Epileptic EEG patterns recognition through machine learning techniques and relevant time–frequency features

Epileptic patterns: database and visualization

As part of the present study, we seek to evaluate and compare a selection of ML approaches so far available. For this purpose, we consider using the same clinical database already applied in our previously conducted studies [23, 35, 36]. More specifically, the intracranial EEG database was recorded at the Montreal Neurological Institute and Hospital (MNI) in Canada. The database contains EEG signals from an epileptic patient with intractable epilepsy. The recorded iEEG signals were sampled at a frequency of 2 KHz and involve 24 bipolar iEEG channels recorded during inter-ictal periods from the deepest contacts targeting the bilateral mesial temporal lobe (MTL) structures. During the recording process, a low-pass anti-aliasing filter with a cutoff frequency of 500 Hz was applied. For the experiments and simulations to be effectively conducted, a selection of techniques, approaches, methodologies, and codings were implemented using the Anaconda-Python environment and the Google Colab-based cloud platform. Noteworthy also, is that in our study context, both of the spikes and HFOs identification processes have been performed by an experienced neurologist’s visual annotation. Actually, the visual inspection of the intracranial EEG recording has revealed the presence of 416 distinguishable epileptic HFOs and 269 clearly identifiable epileptic spikes in total. These visually marked results should serve as a benchmark or gold standard, useful for evaluating the compared ML models respective performances. In addition to the experts marked intracranial EEG data, we have also incorporated simulated scalp EEG recordings to evaluate the tested ML approaches’ effectiveness in terms of Gaussian noise effect. Nonetheless, scalp EEG recordings often suffer from the persistence of various artefacts and inconsistencies, likely to reduce their reliability as gold standards. To address such a problem and establish a rather dependable benchmark effectively replicating the output of scalp EEG, we have considered introducing Gaussian noise to our intracranial EEG data, to be combined with the data at signal-to-noise ratios (SNR), ranging from 20 to −10 dB. Indeed, this methodology has enabled the establishment of a consistent and reliable benchmark, useful for validating the investigated ML models’ performance [36, 37]. To mitigate the potential bias associated with examining specific data, we have undertaken to apply the following criteria to select the appropriate channels useful for conducting the present research. Firstly, we have ensured that the channels turn out to display distinct interictal HFOs and spikes upon initial review. Secondly, only channels showcasing both frequent and infrequent epileptic events have been considered to capture a comprehensive range of activity. Moreover, the channels with varying background levels have been selected to incorporate diversity. It is worth mentioning, at this level, that the patient has provided informed consent in compliance with the MNI research ethics board.

Main parts of epileptic patterns’ recognition via machine learning techniques

Overall, the proposed methodology involves the following stages: data pre-processing, data splitting, feature extraction, model training, hyperparameter optimization, and model evaluation using testing data, which depict the automatic detection process of epileptic patterns by means of machine learning techniques. It is worth recalling that ML stands for a popular supervised learning model that initially involves a training phase followed by a testing phase. It is actually at this level of research that a breakdown of the methodology takes place. Indeed, while the first step (Step 1) involves a processing of the EEG data, the second step (Step 2) involves the EEG data splitting process, wherein, the available data is divided into training and testing subsets using k-fold cross-validation techniques. In effect, such a procedure is intended to ensure that the model is actually well trained and evaluated in terms of separate data. As to Step 3, it refers to the features’ extraction stage. Once data are split, relevant features are computed and extracted out of each labeled EEG event. Capturing significant characteristics of epileptic patterns, these features could include statistical measures, a spectral analysis, a time–frequency analysis, or any other temporal measures likely to help collect important information from the EEG signals. After the features are extracted, the dataset is transformed into a special format fit for being processed via ML algorithms. It is now represented under the form of a matrix or a set of vectors, wherein, each row represents an EEG event and each column stands for a single feature. Regarding Step 4, it designates the training phase, wherein, the ML model is trained on the EEG labeled data, which include both the extracted features and the corresponding class labels indicating whether an event is actually epileptic or not. The model is made to learn to recognize patterns and make predictions based on the input features. The ML algorithm applied in this specific methodology is aimed to construct the most effective ML model fit for classifying EEG data into two binary classes: (HFO, background) and (spike, background). Concerning Step 5, it involves the hyperparameters’ optimization process. It is actually by carefully tuning these hyperparameters that researchers and practitioners usually strive to improve the ML models’ accuracy, robustness, and ability to generalize latent data for the sake of effectively detecting epileptic patterns. Then comes Step 6, or the model evaluation stage, which follows the ML model training and optimization process, to evaluate its performance as to the previously detected latent data, often referred to as test data. This step serves to ensure the model’s generalization capacity and assess its capability in detecting epileptic patterns. Our suggested ML based methodology is illustrated through Figure 1, below, depicting a straightforward flowchart of the steps involved in the automatic detection of epileptic patterns via machine learning techniques. Ultimately, it is necessary to carefully evaluate the ML model’s suitability and effectiveness, along with its applicability to specific datasets, as a final step of the ML process.

Figure 1:

Flowchart illustrating the advanced ML methodology.

Machine-learning algorithms

ML is a data-driven process, liable to potentially model rather complex EEG data patterns, as compared to conventional methods. ML algorithms have been applied to draw significant information from EEG data for the purpose of distinguishing the various epilepsy associated brain statuses, and accurately identifying the epileptogenic zone. These algorithms utilize computed features to train ML classifiers, enabling them to identify the various data born EEG patterns, and make relevant predictions based on newly observed measurements. More recently, several machine learning methods have been widely used in this respect. Worth citing among these methods, one could well mention the support vector machines (SVM), multi-layer perceptron (MLP), Gaussian Naive Bayes (GNB), K nearest neighbors (KNN), decision tree (DT), AdaBoost (ABt), random forest (RF), logistic regression (LR), quadratic discriminant analysis (QDA), Gradient boosting (GB), and extra trees (ET).

The support vector machines (SVM): it is a widely recognized popular supervised learning algorithm fit for dealing with both of the classification (SVC) and regression (SVR) associated problems. SVM serves to retrieve an optimal hyperplane within a multidimensional space enabling to effectively separate the relevant features. In this respect, two primary types of SVC kernels are commonly used: the linear kernel, and the radial basis function (RBF) kernel. In regard to our specific study context, a preliminary test has demonstrated that the RBF kernel proves to perform rather effectively than the kernel option. Noteworthy, also, is that the RBF kernel involves two crucial parameters: regularized coefficient C and gamma. While gamma designates the kernel spread process, influencing the decision region, the C coefficient stands for the SVC classifier specific parameter representing a penalty for misclassifying a data point [40, 41]. With respect to our study case, the gamma value is set to a fixed value of 0.01, while the C parameter is set to the value of 1,000. Our choice of these two parameter values relates to our previously conducted research reference findings relevant to EEG patterns detection.

Multi-layer perceptron (MLP): The multi-layer perceptron (MLP) is characterized with the wide range of parameters it displays, the most significant among which one could cite the hidden layers’ size, the activation function, the weight optimization related solver, and the learning rate. The MLP associated challenge, however, lies in determining the number of hidden layers, whereby, information could flow easily from the input layer to the output layer, as well as the number of neurons persistent in each hidden layer. While it might seem that more hidden layers could provide more features and yield better classification results, a practical limit still persists. Indeed, increasing the number of hidden layers could result in the emergence of a serious overfitting problem, therefrom errors, such as false positives. Unfortunately, there is no established theory or straightforward method to determine the optimal architecture in terms of number of hidden layers and neurons in each layer. Inversely, however, using an insufficient number of hidden layers would result in an underfitting issue, wherein, the MLP turns out to be unable to model complex data, leading to poor performance. It is therefore essential to conduct a validation process to select the most appropriately fit MLP architecture [42, 43]. Accordingly, our administered preliminary test has revealed that the optimal internal architecture fit for processing the neural network should involve a single input layer, a single output layer, along with two hidden layers respectively enclosing 10 and five neurons. As to the implemented activation function, frequently applied in the relevant literature, it has been the relu. Regarding weight optimization procedure, we have opted for a learning rate of 0.0001 and the Adam solver.

Gaussian Naïve Bayes (Bayesian neural network: NN): in this respect, the wide range of applied features are ranked under the set of F= {F₁, F₂ … , F_N}, while the relating classes have been determined as C= {C₁, C₂ … , C_M}. At this level, the Naive Bayes theorem [44] turns out to be computed via the following equation:

As the classification context is of a binary type (0,1), the set probabilities of class 0 and class 1 are equally computed as follows:

As the evidence P(F_i)is exclusively applied for the probability normalization purposes, it is then dropped and considered constant. Subsequently, for each pair (F_i, C_j) the sample mean µ_ij and standard deviation σ_ijare computed. In regard to the testing phase, a new test point bearing the features of vector F_iis classified into a specific class C_jonce the likelihood probability product turns out to be maximized as:

K nearest neighbors (KNN): as one of the simplest ML techniques, it assumes that similar things are too close to each other. In this respect, the selected K value and distance metric stand as two major considerations when using the KNN technique, which rests on a number of steps: Firstly, select a K number of neighbors, and there is no predefined statistical method applicable for determining the most optimally favorable value of K. However, using error curves for various K values related data training and validation purposes could help in effectively determining the optimal K value. With respect to our particular study context, the optimal K value has been fixed at the range of three. Then, we proceed with computing the Euclidean distance of the chosen K number of neighbors is, as maintained through the following equation:

Take the K nearest neighbors as per the calculated Euclidean distance. Among these k neighbors, count the number of data points in each category. Finally, assign the new test point to that category for which the number of neighbors is the maximum [44], [45], [46].

Decision tree (DT): A decision tree is an important and well-established machine learning technique that has been used for a wide range of applications, especially for classification problems. Initially, the tree starts with a root node, preceded with a series of branches with intersections dubbed nodes, and ends up with leaves each corresponding to a class to predict. Tree depth is referred to as the maximum number of nodes persisting prior to reaching a leaf node. Within the binary classification context, the hierarchical structure of a decision tree rests on computing the information gain between the source S and feature A, such as:

Computation of the Shannon entropy is performed via the following equation:

wherein, S designates the samples’ set, p+ denotes the proportion of positive samples and p-refers to the proportion of negative samples, while P(S_i) denotes the probability of a persistent feature A associated symbol. Noteworthy, however, is that the Gini index might also be used instead of the entropy. It is determined by deducting the squared probabilities’ sum of each class from the unit. Mathematically, the Gini index is expressed as:

Where: P_i denotes the element’s probability for being classified under a distinct class. Thus, entropy turns out to be a rather complex procedure to undertake, owing mainly to its logarithms’ implementation requirement, therefore, the Gini index computation process proves to be a rather prompt undertaking [23, 47].

The AdaBoost (ABt) classifier: it proceeds with constructing a certain number of decision trees, N, throughout the training process. As a first step, the algorithm initially undertakes to construct a decision tree, wherein, any misclassified instances turn out to be identified as errors by the modeling procedure, to be subsequently utilized as inputs. This iterative process is reiterated until errors are effectively minimized, and data are accurately predicted. Hence, N models, or decision trees, are established based on the identified errors. This principle of utilizing error-correcting models stands as a common process; frequently applied by all types of boosting algorithms [48]. With respect to our special study case, the optimal number of estimators or decision trees has been set to twenty.

Random forest (RF): It is a supervised ML algorithm widely used for classification and regression problems solving purposes. It proceeds by constructing multiple decision trees on different samples of the dataset. Rather than relying on a single decision tree, the RF undertakes to combine the entirety of the trees emanating predictions to construct a final majority-vote based prediction. Once faced with a new test point, the algorithm undertakes to determine each decision tree relevant prediction and assigns the new data point to the majority of votes receiving category. Thus, increasing the number of trees in the random forest modeling process generally helps in improving accuracy and mitigating overfitting issues [44, 49, 50]. In regard to our study case, we have considered setting up the optimal number of estimators or decision trees to twenty.

Logistic regression (LR): it is a linear regression model fit for modeling regression tasks, though inappropriate for classification modeling purposes. Noteworthy, however, is that the LR stands as a powerful statistical algorithm specifically designed for predicting binary classes [51]. It helps in modeling the probability associated with two possible outcomes, and is commonly used for dealing with binary classification problems. It can also be further extended to handle multi-class classification tasks using the “one vs. all” framework. Rather than directly predicting classes, the LR undertakes to compute the probability of an event’s occurrence, thus providing a valuable tasks’ classification framework.

Discriminant analysis (DA): DA generally provides a three-fold analysis [52], respectively dubbed as a linear discriminant analysis (LDA), a quadratic discriminant analysis (QDA), and a regularized linear discriminant (RDA) analysis. In the context of our study case, the administered preliminary test reveals well that the QDA appears to be rather suited and effectively performing in detecting epileptic patterns. It is worth recalling, in this respect, that the QDA can be drawn from simple probabilistic models enabling to compute the class distribution conditional upon the data P(X|y=k)regarding each class k. Subsequently, predictions can be determined for each feature by means of the Bayes rule:

Gradient boosting (GB): Gradient boosting classifiers stand for a set of machine learning algorithms that combine multiple weak learning models, creating a strong predictive model. Each model in the sequence learns to correct the previous model made errors. The entirety of the models emanating predictions are then combined using simple averaging statistics. Actually, the GB models associated popularity is owed mainly to their noticeable effectiveness in classifying complex datasets [53], [54], [55], wherein, regression trees represent the base estimators even in classification tasks. Comparatively, however, the AdaBoost was the first boosting algorithm specifically designed with a proper loss function. In effect, gradient boosting is a rather generic algorithm allowing to retrieve approximate solutions to the additive modeling problem, rendering it more effectively flexible than the AdaBoost framework. Concerning our study case, an optimal number of 50 estimators or decision trees has been opted for.

Extra trees (ET) algorithm: it is a widely used random-forest type of machine learning algorithm that helps in jointly combining a set of multiple decision trees’ derived predictions. Although ET algorithm applies a rather simple decision-tree constructing approach, compared to random forest, it is often liable to achieve similar or even more effective performance, through its ability to create several unpruned decision trees from a single training dataset [56]. Regarding regression tasks, predictions are made by averaging the decision trees drawn predictions, while majority-voting process is used for classification task purposes. Concerning our specific study case, the number of estimators, or decision trees, has been set to 80 as the ideal option.

The ML models’ performance evaluation

Several applicable metrics are available for assessing the ML approaches achieved performance. Regarding the EEG analysis context, however, the gold standard is typically applied as the spikes and high-frequency oscillations (HFOs) provided EEG data relating annotations, skilfully performed by an experimented neurologist or epileptologist. On assessing the performance of automated detection algorithms, the detection results need be compared to the visual inspection results, which entails calculating the true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) parameters respective values. The true positives (TPs) rates correspond to the number of cases wherein both of the algorithm and the neurologist appear to correctly detect the epileptic pattern. As to the false positives (FPs), they refer to the number of cases wherein the algorithm proves to wrongly identify a background activity as an epileptic pattern, correctly identified by the expert. Concerning the true negatives (TNs), they designate the number of cases wherein both of the expert and the algorithm turn out to correctly identify the epileptic background activity. As regards the false negatives (FNs), they highlight the cases wherein the algorithm mistakenly classifies an epileptic pattern as a background activity, correctly identified by the specialist expert. Based on these four parameters, various evaluative metrics [57] turn out to be applicable to assess the various ML classifiers respective performance, such as:

Accordingly, the sensitivity metric serves to measure the algorithms’ capacity level to accurately detect, identify and classify any persistent spikes or high-frequency oscillations (HFOs), by quantifying the algorithm’s capability in effectively capturing such events, while ensuring that they are not missed or falsely identified. Hence:

The specificity metric provides useful information regarding the algorithm’s capability to accurately identify true negative events. More specifically; the balanced classification rate (BCR), or balanced accuracy metric, provides insights as to the algorithm’s ability in detecting the epileptic patterns and rejecting the non-epileptic patterns in the ensuing detection results, particularly, on dealing with unbalanced data. In fact, balanced accuracy denotes a metric reflecting the average value between sensitivity and specificity [57], and is computed as:

It is important to note that the BCR yielded value ranges from 0 to 100 %. Accordingly, a balanced accuracy of 100 % indicates a perfect classification performance, wherein, both sensitivity and specificity are actually maximized.