org/10.1038/s41598-023-35004-y] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37193736]

29. Merity, S. Single headed attention RNN: Stop thinking with your head. arXiv; 2019; arXiv: 1911.11423

30. Michel, P.; Levy, O.; Neubig, G. Are Sixt. Heads Really Better Than One?. arxiv; 2019; arXiv: 1905.10650

31. Liu, L.; Liu, J.; Han, J. Multi-head or single-head? An empirical comparison for transformer training. arXiv; 2021; arXiv: 2106.09650

32. Murray, D.M.; Boylan, G.B.; Ryan, C.A.; Connolly, S. Early EEG findings in hypoxic-ischemic encephalopathy predict outcomes at 2 years. Pediatrics; 2009; 124, pp. e459-e467. [DOI: https://dx.doi.org/10.1542/peds.2008-2190]

33. Amin, A.; Anwar, S.; Adnan, A.; Nawaz, M.; Howard, N.; Qadir, J.; Hawalah, A.; Hussain, A. Comparing oversampling techniques to handle the class imbalance problem: A customer churn prediction case study. IEEE Access; 2016; 4, pp. 7940-7957. [DOI: https://dx.doi.org/10.1109/ACCESS.2016.2619719]

34. Thabtah, F.; Hammoud, S.; Kamalov, F.; Gonsalves, A. Data imbalance in classification: Experimental evaluation. Inf. Sci.; 2020; 513, pp. 429-441. [DOI: https://dx.doi.org/10.1016/j.ins.2019.11.004]

35. Daud, S.N.S.S.; Sudirman, R.; Shing, T.W. Safe-level SMOTE method for handling the class imbalanced problem in electroencephalography dataset of adult anxious state. Biomed. Signal Process. Control; 2023; 83, 104649.

36. Rakhmatulin, I.; Dao, M.S.; Nassibi, A.; Mandic, D. Exploring convolutional neural network architectures for EEG feature extraction. Sensors; 2024; 24, 877. [DOI: https://dx.doi.org/10.3390/s24030877]

37. Saqib, M.; Zhu, Y.; Wang, M.; Beaulieu-Jones, B. Regularization of deep neural networks for EEG seizure detection to mitigate overfitting. Proceedings of the 2020 IEEE 44th Annual Computers, Software, and Applications Conference (COMPSAC); Madrid, Spain, 13–17 July 2020.

38. Pan, Y.; Dong, F.; Wu, J.; Xu, Y. Downsampling of EEG signals for deep learning-based epilepsy detection. IEEE Sens. Lett.; 2023; 7, 7007104. [DOI: https://dx.doi.org/10.1109/LSENS.2023.3332392]

39. Saideepthi, P.; Gaur, P.; Chowdhury, A.; Nyati, K.; Dwivedi, T.; Sharma, S.; Pachori, R.B. Sliding window along with EEGNet based prediction of EEG motor imagery. IEEE Sens. J.; 2023; 23, pp. 17703-17713. [DOI: https://dx.doi.org/10.1109/JSEN.2023.3270281]

40. Peng, R.; Zhao, C.; Jiang, J.; Kuang, G.; Cui, Y.; Xu, Y.; Du, H.; Shao, J.; Wu, D. TIE-EEGNet: Temporal information enhanced EEGNet for seizure subtype classification. IEEE Trans. Neural Syst. Rehabil. Eng.; 2022; 30, pp. 2567-2576. [DOI: https://dx.doi.org/10.1109/TNSRE.2022.3204540] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36063519]

41. Deng, X.; Zhang, B.; Yu, N.; Liu, K.; Sun, K. Advanced TSGL-EEGNet for motor imagery EEG-based brain-computer interfaces. IEEE Access; 2021; 9, pp. 25118-25130. [DOI: https://dx.doi.org/10.1109/ACCESS.2021.3056088]

42. Li, C.; Huang, X.; Song, R.; Qian, R.; Liu, X.; Chen, X. EEG-based seizure prediction via Transformer guided CNN. Measurement; 2022; 203, 111948. [DOI: https://dx.doi.org/10.1016/j.measurement.2022.111948]

43. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A.N.; Kaiser, Ł.; Polosukhin, I. Attention is all you need. Adv. Neural Inf. Process. Syst.; 2017; 30, pp. 5998-6008.

44. Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Houlsby, N. et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv; 2020; arXiv: 2010.11929

45. Kuang, D.; Michoski, C. KAM-A kernel attention module for emotion classification with EEG data. International Workshop on Interpretability of Machine Intelligence in Medical Image Computing; Springer Nature: Cham, Switzerland, 2022; pp. 93-103.

46. Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; Salakhutdinov, R. Dropout: A simple way to prevent neural networks from overfitting. J. Mach. Learn. Res.; 2014; 15, pp. 1929-1958.

47. She, H.-Q.; Sun, Y.-F.; Chen, L.; Xiao, Q.-X.; Luo, B.-Y.; Zhou, H.-S.; Zhou, D.; Chang, Q.-Y.; Xiong, L.-L. Current analysis of hypoxic-ischemic encephalopathy research issues and future treatment modalities. Front. Neurosci.; 2023; 17, 1136500. [DOI: https://dx.doi.org/10.3389/fnins.2023.1136500]

48. Ismael, A.M.; Alçin, Ö.F.; Abdalla, K.H.; Şengür, A. Two-stepped majority voting for efficient EEG-based emotion classification. Brain Inform.; 2020; 7, 9. [DOI: https://dx.doi.org/10.1186/s40708-020-00111-3]

49. Wang, Z.; Wang, Y.; Zhang, J.; Hu, C.; Yin, Z.; Song, Y. Spatial-temporal feature fusion neural network for EEG-based emotion recognition. IEEE Trans. Instrum. Meas.; 2022; 71, 2507212. [DOI: https://dx.doi.org/10.1109/TIM.2022.3165280]

50. Song, X.; Yan, D.; Zhao, L.; Yang, L. LSDD-EEGNet: An efficient end-to-end framework for EEG-based depression detection. Biomed. Signal Process. Control; 2022; 75, 103612. [DOI: https://dx.doi.org/10.1016/j.bspc.2022.103612]

51. Anguita, D.; Ghelardoni, L.; Ghio, A.; Oneto, L.; Ridella, S. The ’K’ in K-fold Cross Validation. Proceedings of the ESANN 2012; Bruges, Belgium, 25–27 April 2012; Volume 102, pp. 441-446.

52. Padfield, N.; Ren, J.; Qing, C.; Murray, P.; Zhao, H.; Zheng, J. Multi-segment majority voting decision fusion for MI EEG brain-computer interfacing. Cogn. Comput.; 2021; 13, pp. 1484-1495. [DOI: https://dx.doi.org/10.1007/s12559-021-09953-3]

53. Raurale, S.A.; Nalband, S.; Boylan, G.B.; Lightbody, G.; O’Toole, J.M. Suitability of an inter-burst detection method for grading hypoxic-ischemic encephalopathy in newborn EEG. Proceedings of the 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC); Berlin, Germany, 23–27 July 2019; IEEE: New York, NY, USA, 2019; pp. 4125-4128.

54. Lashgari, E.; Liang, D.; Maoz, U. Data augmentation for deep-learning-based electroencephalography. J. Neurosci. Methods; 2020; 346, 108885. [DOI: https://dx.doi.org/10.1016/j.jneumeth.2020.108885]

55. Veselov, D.; Andriyanov, N.; Trung, L. Detection of Intracranial Hemorrhage by Artificial Intelligence and Deep Learning Methods. Proceedings of the 2024 X International Conference on Information Technology and Nanotechnology (ITNT); Samara, Russian, 20–24 May 2024; IEEE: New York, NY, USA, 2024; pp. 1-6.

56. Allen, B. The promise of explainable ai in digital health for precision medicine: A systematic review. J. Pers. Med.; 2024; 14, 277. [DOI: https://dx.doi.org/10.3390/jpm14030277]