According to the international league against epilepsy (ILAE) statement, quoted in 2014, epilepsy is a transient occurrence of signs and symptoms due to abnormal, excessive, or synchronous neuronal activity in the brain [1]. Epilepsy can be diagnosed by three main symptoms: at least two seizures have occurred over 24 h, reflex seizures have occurred twice or more over 10 years, and/or epilepsy syndrome [1]. Epileptic seizure is the most common symptom [2], that is caused by a disrupting episode with the brain’s activities, and not all epileptic patients have seizures [3]. Approximately 50 million people over the world have epilepsy [4], that is 1% of the overall population. According to the Saudi epilepsy society, 0.654% of people in Saudi Arabia suffer from this disease. Epilepsy affects all ages, however, the symptoms and signs differ by age group. For example, in newborns, the symptoms involve a lack of oxygen during delivery and/or abnormal brain development. In infants symptoms are brain tumors and/or genetic disorders [5]. Therefore, an accurate diagnosis is important. As childhood is the stage of brain formation, epilepsy occurs more dynamically in the nervous system and may interfere with the brain development. That could affect the individual in several ways, such as the failure to develop skills, a slower growth rate, and the possibility of losing previously developed skills [6]. Every 4.8/1000 children worldwide suffer from epilepsy. In the US, 0.94/1000 children under 18 suffer from epilepsy [7]. In recent years, epilepsy diagnosis in infants and children has been improved, and investigation of new methods has become hot topic for researchers. EEG signals can reflect the state of temporal brain activities. It is a complex and nonlinear interconnection between a billion neurons [8]. In addition, it is common to diagnose epilepsy by analyzing the EEG data; in this regard, several studies have focused on aiding epileptic patients to find suitable treatments. Moreover, it involves EEG signal processing prior to application of the identification methods. Hence, the accuracy of the system also depends on the way the EEG signals are processed. In this regard, time and transform domain (frequency transform, wavelet transform etc.) signal processing techqniues have been investigated in the literature [9]. Efficacy of the signal processing technique greatly enhances the efficiency of epilyptic pediatric identification approaches. So to identify such a critical disease at the early stages, there is a dire need for investigating the machine learning (ML) techniques. Current study aims to investigate the ensamable based ML techniques for pediatric epilyse identification.

The organization of the paper is as follows: Section 2 contains a literature review, proposed methods are given in Section 3, Section 4 contains results and discussion while Section 5 concludes the paper.

ML has countless applications in almost every field of study and human life. However, in the medical sector, its significance is matchless. Other than pediatric epilepsy, ML has been investigated for detection and prediction of several fatal diseases like COVID19 [10], cardio diseases [11], diabetes [12], Parkinson [13], and many others. In [14,15], the researchers applied the SVM algorithm to classify seizure onset in pediatric epilepsy patients. The dataset from the CHB-MIT [16]contained 23 patient’s files of EEG recording and all patients under eighteen years. It is consisted of eighteen channels containing 163 seizures that were separated into records of one hour each. In [17], the authors designed an advanced seizure prediction via preictal relabeling (ASPPR) algorithm to predict epileptic pediatric seizures. The model achieved 96.30% accuracy for predicting seizures between 1 to 6 mins, 96.13% accuracy for 8 to 13 mins, 94% accuracy for 14 and 19 mins, and 94.2% accuracy for 20 and 25 mins. In [18], the authors proposed a technique to extract features from the epileptic pediatric dataset from the CHB-MIT EEG database. In their proposed method, EEG signals are mapped into 2D space that leads to texture image and the gray-level image domain. Furthermore, they compared their results with other methods and achieved 97.74% accuracy using the SVM classifier with linear kernels.

The data mining of EEG signals was analyzed by using a time series approach, which calculated peak points (the lowest point in a part of the signals) and valley points (the highest point in a part of the signals) and then calculated the distances between them. The study consisted of three experiments; all of them predicted onset seizures early with minimum latency. Identifying seizures in children is different than in adults because the seizures in children have lacked the characteristics of EEG features. The study in [19] aimed to find the difference between the type of spike called Rolandic in two groups of epilepsy, patients with benign focal epilepsy and patients with structure focal epilepsy, the author used an algorithm with three stages: spike detection, determination parameters, and classification. The proposed method achieved 75% with an Artificial neural network (ANN). Some studies that focused on newborn data, proposed alternative methods for detecting seizures [20–25] and achieved a high-level accuracy compared with previous studies. The number of electrodes may affect pediatric seizure detection. In [26], authors proposed a collective network of binary classifiers (CNBC) using multi-dimensional particle swarm optimization (MD PSO). The proposed method achieved 93% accuracy and it was applied on long-term EEG data for seizures extraction. The authors in [27] detected cognitive impairment in children with epilepsy by using network analysis, and their proposed method achieved 85% accuracy.

Summary of the state-of-the-art ML related techniques in detection of epilepsy in babies and children are enlisted in Tab. 1. According to the comprehensive and systematic literature review, many studies in detecting pediatric epilepsy are done, but the performance still needs to be improved. The highest achieved accuracy of 96% is claimed by [14,15] for the CHB-MIT EEG database. Apart from the detection/prediction accuracy, there is another challenge to identify the right candidate ML algorithm for future and advance prediction of pediatric epilepsy. Since there is a variety of ML algorithms with their own strengths and weaknesses against various application domains. To answer these questions, the current study is performed. The study aims to investigate several ML and ensemble learning algorithms to potentially enhance the performance of EEG data classification. For evaluation of these techniques, several performance metrics are targeted including accuracy, sensitivity (true positive rate) and specificity (true negative rate), etc.

Original EEG signals contain noise from two environments, the first is external environments, such as electrode displacement and cable movement, and the second is internal sources, such as muscle movements and eyes blink [28]. These noises affect the signal and as a result the signals may incorrectly classified. Therefore, preprocessing the EEG signals prior to classification is essential and inevitable. The EEGLAB is an open-source toolbox for preprocessing EEG data [29]. It is compatible with MATLAB and is used to remove all noises and artifacts from the EEG signals. EEG signals are comprised of five sub-band frequencies: “delta” (0.5–4 Hz), “theta” (4–8 Hz), “alpha” (8–13 Hz), “beta” (13–30 Hz), and “gamma” (> 30 Hz), given in Tab. 2. The higher frequencies are considered as abnormal [30]. These five frequency bands present accurate information regarding brain signals. Therefore, using specific techniques to decompose signals to classify them is an optimal method, and the wavelet transform is a famous technique for this purpose [30].

After extracting useful features from the EEG signals, seven algorithms were applied to classify the data into “epileptic” and “non-epileptic” seizures, and the performance parameters were calculated to evaluate the model. This section presents the results as well as a comparison of the study with previous studies that used the same dataset. The data were divided into two sets: training and testing, with 90% and 10% division, respectively for each classifier. Then, the sensitivity, specificity, and accuracy for each classifier were calculated for the entire dataset and averaged. The results of the performance measurement parameters for all patients equal and less than 7 years old are presented in Tab. 6 and depicted in Fig. 4 as well. It can be observed that the best performance was achieved with Ensemble learning which achieved 100% accuracy overall and indicating that the model predicted all classes successfully. SVM achieved good accuracy which is 98% overall. The lowest accuracies were achieved with LD and LR, with 70.2% and 73% overall, respectively. NB, KNN, and DT classifiers achieved satisfactory accuracies with 87.5%, 90.2%, and 92.3%, respectively. In the KNN model, the number of neighbors (K) that achieved the best accuracy was 3 (while 1 and 2 achieved the same accuracy but lower than 3). When the ensemble learning model was applied, the bagged tree method was used to solve the classification problems and the DT was used for building the learner type of ensemble classifier.

There are various ways to assess classifier performance like scatter plots, confusion matrices, and receiver operation characteristic (ROC) curves. Presented below are the three assessment methods as an example for patient 16. The scatter plot of the LD model is presented in Fig. 5. Further, the classifier result after training the data is presented, in which the correct and incorrect classes are denoted with a dot and as X, respectively, while the different colors refer to different classes. For the LD model, confusion matrix is presented in Fig. 6 as an example and the classifier performance for each class is presented. The rows present the predicted classes, while the columns display the actual classes. The blue cells denote that the classifiers are performing well, while the red cells denote that they are not. The classifier predicted the true negative classes correctly and predicted twelve of the false-negative classes incorrectly. It appears that the classifier peforms well in the true classes and vice versa.

The ROC curve is another method to assess model performance. It provides the TP rate versus the FP rate of the classifier in the chart. For example, Fig. 7 illustrates the ROC curve of the NB model for patient 16 as an example. The red marker displays the value of the “true positive” and “false positive” of the classifier. The blue area under the curve indicates the quality of the classifier. A larger area under the curve indicates better performance and vice versa. The accuracies of the seven models were satisfactory, indicating that the proposed method worked well on the EEG data. Ensamble learning and SVM model achieved high accuracies overall with 100% and 98%, respectively. Based on the previous studies presented in the literature review, the SVM classifier demonstrated high performance. From the experiment, it is observed that the outperformed accuracies were obtained by the Ensemble learning model, which achieved 100% overall which is the highest accuracy achieved among the seven models. On the other hand, the LD and LR models achieved the lowest accuracies among the algorithms, with 70.2% and 73% for LD and LR overall, respectively. In comparing the results with other studies that used the same EEG dataset the Ensemble learning method produced better sensitivity, specificity, and accuracy, at 100%. According to Tab. 7, proposed model significantly outperformed recent studies, and with lesser complexity. Satirasethawong et al. [44] proposed an algorithm for evaluating their methods of epileptic seizure classification. This algorithm was developed based on static methods. They used one performance parameter, namely sensitivity (in [45,46] they also used only the sensitivity as a performance parameter), and it achieved good results, at 88.50%. The results in [47] were superior to those of [48], and the KNN performed well at 98%. Compard to the proposed work, no methods achieved 100%, but overall, their results were satisfactory. In [49–51], the authors used a Convolutional neural network (CNN) classifier, and this demonstrated a significant performance reach to 99% in [51] but with a huge complexity. Nonetheless, complexity of the proposed method is adequate enough for the application at hand. Because major time investment is required for processing of EEG signals. Once it is done, identification task is of moderate complexity and appropriate for such the soft real time application. Moreover, the techniques are applied on offline dataset.

Epilepsy is a neurological disease that affects approximately 1% of the world’s population. Seizures are a common symptom of epilepsy, and recently they have been used to diagnose the disease in computer science research by using machine learning techniques. To observe the seizures, the brain activity is recorded EEG signals, which is used as data in epilepsy classification processes. The goal of this work is to enhance the performance of detecting pediatric epilepsy by analyzing EEG data and classifying it into “epileptic” and “non-epileptic” seizures by applying various machine learning techniques as well as ensamble learning technique. To achieve this goal, the patients with 7 years old and less from the EEG CHB-MIT scalp database were analyzed by using EEGLAB (a toolbox in MATLAB) to remove the artifacts from the EEG signals. DWT with Daubechies family order was used since the signals were non stationary. In the current study, seven algorithms were used: KNN, DT, LD, SVM, ensemble learning, NB, and LR. The Ensemble learning and SVM models in this work outdo the performance over other models in diagnosing epileptic seizures in the literature. Future work could include creating a new method and tools to analyze signals accurately and easily. This would assist in diagnosing epileptic seizures and other neurological diseases in a shorter time and with fewer processing steps. In addition, further advanced algorithms could be applied, including deep learning and extreme learning machines, etc. More importantly, the research should be carried out on the detection of various types of seizures rather than just presence or absense.