Epilepsy, a neurogenic disorder characterized by abrupt and transient disturbances in the body, manifests through sudden electrical bursts within the brain. These recurrent electrical discharges are commonly referred to as “epileptic seizures” or “ictal events” and more colloquially as “fits.” According to the survey, more than 50 million people worldwide are affected by epilepsy, representing approximately 2% of the global population (Moshé et al., 2015; World Health Organization, 2024). Consequently, the diagnosis of epilepsy is one of the utmost concerns. The most prevalent and reliable method for diagnosing epilepsy to date is recording the brain signals, primarily electroencephalography (EEG). Electrocorticography (ECoG) is also important and is particularly used for surgical intervention. Brain signals are primarily visually inspected by trained neuro-clinicians or neurophysiologists (Banerjee et al., 2009; Duque-muñoz et al., 2014). However, even standard EEG recordings for diagnosing epilepsy can last between 30 min to 6 h, rendering the visual scrutinization process very time-consuming. EEG data are often contaminated by motion artifacts, background noise, and interfering patterns from other neurological disorders. In developing countries, where the availability of neurophysiologists is low, diagnosing epilepsy using EEG becomes even more challenging and prone to manual errors (Banerjee et al., 2009). This situation makes the diagnosis of epilepsy using EEG very difficult and increases the likelihood of manual errors (Schuyler et al., 2007; Gandhi et al., 2010, 2012). Misdiagnosis of epilepsy often leads to the administration of improper drugs, which can prove to be life-threatening for patients. Hence, there is a dire need to develop an accurate, computationally fast, and robust tool for the diagnosis of epilepsy.

Significant research efforts have been dedicated to developing expert systems for the automated detection of ictal patterns or epileptic seizures in EEG. This section discusses some of the key findings reported in this area. Early breakthrough studies, such as those by Gotman (1999), relied on mimetic techniques that utilized descriptions provided by experienced neurophysiologists, including attributes such as crest, sharpness measures, time durations, inclinations, and more. However, this method proved to be inaccurate due to the heterogeneity among ictal patterns. Subsequent automated diagnosis methods for epilepsy involved the application of various frequency-domain techniques, such as the fast-Fourier transform (Polat and Güneş, 2007), and time-domain techniques, such as the empirical mode decomposition (Sharma and Pachori, 2015). Approaches based on FFT failed to capture the correct onset of ictal events due to their assumption of EEG as stationary despite its original non-stationary nature. The introduction of time–frequency domain techniques such as the short-time Fourier transform (Tzallas et al., 2009), especially wavelets (Gandhi et al., 2011; Swami et al., 2014; Faust et al., 2015; Edakawa et al., 2016; Swami et al., 2016a), aided in the development of many automated seizure detection models (Acharya et al., 2013).

Researchers have explored a wide variety of features for characterizing ictal patterns in EEG, including combinations of lower- and higher-order statistical parameters such as standard deviation (Gajic et al., 2015), kurtosis (Acharya et al., 2013), chaotic parameters such as correlation dimension and Lyapunov exponents (Tzallas et al., 2009; Acharya et al., 2013), Shannon entropy (Gandhi et al., 2011; Swami et al., 2017), energy (Swami et al., 2014, 2016a), and many more. Earlier, there was a general assumption that increasing the number of feature sets would improve the machine learning (ML) model’s accuracy. However, ample evidence suggests that increasing the dimensionality of the feature matrix could increase the computational cost of the expert system, while some features may even decrease the accuracy of the ML model. Hence, feature engineering for epilepsy diagnosis can present a paradox. Consequently, many researchers opt to exclusively utilize deep learning (DL)-based methods. While these methods perform optimally when trained with sufficiently large annotated/synthetic datasets (Pascual et al., 2020; Srinivasan et al., 2023; Dash et al., 2024), practical applications often encounter scarcity of such datasets and/or face challenges with the “black box” nature of the model. This opacity seldom instills confidence in clinicians to adopt new computer-aided diagnosis (CAD) systems. Therefore, identifying the underlying issues hindering the adoption of CAD in clinical settings is paramount.

Some seminal research efforts on optimal feature selection have yielded noteworthy results (Gandhi et al., 2012). However, the majority of research endeavors focusing on feature optimization and selection have centered around a single objective function (Acharya et al., 2013; Dhiman and Saini, 2014; Dash et al., 2024), primarily aimed at enhancing the accuracy of expert systems. This often results in the development of models with either exceedingly slow computation times and high accuracy or rapid models with lower accuracy. Feature selection methods could provide a faster alternative (Nara et al., 2016; Krishnan et al., 2024); however, those methods usually do not solve multiple objectives. Additionally, the scalability issues of these models are frequently overlooked and often lead to the selection of a maximum number of features. This is an important issue for realizing practicability in clinical settings (Swami et al., 2018; Tirumani et al., 2018). A compromise between sensitivity and specificity rates has often been observed (Mormann et al., 2007; Swami et al., 2016a), rendering the replication and practical application of results in clinical settings challenging. The scarcity of extensive, annotated datasets further exacerbates these challenges. This research seeks to bridge the gaps between these extremes.

Moreover, much of the literature in epilepsy research lacks a clear delineation of the procedure for constructing optimization functions, hindering future replicable research. The present study aims to demonstrate the application of evolutionary transfer optimization (ETO) (Tan et al., 2021) through an evolutionary multi-objective optimization (EMO) approach. The transfer optimization methodology is employed to train specific datasets, with testing datasets consisting entirely of out-of-sample signals. In this context, the EMO method employed is the non-dominant sorting genetic algorithm (NSGA)-II, aimed at simultaneously minimizing the number of feature sets and error rates. The knowledge transfer is directed toward minimizing classification error rates while maximizing accuracy and minimizing features, aligning with the concept of Maximizing Accuracy while Minimizing Features (MAMF). This concept can also be applied to address a broad spectrum of not only neurological but also various real-life challenges.

The following section of this article outlines the materials and methods employed. It comprehensively details the datasets utilized in this study and endeavors to present a procedural execution methodology for developing an expert system. Subsequently, the subsequent section of this article presents the results and discussion. Finally, the conclusions section summarizes the significant developments from this study and discusses its future scope.

Datasets from three different repositories were used in this study. The first dataset is freely available in the epilepsy EEG repository of the University of Bonn (UoB) (Andrzejak et al., 2001). The datasets within this repository have become a common benchmark for validating expert systems for detecting epileptic seizures. The datasets considered from this database are named set C, set D, and set E. Each of these subsets consists of intracranial EEG, i.e., electrocorticography (ECoG) segments acquired with a sampling rate of 173.61 Hz from five epilepsy patients, with each segment comprising 4,097 samples lasting for a duration of 23.6 s. The signals in set C were acquired from the region around the hippocampus location opposite the hemisphere of the epileptogenic zone, while the signals in set D were acquired from the epileptogenic zone. Both sets C and D consisted of interictal (non-ictal) events, whereas only the signals in set E consisted of epileptic seizure (ictal) events.

The second dataset considered in this study is available from our repository (Swami et al., 2019). The signals in this repository were collected from 10 epilepsy patients using the Grass Telefactor Comet AS40 machine. The acquisition was conducted at the Neurology & Sleep Centre (NSC) by a trained clinician under the supervision of a neurophysiologist. During acquisition, gold-plated scalp EEG electrodes were positioned according to the international 10–20 electrode placement system. The data collected at 200 Hz from all channels were segmented into signals lasting for a duration of 5.12 s, comprising 1,024 samples. The subsets named interictal and ictal events were considered in this study.

The third dataset considered in this study was collected from the database of Sri Ganga Ram Hospital (SGRH). The signals downloaded from this repository were collected from 20 epilepsy patients (Gandhi et al., 2011, 2012; Swami et al., 2016a). The sampling rate during acquisition was fixed at 400 Hz, and the Grass Telefactor Twin3 EEG machine was used for acquisition. The data from all channels were segmented into signals lasting for a duration of 10 s, comprising 4,000 samples. The subsets with interictal and ictal stages were considered from this database.

Samples of EEG segments from each of the three repositories are shown in Figures 1–3.

The process involves utilizing domain knowledge of the signals/datasets to extract a relevant set of attributes, referred to as features, which can then be fed into the ML classifier. The entire feature engineering procedure of this study is outlined as follows.

The GA is an evolutionary computing algorithm that is based on biological evolution. In multi-objective GA, more than one objective is optimized simultaneously to achieve the best-compromised solution (Deb, 2001; Smith, 2002). In this study, we have used the NSGA-II method for minimization of the number of required features and the error rate simultaneously. The steps involved in NSGA-II for minimizing the required objective functions are as follows:

Here, C A is the mean classification accuracy calculated using equation (7).

Where T P represents true positive values,

T N represents true negative values,

F P represents false positive values, and

F N represents false negative values.

Cross-validation by randomized sub-sampling: It is a statistical cross-validation procedure in which the original input data are randomly subdivided into training and testing sets. This process was iterated 10 times, and an equal number of training and testing sets were formed. During each iteration, the C A of the model was evaluated. Here, a generalized regression neural network (GRNN) (illustrated in the next section) was employed for classification. Finally, the mean C A (in %) was subtracted from 100 to evaluate the measure of mean E r (in %), which formed the second objective of this study.

The rank 1 Pareto front of the University of Bonn (UoB) datasets is depicted in Figures 5, 6. Figure 5 resulted from subjecting the NSGA-II method to 500 iterations with a population size of 10, while Figure 6 was generated using the same method but with a population size of 20. The selected solution after step xi is highlighted in both Figures 5, 6. Similarly, the results of the rank 1 Pareto front for the Neurology & Sleep Centre (NSC) datasets are presented in Figures 7, 8. Additionally, the results for the Sri Ganga Ram Hospital (SGRH) datasets are shown in Figures 9, 10. Figures 7, 9 were obtained when the NSGA-II method underwent 500 iterations with a population size of 10, whereas Figures 8, 10 were generated with a population size of 20 under the same method.

In artificial intelligence (AI), the classifier usually maps the function of input feature space to output class space. This could be mathematically expressed as f : R a → R b , where f represents the function, a is the dimension of input feature space, and b is the dimension of output feature. In neural networks, this mapping is achieved by the simulation of artificial neural clusters, such as the human brain. The GRNN model predicts the output/target class by predicting the probability density functions of the input data. GRNN has a memory-based architecture, and the solution is converged to the regression surface by following an asymptotic curve (Specht, 1991; Tomand and Schober, 2001). The parallel and one-pass learning architecture of the GRNN model is a lot faster than the recurrent neural networks.

The typical architecture of GRNN could be divided into four layers (shown in Figure 11). The input layer is the first layer of the GRNN model, which distributes the input data X among all the neurons after scaling. The second layer is the pattern or hidden layer. It applies the radial basis function to the probability density estimates. The spread of the radial basis function follows a Gaussian-shaped curve and is directly dependent on the value of the smoothing parameter σ . The measured values are passed into the next layer’s neurons, which consists of the summation units (one in the denominator and the other in the numerator). The denominator unit sums input weights d w j for all the samples from the pattern layer’s neurons. Similarly, the numerator unit sums the weights n w j for all the samples with actual targets of the pattern layer’s neurons. The final values of the numerator and the denominator have been indicated by ε 1 and ε 2 , respectively. The final layer of the GRNN model acts like an accumulator, which divides the ε 1 and ε 2 inputs to predict the output value Z ̂ X . This value could also be expressed by assuming function f X Y as the probability density of random variables X and Y . The density estimation is f ̂ X Y for samples Χ i and Y i , where i is the indices of the samples. The final predicted targets Y ̂ Χ for p number of signals is given by equation (8).

Where scaling function δ i 2 = Χ − Χ i T Χ − Χ i .

Based on equation (6), the outputs are similarly updated for Y ̂ ′ Χ values.

Performance parameters: The following parameters were considered for evaluating the performance of the expert system developed:

Classification Accuracy (CA): It is the measure of the expert system to correctly classify the signals in the testing set as given by equation (7).

Sensitivity (SN): It is the statistical measure of the expert system to correctly classify the ictal patterns in EEG. It is evaluated by equation (9).

Specificity (SP): It is the statistical measure of the expert system to correctly classify the non-ictal patterns in EEG. It is evaluated by equation (10).

Mathew’s Correlation Coefficient (MCC): It is a balanced statistical measure that considers both the sensitivity and specificity values of the expert system. It is calculated using equation (11) (Jurman et al., 2012). The value varies from -1 to 1. The closer the value of M C C toward 1, the better the prediction (Jurman et al., 2012).

Computation Time (CT): It is the measure of the total time elapsed for classifying the signals in the testing set. In this study, CT was measured in s.