Seizure Detection in Medical IoT: Hybrid CNN-LSTM-GRU Model with Data Balancing and XAI Integration

Abstract

The brain acts as the body’s central command, overseeing diverse functions including thought, memory, speech, movement, and the regulation of various organs. When healthy, the brain functions seamlessly and automatically; however, disruptions can lead to serious conditions such as Alzheimer’s Disease, Brain Cancer, Stroke, and Epilepsy. Epilepsy, a neurological disorder marked by recurrent seizures, results from irregular electrical activity in the brain. These seizures, which can strain both patients and neurologists, are characterized by symptoms like the loss of awareness, unusual behavior, and confusion. This study presents an efficient EEG-based epileptic seizure detection framework utilizing a hybrid Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) models approach to support automated and accurate diagnosis. Handling imbalanced EEG data, which can otherwise bias model outcomes and reduce predictive accuracy, is a key focus. Experimental results indicate that the proposed framework generally outperforms other Deep Learning and Machine Learning techniques with the highest accuracy at 99.13%. Likewise, an Explainable Artificial Intelligence (XAI) called SHAP (SHapley Additive exPlanations) is utilized to analyze the results and to improve the interpretability of the models from medical decision-making. This framework aligns with the objectives of the Medical Internet of Things (MIoT), advancing smart medical applications and services for effective epileptic seizure detection.

1. Introduction

Today, Artificial Intelligence (AI) has become a significant driver in personalized diagnosis, treatment planning, and disease management within telemedicine applications. This advancement is further enhanced by the integration of Internet of Things (IoT) technology [1,2], as highlighted in [3,4].

Electroencephalogram (EEG) signals are electrical impulses generated by brain activity. Wearable EEG devices enable elderly and disabled individuals to collect EEG data, which can then be processed into control signals through motor imagery, allowing for remote control functions. Since EEG signals exhibit unique patterns in different mental states, machine learning techniques are well-suited for their classification. Many researchers have highlighted the applications of EEG signals and brain–computer interfaces in areas such as smart homes [5] and healthcare devices [6]; however, limited attention has been given to securing and maintaining the privacy of the EEG data itself.

Medical Internet of Things (MIoT) refers to the integration of Internet of Things (IoT) technology into the healthcare sector to enable efficient, real-time monitoring, diagnosis, and management of patients’ health. In practice, the Medical Internet of Things (MIoT) framework centered around EEG data is vulnerable to various security threats, including both passive and active attacks. Passive attacks involve intercepting or eavesdropping on transmitted data, while active attacks include altering or manipulating EEG data. For instance, when a hospital uses a cloud service to predict epilepsy risks in patients, it is crucial to ensure not only accurate predictions but also the security and privacy of the transmitted data. Consequently, securely handling encrypted EEG data has become a significant challenge. Some researchers have started exploring encryption-integrated machine learning solutions in response to these concerns, as discussed in [7,8].

Advances in Artificial Intelligence (AI), Machine Learning (ML), and the Internet of Things (IoT) offer promising potential for enhancing disease classification and prediction in telemedicine applications. ML techniques, for example, can facilitate the early detection of skin conditions with high accuracy and reliability. Recently, IoT and medical systems have emerged as complementary technologies that enhance Medical Internet of Things (MIoT) applications. By integrating healthcare services with IoT devices, real-time monitoring, analysis, and decision-making become possible in complex healthcare systems, bringing substantial benefits to patient care. The MIoT system encompasses multiple tiers, as depicted in Figure 1, and are as follows:

▪

Layer 1: In this layer, the essential role is played by MIoT sensors and actuators responsible for collecting and monitoring various medical and healthcare data.

▪

Layer 2: This layer encompasses the utilization of gateway and edge devices, serving as intermediaries connecting the Wireless Sensor Network (WSN) to cloud servers. These devices form an integral part of the MIoT-based controller platform.

▪

Layer 3: The utilization of cloud computing comes into play here, demonstrating its efficiency in powering intelligent medical IoT systems. The proposed system employs cloud computing servers for the storage and processing of industrial control field data obtained through the MIoT-based controller. The collected medical data are periodically transmitted to the appropriate channel via an IoT protocol like the Constrained Application Protocol (CoAP).

▪

Layer 4: At this layer, the focus shifts towards crafting mobile and web applications that interface with cloud servers to access analytical outcomes generated by applying machine learning techniques to store medical and healthcare data. The primary objective is to provide actionable decision-making insights for healthcare institutions.

In general, the multitier architecture of an MIoT system facilitates seamless data flow, storage, and subsequent analysis.

Imbalanced data pose a significant challenge in EEG seizure detection, where it refers to an unequal distribution of instances across classes in the dataset. In EEG seizure detection, instances representing normal brain activity typically far exceed those depicting seizure events. This imbalance can result in biased model performance and inaccurate predictions, as the algorithm may favor the majority class (normal activity) while struggling to accurately identify the minority class (seizures).

Imbalanced data in EEG seizure detection can have significant consequences. Models trained on imbalanced datasets often default to classifying most instances as belonging to the majority class (normal brain activity) due to its prevalence, thereby overlooking the critical minority class (seizures). This leads to a problematic trade-off where the model’s high accuracy is largely due to its success in identifying the majority class, while it struggles to detect seizure events accurately.

Addressing imbalanced data in EEG seizure detection is essential for achieving accurate and reliable outcomes. Techniques like resampling—either by oversampling the minority class or under-sampling the majority class creating synthetic samples—and specialized algorithms such as ensemble methods can help balance class distributions. These approaches enable the model to effectively learn from both classes, ensuring that seizure instances are not overshadowed by the abundance of normal data, thus improving the overall performance and reliability of seizure detection.

In real-world scenarios, collecting medical data like EEG from devices poses additional challenges, often resulting in a class imbalance that complicates fault diagnosis. Training seizure detection models directly on such imbalanced datasets can lead to overfitting on the minority class and bias toward the majority class. To overcome this, we propose a specialized seizure detection framework, leveraging an optimized K-Nearest Neighbors (KNN) algorithm adapted for imbalanced EEG data, specifically designed to enhance performance in smart medical IoT systems. In brief, this paper provides a robust and intelligent framework for seizure detection considering the context of MIoT devices. With the impressive strides seen in MIoT health care, we emphasize the relevance and practicality of our ensemble model in this field. The contributions of this paper can be summarized as follows:

• The proposed hybrid framework is a merge of deep learning techniques including Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Gated Recurrent Units (GRUs) The suggested hybrid framework is validated for seizure detection in EEG signals.
• The synthetic minority oversampling technique (SMOTE) is applied and is used to balance the dataset by creating additional samples for the minority class. Additionally, the Explainable Artificial Intelligence (XAI) technique, SHAP (Shapley Additive Explanations), is applied to evaluate the importance and effectiveness of the features.
• To validate and determine the effectiveness of the proposed technique in identifying seizures, the comparison result with existing Machine Learning (ML) and Deep Learning (DL) models is conducted. The results demonstrate that the CNN-LSTM-GRU model performed better than other models in the detection of seizures. The proposed technique has a higher accuracy, precision, F1 score, recall, and specificity.
• The proposed model is designed with real Medical IoT (MIoT) applications. So, the model extends the scope of seizure monitoring intelligently due to its scalability and flexibility. The designed approach allows a range of MIoT components to operate efficiently regardless of the deployment environment.

The article is organized as follows: Section 2 reviews previous studies related to the topic, while the proposed system for MIoT is outlined in Section 3. Section 4 details the experimental study and analyzes the results. Finally, Section 5 provides the conclusion on this innovative approach.

2. Materials and Methods

Recently, considerable focus has been placed on the classification and modeling of EEG data. Electroencephalography (EEG) has become an essential tool for remote epileptic seizure detection, and numerous methodologies have been developed for the early assessment of epilepsy [9,10,11]. Recent research has focused on optimizing EEG data classification and enhancing privacy through secure methods. Yong Jiao and colleagues [9] proposed a Sparse Group Representation Model (SGRM) that leverages intersubjective data to develop an accurate motor imagery-based Brain–Computer Interface (BCI), achieving a classification accuracy of 78.2%. Chatterjee and colleagues [10] utilized the Fuzzy Discernibility Matrix (FDM) for feature extraction from EEG signals, employing Support Vector Machine (SVM) and ensemble classifiers for accuracy evaluation, though the method encounters limitations in handling large EEG datasets and multiclass scenarios.

For EEG data classification, a deep Convolutional Neural Network (CNN) approach reached 86.41% accuracy in [11] and 85.62% in [12]. In another study, Coiflets wavelets were used for preprocessing, focusing on amplitude-based feature extraction [13]. Ward and colleagues [14] applied Joint Factor Analysis and I-vectors for feature extraction with Universal Background Models (UBMs) for classification, although this approach did not achieve optimal accuracy.

In [15], multi-feature fusion with CNNs was recommended for automated epilepsy detection. Another study [16] proposed an algorithm for noisy environments, using sample entropy and phase synchronization with Variational Mode Decomposition (VMD) for seizure detection. This model emphasized minimizing false alarms while ensuring cross-patient applicability. To improve efficiency, the dynamic Epileptic-Net model was introduced [17], and in [18], both time- and frequency-domain features were used with machine learning models (e.g., Logistic Regression, Decision Trees, SVM) to detect seizures within the Temple University Hospital (TUH) corpus. Additionally, Ref. [19] used discrete wavelet transform (DWT) and genetic algorithms with machine learning classifiers to enhance epileptic seizure detection accuracy and reduce computation time. Additionally, in [20], a novel method integrating Bi-LSTM, GRU, and Average Pooling demonstrated exceptional performance across binary and multiclass tasks. Also, some hybrid models are used as in [21,22,23].

One of the open research areas in AI and ML is XAI. Although AI, especially deep learning, has become increasingly sophisticated and complex, they are viewed as “black boxes”, making it difficult to understand how they compute the predictions or decisions. XAI provides the tools to explain and interpret these models in users appropriate ways to trust the models [24,25,26]. In the last decade there has been a lot of research focus into XAI in healthcare [27]. In Ref. [28] the revolutionary potential of Explainable AI (XAI) is demonstrated in developing Healthcare 5.0. It highlights the significance of explainability in boosting trust, privacy, and clinical decision-making by proposing a thorough architecture and solution taxonomy for XAI in applications like CT image analysis and ECG monitoring

Recently, considerable research has focused on EEG classification, as demonstrated by the studies discussed above.

Table 1. A comparative analysis of EEG classification techniques.

Reference	Methodology	Key Features	Dataset
Jiao et al. [7]	Sparse Group Representation Model (SGRM)	Uses intersubjective data for BCI	BCI Competition IV (64 EEG channels, 1000 Hz sampling rate, 2 classes, and 7 subjects) [29]
Chatterjee et al. [8]	Fuzzy Discernibility Matrix (FDM) with SVM and Ensemble	Extracts optimal features from EEG signals	BCI Competition-II Dataset-III [30] training and 140 test trials 9 s long each, 3 channels, 128 Hz sampling rate, 2 classes) [29]
Tang et al. [9]	Deep CNN	Applied for EEG data classification	Experimental dataset (private dataset from 2 subject 28 experiments)
Alomari et al. [11]	Coiflets Wavelet and Amplitude Estimation	Feature extraction from EEG signals	PhysioNet EEG dataset (1500 two-minute EEG recordings, obtained from 109 subject) [31]
Ward et al. [12]	Joint Factor Analysis and I-vectors	Feature extraction with UBM classification	PhysioNet’s EEG dataset [31] Temple University Hospital (TUH) (16,986 sessions from 10,874 subjects) [32]
Y. Sun and X. Chen. [13]	Multi-feature Fusion with CNN	Used for epilepsy detection	CHB-MIT EEG Dataset (182 records of 23 seizure cases) [33]
Ru et al. [14]	Noisy Environment Algorithm with VMD	Uses sample entropy and phase synchronization	CHB-MIT EEG Dataset (182 records of 23 seizure cases) [33]
G. Zazzaro and L. Pavone [15]	Inclusion of Includes preictal/ictal (seizure-related) and interictal (non-seizure) EEG	Enabling patient-specific seizure detection and cross-patient model generalizability	Freiburg Seizure Prediction EEG Database (21 subject of seizure patients using 128 channels, 256 Hz sampling rate) [34]
Islam et al. [16]	Epileptic-Net	Dynamic deep learning model for seizures	BONN EEG Dataset (4096 samples) [29]
Khan et al. [17]	Time and Frequency Features with ML models	Logistic Regression, Decision Trees, SVM on TUH corpus	Temple University Hospital (TUH) (16,986 sessions from 10,874 subjects) [34]
Mardini et al. [18]	DWT, Genetic Algorithm, ML Classifiers	Reduces computation for seizure detection	BONN EEG Dataset (4096 samples) [29]
Zhan et al. [20]	Hybrid Deep Learning	Using Bi-LSTM, GRU, and average pooling layer	BONN EEG Dataset (4096 samples) [29]

shows a comparative summary of EEG data classification techniques in medical applications. However, little attention has been paid to applying optimization techniques to improve machine learning models and address the problem of imbalanced data in seizure recognition datasets. These datasets are used to classify EEG signals and aim to automatically detect seizures at an early stage. This paper presents an effective approach for seizure detection and prediction that not only addresses the problem of imbalanced EEG data, which often leads to overfitting, but also incorporates optimization techniques to improve model performance. The proposed system shows promise for deployment in smart healthcare applications.

The use of Artificial Intelligence (AI) in healthcare raises important ethical considerations that must be addressed to ensure its responsible application. One critical concern is bias in AI models, which can arise from imbalanced training data and may lead to unequal outcomes for different patient groups. Ensuring fairness requires diverse, representative datasets and rigorous validation across various populations to mitigate the risk of discriminatory practices. Another vital aspect is informed consent, where patients must be made aware of how their data are collected, processed, and utilized by AI systems. Transparency in AI decision-making, supported by Explainable AI techniques, is crucial for building trust between patients and healthcare providers. Additionally, safeguarding patient privacy and complying with regulatory frameworks such as HIPAA and GDPR are essential to uphold ethical standards. By addressing these considerations, the integration of AI in healthcare can be guided by the principles of fairness, accountability, and respect for patient autonomy.

3. Proposed Seizure Detection for MIoT

This section explains the main stages for our proposed model and demonstrates its construction for handling complex data typically associated with the MIoT.

3.1. Stages of Proposed Framework

Based on smart Medical Internet of Things (MIoT) systems technology which ensures the detection of seizures from EEG signals with maximum reliability, the proposed detection system is founded on a developed theoretical framework, consisting of, but not limited to, statistical learning, optimization theory, and deep learning techniques. We introduced a robust method for the automated identification and detection of seizures by a set of advanced Deep Learning (DL) methodologies, each adopted for its theoretical merits and its relevance to the identification problem while effectively addressing the challenge of imbalanced EEG data, all within the context of MIoT systems. The depicted system structure can be observed in Figure 2, encompassing a series of significant stages aimed at achieving accurate seizure prediction.

The main steps of the proposed system are elaborated as follows:

• Step 1 (Data Acquisition): At this initial stage, the imbalanced seizure datasets are collected, forming the foundation for subsequent phases within the proposed framework.
• Step 2 (Data Preprocessing): All instances of the dataset are imported to make them suitable for further processing within the framework. To effectively address the imbalance problem, the SMOTE technique is applied to handle seizure instances.
• Step 3 (Model Training/Testing): The preprocessed and rebalanced dataset is divided into an 80–20 ratio, where 20% is allocated for the testing phase and 80% for the training phase. This initiates the training phase for hybrid Deep Learning models (i.e., CNN-LSTM-GRU). To evaluate the proposed framework, a subset of randomly selected training data is used for the chosen DL for seizure analysis.
• Step 4 (Feature interpretability): In this segment of the framework, the SHAP (Shapley Additive Explanations) technique is employed to enhance interpretability, providing insights into feature contributions for the prediction. This makes it easier to comprehend how the model is attained and helps in the decision-making process

3.2. Proposed Model Architecture

SMOTE or the Synthetic Minority Over-sampling Technique can be said to be an effective algorithm in solving class imbalance issues in EEG data specifically for seizure classification. SMOTE improves the proportion of minorities to majorities in the given dataset by creating realistic synthetic samples rather than repeating the subjects of the minority class. It does so by using the existing samples of the minority class and drawing new points in the feature space defined by these samples. If this technique is utilized when dealing with EEG data, it will prevent the under-representation of seizure-related patterns, improving the efficiency of training the machine to classify both seizure and non-seizure instances thus raising the general classification performance.

Convolutional Neural Networks (CNNs) are applied due to their unique strength of feature extraction of the raw EEG data. This ability is important in mapping the complex and decorated patterns of the EEG signals associated with seizures, such as identifying some small changes that may be related to seizures. Whereas Recurrent Neural Networks (RNN) called LSTM and GRU are developed to avoid the loss gradient result, these architectures can represent temporal dependencies. Their capability to remember information over long sequences makes them suitable for continuous and time-varying data such as EEG, hence facilitating the monitoring and classification of seizures.

Our proposed technique is the combination of CNN with LSTM and GRU. The model has two types of inputs with two branches that connect to a final concatenation layer. In the case of the CNN branch, the first layer is an input layer where the dimensions are determined by the training dataset, followed by a 1D convolutional layer that applies the relevant filters to capture and retrieve local structures in EEG data. This is then followed by a MaxPooling layer designed to down-sample the convoluted feature maps or the activations and highlight the most critical features. The features are then flattened, ready for concatenation with the LSTM and GRU feature branches. So, the architecture of the hybrid model is as follows: The CNN branch has 3 Layers (Conv1D -> MaxPooling1D -> Flatten). It uses ReLU activation in the convolutional layer for non-linearity. The LSTM branch has 1 Layer u which uses ReLU activation in the LSTM layer. The GRU branch has 1 Layer. Similarly to LSTM, ReLU activation is used in the GRU layer. The concatenation layer combines the outputs of CNN, LSTM, and GRU into a single feature vector. The output layer has a dense layer with 1 unit and sigmoid activation for binary classification. The compilation details optimizer is Adam with loss function: binary cross-entropy and metrics are the accuracy.

The temporal dependencies within the model are managed in the second branch utilizing the LSTM and GRU layers. The LSTM layer is intended to learn the long-term dependencies thanks to its ability to learn sequences with varying complex dynamics. The GRU layer works similarly and is also utilized to learn varying patterns but in a rather more simplified manner. The output features generated from all three branches, the CNN, LSTM, and GRU, are merged through a concatenate layer, thereby allowing the fusion of the selected features collected from the CNN and the other selected features captured from the LSTM and GRU. The concatenated resulting output is further passed to the final dense layer which is fitted with a sigmoid activation function to enable it to perform a binary classification: seizure or no seizure. The construction is undertaken using the Adam optimizer and binary cross-entropy loss as the compilation of the model for 100 epochs.

4. Experimental Results

This section presents a description of the Epileptic Seizure Recognition Dataset, along with the performance metrics and results analysis of the proposed system.

4.1. Epileptic Seizure Recognition Dataset

The epileptic seizure dataset used in this study is sourced from [35], with its structure detailed as follows: it consists of 11,500 samples, each containing 178 features, following a normal distribution. The dataset is categorized into two binary classes, represented as Y = {0, 1} A comprehensive description of this dataset is provided in

Table 2. Dataset Description.

Class Type	Classes Description	Number of Cases
1 (Seizure)	Recording of seizure activity	2300
0 (Normal)	Normal activity	9200

and Figure 3.

The original dataset from the reference consists of 500 patients. For each patient, a 23.6 s recording of brain activity. The recording is sampled into 4097 data points, with each point representing the EEG signal value at a specific moment in time. The dataset includes recordings from 500 patients, with each patient contributing 4097 data points captured over the 23.6 s duration.

4.2. Evaluation Metrics

The effectiveness of Deep Learning models in classification and prediction tasks is commonly evaluated using metrics like accuracy. These evaluation metrics, including accuracy, can be derived by analyzing parameters from the confusion matrix, as illustrated in

Table 3. The parameters of the confusion matrix.

	Predicted Seizure	Predicted Normal
Actual Seizure	TP	FP
Actual Normal	FN	TN

. The accuracy, precision, and recall metrics are calculated using Equations (1)–(4), respectively.

After calculating the parameters, the performance metrics can be computed as follows [36,37,38]:

$$Accuracy={\displaystyle \frac{TP+TN}{TP+FP+FN+TN}}$$

(1)

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(2)

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(3)

$$Specificity={\displaystyle \frac{TN}{FP+TN}}$$

(4)

$$F1-score=2\times {\displaystyle \frac{Precision\times Recall}{Precision+Recall}}$$

(5)

4.3. Results Analysis

In the initial phase, EEG signals are preprocessed to improve their quality. This process begins with the application of the min–max normalization technique to standardize the dataset. First, the minimum and maximum values in the dataset are identified. Each data point is then scaled relative to these extremes to set the maximum value to 1 and the minimum to 0, placing all other values within the [0, 1] range.

To correct the dataset’s class imbalance, SMOTE is applied. Class imbalance refers to the situation when one or more classes have an extremely lower number of sample data than the other(s); this could make the model biased towards the majority class. With SMOTE, the seizure activity class is expanded by synthetic means instead of just randomly respawning some members of that class. This means that the feature space of the minority class instances is located, similar ones are paired up, and synthetic ones are created in between those pairs. This way the classes in the training set become balanced, improving the capability of the model to learn the distinguishing characteristics of the minority class thus improving the generalization of the model.

The design of the hybrid CNN-LSTM-GRU-based algorithm for the detection of seizures from EEG signals takes advantage of Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU) to consider all features from the EEG data.

Figure 4 shows the training and validation accuracy of our proposed framework for 100 epochs. The training accuracy rises sharply and flattens out at almost 100%, indicating that the model fits the training data very well. The validation accuracy increases in the same manner but stabilizes at a value slightly less than the training accuracy, around 99%. The above scenario of the training and validation accuracy indicates that there is no overfitting as the model has generalized well. The high accuracy level of 99.13% achieved within the framework demonstrates the strength of this framework in seizure detection using EEG signals.

From the ROC curve shown in Figure 5, our proposed framework achieves an accuracy of 99.15% which indicates exceptional performance in distinguishing between classes. The curve closely follows the top left corner of the graph, reflecting that the model accurately identifies almost every positive instance without incorrectly classifying any negatives as positives. Such a result signifies an ideal balance between sensitivity and specificity, suggesting the model’s highly reliable performance in classification tasks.

The XAI technique with the use of SHAP [25] is employed to enhance interpretability, providing insights into feature contributions for the prediction for the different model as shown in Figure 6 and Figure 7. This makes it easier to comprehend how the model is attained and helps in the decision-making process. So, the SHAP summary plot provides a comprehensive view of feature importance and its influence on seizure classification. The features are ranked by their impact, with X177 being the most influential in the proposed model. The local explanation and SHAP plot play a significant role in showing how the model came up with the prediction of a seizure, determining the specific features that contribute to the model’s output, enhancing interpretability, and identifying the key drivers in seizure detection.

4.4. Comparison Results

The comparison of the different DL models with the proposed model is conducted to evaluate the effectiveness of the proposed framework. The comparison results of confusion matrices are depicted in Figure 8. These confusion matrices, for the CNN, LSTM, GRU, and the proposed model show the superiority of the hybrid model in the detection of seizure conditions based on EEG readings. The proposed framework has been rated as the best among all, with the fewest false negatives appearing (11) and false positives present (21). The LSTM model follows with 46 misclassifications, and both the CNN and GRU achieve lower accuracies with 68 and 61 misclassifications, respectively. The proposed model significantly reduces false negatives and provides higher sensitivity and specificity, making it efficient for capturing seizure events alone without much risk of normal being classified as seizures. Overall, it can be noted that the CNN-LSTM-GRU model led to great progress in seizure detection as it guarantees proper functions that involve differentiating seizure and normal conditions present in EEG data.

For the value of accuracy if changing the test size, if we split the test to be 25% then the accuracy will be 99.15%, but if we reduce the test to 20% the accuracy will be 99.02% and if we increases the test to 30% the accuracy will be 98.84%.

To evaluate how the hybrid CNN-LSTM-GRU model stacks up against DL and traditional ML models in terms of accuracy, precision, recall, F1 score, and specificity, we have compared them.

Table 4. A comparison between the proposed system, ML, DL, and our proposed hybrid model with and without SOMTE.

Algorithm	Accuracy	Precision	Recall	F1-Score
LR	0.6207	0.65	0.64	0.64
DT	0.9234	0.92	0.92	0.92
NB	0.8891	0.90	0.89	0.89
SVM	0.9647	0.96	0.96	0.96
RF	0.9707	0.97	0.97	0.97
KNN	0.9834	0.98	0.98	0.98
CNN	0.964	0.985	0.977	0.981
LSTM	0.9848	0.986	0.988	0.987
GRU	0.9788	0.983	0.984	0.983
Proposed Hybrid Model Without SMOTE	0.9787	0.9786	0.9787	0.9785
Proposed Hybrid Model With SMOTE	0.9913	0.9913	0.9913	0.9913

shows the comparison results. It illustrates and summarizes the performance of different algorithms and techniques deployed in Machine Learning and Deep Learning datatypes for the binary classification of seizure in EEG datasets based on measures such as accuracy, precision, recall, F1 score, and specificity, among others as shown in Figure 8 and Figure 9. Where Logistic Regression (LR), Decision Tree (DT), Naive Bayes (NB), Support Vector Machine (SVM), Random Forest (RF), K-Nearest Neighbors (KNN), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and the hybrid model are studied. Here, the so-called traditional Machine Learning models, like LR and DT, show moderate to high results, though, in terms of effectivity, LR should be placed last at 62.07% accuracy. However, the DT model has a dramatically increased rate of accuracy at 92.34% with all the metrics being equal across the board. The Naive Bayes algorithm also shows quite pleasing outcomes and performance results with its accuracy standing at 88.91% and specificity at 96.8%, meaning that it can classify the classes effectively.

Figure 9 presents an evaluation comparison using a cluster bar chart among different ML algorithms, CNN, LSTM, GRU, and our proposed hybrid approach with and without SOMTE technique. It appears that Deep Learning approaches like, CNN, LSTM, or GRU, tend to perform better than conventional Machine Learning methods on most of the metrics, with CNN recording 96%:74% accuracy and LSTM achieving the highest accuracy among all the deep learning models of 98.48%. Since both LSTM and GRU belong to the family of recurrent neural networks, they achieved high recall and F1-score metrics, which indicates that they are good for time-sequence-related data. The hybrid model proposed in this paper achieves the best overall results with an accuracy of 99.13% and a precision, recall, and F1-score value of 99.13%, along with a specificity of 98.87%. These parameters show that the proposed system is the best-in-class discrimination and therefore arguments about sensitivity and specificity even-out. This additionally shows that the model would work well in classification tasks that can classify the seizure from EEG signals effectively. This result depicts how successful the combined model is in utilizing the advantages of different systems to deliver the best results.

To compare our approach with the recent state-of-the-art methods, four recently published methods referenced at [30,39,40,41,42] are used. These methods are all hybrid Deep Learning-based approaches that predict seizure epilepsy utilizing EEG signal.

Table 5. A comparison between the proposed system with previous studies in the same dataset.

Paper	Accuracy
Alalayah et al. [39]	98.98%
Akyol, Kemal [40]	97.17
Kunekar et al. [30]	97.10
Kode et al. [41]	99.00%
Woodbright et al. [42]	98.65
Proposed model	99.13%

shows this comparison.

4.5. Limitations and Challenges

There are many limitations and challenges in this study, such as the following:

▪

Lack of External Evaluation: The framework’s performance, while impressive, is primarily evaluated using internal validation. The absence of external evaluation on diverse datasets limits its generalizability and raises concerns about its applicability to broader populations.

▪

Synthetic Data Bias: Although SMOTE effectively addresses class imbalance, it may introduce biases by generating synthetic samples that do not fully capture the variability and complexity of real-world data.

▪

Sensitive Medical Data Handling: Managing sensitive medical data presents significant challenges, necessitating strict compliance with data privacy regulations such as HIPAA or GDPR and the implementation of robust anonymization and security measures.

▪

Computational Complexity: The hybrid model’s complexity, combining CNN, LSTM, and GRU architectures, may result in high computational demands, potentially limiting its deployment on resource-constrained devices, especially in real-time Medical IoT (MIoT) environments.

▪

Scalability and Integration: Integrating the model into existing MIoT frameworks and ensuring scalability across diverse medical applications and hardware configurations remains challenging.

Addressing these limitations in future work is critical to enhancing the model’s robustness, generalizability, and practical utility in real-world medical scenarios.

5. Conclusions

Being the control center of the body, the brain is fundamental to the execution of cognitive tasks, retention of information, articulation, movements as well as the homeostasis of body systems. This paper presents an effective EEG-based system for the early detection of epileptic seizures from an EEG signal, utilizing the hybrid Deep Learning model of the CNN, LSTM, and GRU models. The proposed framework applies the SMOTE technique to efficiently solve the challenge of imbalanced EEG data for undistorted and correct classification. The model achieved 99.15%, ensuring that the accuracy surpassed the Deep Learning and Machine Learning models. The specialists were able to understand the model better using Explainable AI techniques such as SHAP which made the process of medical decision-making more user-friendly. This framework aligns with the objectives of the Medical Internet of Things (MIoT) scheme as it encourages the development of better medical applications toward timely and accurate epileptic seizure detection.