|Year : 2021 | Volume
| Issue : 2 | Page : 41-48
Applications of artificial intelligence in epilepsy
Pradeep Pankajakshan Nair1, Rajeswari Aghoram1, Madhuri Laxman Khilari2
1 Department of Neurology, JIPMER, Puducherry, India
2 Consultant Neurologist, Apollo Health City, Hyderabad, Telangana, India
|Date of Submission||02-May-2021|
|Date of Acceptance||03-Oct-2021|
|Date of Web Publication||30-Dec-2021|
Dr. Pradeep Pankajakshan Nair
Source of Support: None, Conflict of Interest: None
Epilepsy is a common neurological condition characterized by a predilection for recurrent seizures. It affects 3.0–11.9 persons per 1000 in India. The advent of machine learning and artificial intelligence (AI) has allowed us to harness computing power to evaluate enormous amounts of data to provide more definitive answers to many vexing questions in epilepsy such as the nature of a paroxysmal event, prediction of seizure, response to therapy, etc. In this article, we present an overview of AI and machine learning approaches to the diagnosis and management of epilepsy. We performed a MEDLINE search with both keywords (AI, epilepsy, Epilepsy, Machine learning, seizure) and MeSH terms (AI, Seizures) combined with Boolean operators. We present a narrative summary of the results. We initially discuss basic concepts regarding AI and its divisions, followed by a discussion of the role of AI in epilepsy from published studies particularly in the areas of diagnosis and classification of epilepsy; seizure detection and prediction; epileptogenesis; and management of epilepsy. Despite the growing popularity of AI in epilepsy, it should be remembered that these approaches are not without drawbacks. All machine learning approaches are data expensive and require a large computational capacity. This also has a bearing on the time taken for the development of these algorithms. AI is here to stay and influence all aspects of care for people with epilepsy (PWE) and it is necessary to equip ourselves to interface with these smart systems. This balance will help provide the best possible care to PWE.
Keywords: Artificial intelligence, epilepsy, machine learning, support vector machine, wearable electronic devices
|How to cite this article:|
Nair PP, Aghoram R, Khilari ML. Applications of artificial intelligence in epilepsy. Int J Adv Med Health Res 2021;8:41-8
| Introduction|| |
Epilepsy is a common neurological condition with an estimated prevalence of 3.0–11.9 per 1000 in India. Although this disease is very old, much about it remains poorly understood. Many questions continue to linger in the minds of physicians caring for these patients such as:
- Many conditions present with episodic neurological complaints. Interictal evaluations in such patients may be completely normal. Is there a way to identify if a given event is actually a seizure or not?
- Even with therapy using anti-seizure medications, about one-third of people with epilepsy (PWE) continue to experience seizures. Can we predict the response to anti-seizure medications?
- The occurrence of seizures in an individual is unpredictable. Can we predict these seizures and if so can we terminate them?
- Evaluation of people with refractory epilepsy for possible surgical candidacy and identifying epileptic networks requires multiple modalities of information (such as electroencephalogram [EEG], magnetic resonance imaging [MRI], and positron emission tomography [PET]) to be integrated and interpreted. Is there a way to do this accurately to delineate an epileptic network?
The answers to these and other such questions require sifting through large quantities of data. The advent of machine learning and artificial intelligence (AI) has allowed us to harness computing power to evaluate enormous amounts of data to provide more definitive answers. In this review, we begin by introducing the concepts of machine learning and AI. We then discuss how these techniques have been helpful to deal with the questions we raised earlier.
| Search Strategy|| |
We performed a MEDLINE search with both key words (such as AI, machine learning, prediction, diagnosis, epilepsy, seizure) and MeSH terms (AI, Machine learning, Epilepsy, Seizures) combined with Boolean operators. We did not use any filters or restrict by dates. We present a narrative summary of the results.
| Artificial Intelligence and Machine Learning|| |
AI is a branch of science which utilises constructs of human intelligence as algorithms in computers to help the machine solve problems in a more human way. The term AI was coined by John McCarthy in 1956. According to him AI “is the combination of science and engineering to make intelligent devices for human welfare.” It can be the combination of learning, perception, problem solving, linguistic logic, and reasoning. AI draws contributions from multiple disciplines such as philosophy, logics/mathematics, psychology, cognitive science, computation, neuroscience, etc. The use of AI in medicine has increased of late due to the easy availability of low-cost parallel computing systems.
Machine learning is a subfield of AI which is concerned with the programs that learn from experience in the absence of explicit instructions. [Figure 1] shows the general process flow for the use of these methods. Python, R, Matlab, Octave, Julia, C++ are some of the programming languages that support machine learning development. An integrated development environment (IDE) that is specific for each language, for instance, R studio for R, Pycharm, iPython or Jupyter for Python, etc., is often used. These IDE provide a more user-friendly interface for coding and piloting of such algorithms. Machine learning can be broadly classified into:
- Supervised learning with or without reinforcement
- Unsupervised learning with or without reinforcement
- Deep learning.
| Supervised Learning|| |
This method is based on comparisons between a set of labeled data (training data), and new or unlabeled data to predict, find, and speculate on the relationship between the computable data. This is similar to teaching someone to drive a car. The statistical methods commonly employed are regression, classification, etc., that help analyze data and recognize patterns. The algorithms available for supervised learning include: k-Nearest Neighbours (kNN), decision trees, naive Bayes, logistic regression, and support vector machines (SVM).
When multiple classifiers are possible, ensemble techniques may be used. An ensemble technique is a supervised learning method where several possible classifiers are weighed and combined to get a single classifier which outperforms individual classifiers. The ensemble method has been investigated from the late seventies by combining two linear regression models and also form the basis of award-winning adaBoost algorithm., A typical ensemble method consists of the following building blocks: training set, base inducer, diversity generator, and combiner. This technique has been used in identifying bad channels of intracranial EEG recording and locating epileptic brain regions., Jukic et al. suggested that ensemble machine learning methods can be used to differentiate the EEG signals from epileptogenic and nonepileptogenic brain regions with accuracy up to 98.9%.
| Unsupervised Learning|| |
Unlike in supervised learning, the task of feature extraction is also entrusted to the computer. Thus, after providing the data, the machine is asked “what can you tell me about the data?” or “what are the best groups you can make out of this data?” or “what features occur together most frequently in the data?” The algorithm that is widely used in unsupervised learning is k means clustering.
| Deep Learning|| |
Deep learning algorithms are also unsupervised methods and use the artificial neural network (ANN), that are modeled on human neural networks. They are different from traditional machine learning in that deep learning acquires data and derives latent data relationships without manual guidance [Figure 2]. ANN use a layered approach with one input and output layer, with many hidden layers that work to improve learning with the addition of more data. Multiple ANN networks are already available. If there are more than two hidden layers constituting ANN, it is specifically called deep neural network (DNN), and using those models to achieve the learning function is called deep learning. Deep learning approach has a significant advantage in automatically discovering discriminative features from data and learning compared to the traditional machine learning approach, which requires an additional process to extract the features manually and apply them as inputs.,,, In other words, training data can effectively serve as programming code for a deep learning model.
|Figure 2: Comparison between traditional machine learning technique and deep learning technique|
Click here to view
In addition to the general DNN model (i.e., multi-layer perceptron), various transformed models, such as convolutional neural network (CNN) specialized for the analysis of image data and recurrent neural network (RNN) specialized for the analysis of time-series data, have been actively employed., These have been further combined to develop hybrid models.
| Artificial Intelligence in Epilepsy|| |
The following are a few areas in epilepsy where AI has been found to be helpful:
- Automatic seizure detection and prediction
- Diagnosis and classification of seizures and epilepsies
- Understanding the epileptogenesis
- Optimizing the medical and surgical management including the prediction of seizure freedom and delineation of epileptic networks
- Development of wearable electronic devices for PWE.
| Automatic Seizure Detection|| |
An ideal seizure prediction model should have (a) sufficient buffering time preceding the clinical seizure activity in order for it to be useful, (b) minimal false positive alarms and (c) a seizure aborting intervention that is well tolerated and can be coupled to detected ictal activity in real time.
The usual phases of a seizure prediction system consist of (i) data acquisition, (ii) preprocessing of EEG signals (iii) feature extraction (iv) classification, and (v) validation of results in the postprocessing step. Preprocessing of the EEG signals is to remove the noise and to get a better signal-to-noise ratio. This involves noise removal, bad channel detection, and referencing. This may be achieved by converting signals from multiple EEG channels into a surrogate channel by averaging or by applying common spatial pattern filtering. Other techniques applied include the Butterworth bandpass filter, notch filter, wavelet transform, and empirical mode decomposition.,,,, The next step is feature extraction that can be the time domain, frequency domain, or both. These features may be linear or nonlinear. Some of the classification methods include kNN classifier, support vector machine, CNN, and extreme learning machine. Postprocessing methods include k fold cross-validation, Kalman filter, moving average filter, and Poisson process random predictor. Mahjoub et al. and Hussain demonstrated a higher accuracy of seizure detection by employing advanced pre-processing methods, different feature extraction techniques, and advanced parameter optimization approach.,
Automatic seizure detection was tried with EEG data adopting supervised ML techniques such as k NN, SVM, and deep ML techniques.,, Both scalp EEG and invasive EEG have been used. These map spatio-temporal characteristics of the EEG to classify it as ictal or interictal. Daoud and Bayoumi developed a model based on CNN, RNN, and deep convolutional autoencoder to learn the spatial and temporal features, respectively from raw EEG data. They utilized long-term surface EEG (23 channels) recordings of 8 pediatric patients with epilepsy. The model utilized the CNN for extracting spatial features and RNN to extract the temporal features. An equal number of interictal and preictal samples were used for training. The EEG signals were divided into nonoverlapping five seconds segments, and each segment contributed a training batch. The training was done N separate times, where N is the number of seizures for a specific patient. Eighty percent of the training data was used for the training set and 20% for the validation set. The model could predict the seizure 1 h before the onset with an accuracy rate of 99.6% and a false alarm rate of 0.004/h. The current approach is to develop patient-specific models rather than generalized models. These models identify ictal activity patterns akin to the fingerprint of a given patient. However such ML models require large amounts of training data and this is an important limitation for patient-specific models because the large number of seizures needs to be recorded from the individual to train the model.
In addition to EEG, data researchers have tried seizure prediction and seizure classification utilizing video recordings, surface electromyography (EMG), and mobile technology. Video data of the extremity movements in neonates was processed through neural network algorithm achieving a specificity and sensitivity of around 90% to identify neonatal seizures. When combined with EEG data, the sensitivity and specificity further improved. Using infrared camera data fed into ANN, efforts have been made to classify the nocturnal seizure recordings into tonic, tonic-clonic and focal motor seizures. The other data explored is accelerometer recordings from extremity using SVM algorithm achieving a sensitivity above 90%, false positivity rate of <1 per day and a clinical latency range of 10–30 seconds., Random forest and k-NN algorithms have also been used for interpreting the kinetic data., Emami et al. proposed a CNN based seizure detection model which learned the seizure and nonseizure states automatically and could achieve a reasonably high positive rate of >70%. The joint graph structure and representation learning network compiles the available information in EEG, task-related learning attained through gating and attention mechanisms helping the model focus on certain areas of input, for example, the intracranial EEG signals. This model outperforms the CNN and Graph CNN models; it can have implications in subtle preictal changes.
Kiral-Kornek et al., went one step ahead and created wearable devices (WD) with low power consumption with neuromorphic chips for the prediction of seizures.
| Diagnosis and Classification of Epilepsy|| |
Diagnosis and classification of seizures, epilepsy, and epilepsy syndrome form the cornerstone in the management of epilepsy. Traditionally, this is achieved by interpretation of multitude of information including clinical history, imaging, EEG, and genetic data. This is often time-consuming, laborious and difficult even for experienced epileptologists. ML approaches can play a potential role in simplifying this process. Recently in addition to classic ML techniques such as SVM, kNN, and deep learning algorithms such as CNN, hybrid methods such as the combination of CNN and RNN have been studied for this purpose.,, Liu et al. extracted the temporal and spatial features of ictal EEG from time Fourier transform for the CNN/RNN hybrid model and this model scored F1 scores of 97.4% and 97.2% in two data sets.
Resting-state functional MRI (fMRI) data fed into the SVM algorithm could achieve a specificity of 82.5% and sensitivity of 85% for classifying brain networks in epileptic persons and controls Amarreh et al. utilized diffusion tensor imaging data into SVM algorithm and could accurately distinguish patients with active epilepsy from those in remission and controls. Soriano et al., utilized resting-state magnetoencephalography (MEG) data and ANN to classify focal and generalized seizures achieving a sensitivity of around 88% and specificity of 86%. In addition to the imaging data, other clinical data have also been utilized for an epilepsy diagnosis. These include EMG recordings in ANN for juvenile myoclonic epilepsy, genetic-based data mining techniques and clinical semiology for differentiating temporal and extratemporal epilepsies, evoked potential data, and SVM algorithm for diagnosing generalized epilepsy.,,
| Understanding Epileptogenesis|| |
Epilepsy is being increasingly understood as a network disorder. The behavior and evolution of such networks is an important area of study as it improves the understanding of epileptogenesis. Traditionally animal models, lesional models, and cell-based models have been used. However, these are inherently limited in that they are only partial reflections of the epileptic brain. We can, with the help of ML technology, integrate multimodal data from neuroimaging, EEG, MEG, functional imaging to develop biophysical models of epileptic brains that can be used to simulate the effects of interventions. Jirsa et al. and Hashemi et al. build such models called “virtual epileptic patient” that can be personalized and can also predict seizure propagation and epileptogenicity.
| Optimizing Management|| |
Antiseizure medications are the mainstay of therapy of epilepsy. However, treatment decisions in epilepsy are based on multiple other factors such as type of epilepsy, frequency of seizures, etiology of epilepsy, genetic factors, etc. Tailoring the optimum management technique to a given patient with epilepsy is the aim of personalized or precision medicine. The AI and machine learning techniques have also been used for medical decision-making in epilepsy. Aslan et al. used ANN and 7 clinical parameters from 302 patients and achieved an accuracy of 91.1% for predicting seizure freedom, seizure reduction, or change in seizure frequency. Kimiskidis et al. used data from paired-pulse Transcranial magnetic stimulation-EEG recordings in a Naïve Bayes algorithm for identification of epilepsy and prediction of treatment response. He could achieve a mean sensitivity of 86% and specificity of 82% in distinguishing patients with genetic generalized epilepsies from controls, as well as a mean sensitivity of 80% and specificity of 73% in predicting seizure freedom at 12 months of follow-up. An et al. demonstrated that random forest algorithm trained with around 635 features from 175 to 735 records, could achieve an area under curve (AUC) of 76.4% and could identify patients with drug-resistant epilepsy at an average of 1.97 years before failing a second medication trial.
Machine learning algorithms have also been tried for predicting the response to individual medication using data from drug dispensing databases, EEG profile, and genetic profiles. Devinsky et al. trained a random forest classifier using clinical data from 34,990 patients to test the efficacy and tolerability of the medication regime.
One of the biggest deep learning applications is in molecular diagnostics and pharmacogenomics. Almost 70% of epileptic encephalopathies are attributed to genetic mutations. We can use it to predict an individual's response to a drug based on genetic configuration or mutations. A technique where raw data is fed into a 3 D array, from there onto neural networks, and then modeled to incorporate experimental data is often used. A large subset of generalized epilepsy has underlying genetic etiology. Therapeutic implications for conditions like Dravet syndrome and Tuberous sclerosis, where genetic mutations dictate the precision medicine pathways are known today. Petrovski et al. demonstrated a sensitivity of 91% and specificity of 82% to predict seizure freedom at 1 year of follow-up in a group of newly diagnosed persons with epilepsy using data of five single nucleotide polymorphisms from 115 patients using a k-NN algorithm. The sensitivity and specificity were around 81% for PWE on chronic therapy.,
ML algorithms can also be used to identify candidates for surgery and predict surgical outcomes, particularly after temporal lobectomy. Grigsby et al. tried to predict the anterior temporal lobectomy outcome using ANN algorithm and encoded clinical, electrographic, neuropsychologic, imaging, and surgical data from 65 patients. He could achieve a sensitivity of 80.0% and specificity of 83.3% in predicting Engel I outcomes and 100% and 85.7%, respectively, for Engel I or II outcomes taken together. Armañanzas et al. compared k-NN and Naïve Bayes algorithms in predicting Engel I versus II–III outcomes using presurgical data including neuropsychology battery and both classifiers could achieve a predictive accuracy of around 89.5%. Memarian et al compared SVM and Naive Bayes algorithms using preoperative clinical, electrophysiologic, and structural MRI data from 20 patients and SVM scored the highest accuracy of around 95% in predicting Engel I outcome. Resting-state fMRI with SVM algorithm could predict favorable outcome after vagal nerve stimulation with AUC of 86%. However, there are a few machine learning studies which have failed to demonstrate any advantage for predicting seizure freedom.,
The success of epilepsy surgery lies in the identification of epileptic networks in an individual.
Delineating epileptic networks
Detection and lateralization of temporal lobe epilepsy
Bakken et al. analyzed the MRI spectroscopy data with ANN algorithm for lateralizing the temporal lobe epilepsy (TLE). 18F-fluorodeoxyglucose-PET images data have also been studied for this purpose using ANN achieving 85% agreement with the expert review. Using other machine learning techniques researchers have tried to utilize data from single-photon emission computerized tomography, MRI sequences such as Diffusion tensor imaging (DTI), Diffusion-weighted imaging, diffusion kurtosis image. T1-weighted Fluid-attenuated inversion recovery (FLAIR), resting-state MEG for detection, and lateralization of TLE.,,,,,
Detecting epileptogenic zone
Epileptogenic zone is the region that is responsible for the onset and propagation of seizures and is necessary to sustain epileptiform discharges during a seizure. Epilepsy surgery aims at seizure freedom by removing this epileptogenic zone. Invasive EEG often relies on detecting high-frequency activity such as high-frequency oscillations (HFOs) to identify the epileptogenic zone. In implanted StereoEEG, fast activity could appear to originate simultaneously from multiple areas resulting in erroneous and subjective interpretations of the epileptogenic zone. Differentiation of epileptogenic zone from the region of propagation could also be difficult if relying on the fast activity alone. A machine learning system developed to recognize the onset of seizure in the form of sharp transients or spikes, suppression of lower frequency waveforms, and multiband concurrent fast activity resulted in near-accurate estimation of the epileptogenic zone.
Detecting seizure onset zone
Many groups have studied machine learning techniques to predict the seizure onset zone (SOZ). Elahian et al. built a model using logistic regression algorithm and trained the model with the data from intracranial recordings of seizures. The number of nonresected SOZ electrodes predicted the postoperative outcome. Varatharajah et al., studied a model which utilized interictal instead of ictal data for the prediction of SOZ. They trained the SVM algorithm with 3 biomarkers from iEEG recordings such as HFO, epileptiform discharges, and phase-amplitude coupling. When compared to clinically localized SOZ, this system could identify the SOZ accurately (AUC of the receiver operating characteristic curve = 0.73).
Identification of epileptic lesion on magnetic resonance imaging
Focal cortical dysplasia (FCD) can be missed in an MRI. In a study, 21% of type II FCDs had unremarkable MRI. In a proof of concept study, automated surface-based MRI morphometry along with machine learning demonstrated a sensitivity of 92% and specificity of 96% for FCD type II detection. In this study, surface-based quantitative approach was tested in 11 patients and a number of controls against expert lesion-tracing, location-specific intracranial EEG, histopathology, and surgical outcome.
Neurostimulation therapy for epilepsy is heavily dependent on machine learning methods. In fact, Neuropace is a Food and Drug Administration (FDA) approved implantable responsive neurostimulation device for nonresectable focal epilepsy. It records invasive EEG through implanted electrodes, that are processed in real time and if the ictal activity is detected it delivers small electric shocks to abort ictal activity. The site, stimulus intensity, frequency, etc., of other therapeutic neurostimulation devices such as vagal nerve stimulators can be predicted using ML techniques and so can response to therapy.
Apart from seizure freedom, such machine learning techniques have also been used to predict other important outcomes like language impairment in the case of FCD, total disease duration, and emergency department visit rate.,,
| Wearable Devices (WD)|| |
There are thousands of WD in the market for the measurement of various health parameters and it is increasing exponentially over time. WD can have applications in epilepsy also. Persons with generalized tonic-clonic seizures (GTCS) and bilateral tonic-clonic seizures are at great risk for injury due to seizure per se. Every year around 25% of persons with GTCS develop serious seizure-related accidental injury and GTCS increases the risk of sudden unexpected death in epilepsy multiple folds. Unpredictable nature of the seizure can be distressing and disabling for the individuals and caregivers. Furthermore, the patients and caregivers may miss the seizures which can affect the therapeutic decisions. In all these scenarios, automated seizure alarms and reliable seizure detection systems could be useful.
Although over 3000 articles are available in the literature on WD, the accuracy of these devices is not adequate to integrate them into formal medical practice. Another issue associated with the development of WD includes the technical feasibility including the need to hide the equipment, electrodes, etc. Broadly, the WDs are EEG or non EEG-based systems. Currently, the seizure detection WDs which are supported by phase 3 trials use non EEG data and biomarkers from exploratory studies and not from ML. Currently, the applicability of ML in the field of non EEG-WDs is limited to GTCS. The various non EEG parameters utilized by the WDs include heart rate, surface EMG, accelerometery, near-infrared spectroscopy, motion/pressure data, electrodermal activity, respiration and noise or a combination of these parameters.
Closed-loop systems such as responsive neurostimulation (RNS) which detect seizures and trigger a therapeutic stimulation is a promising treatment modality in PWE. A commercially available RNS, Neuropace, is approved by FDA for use as adjunctive therapy for reducing seizure frequency in individuals ≥18 years with refractory partial-onset seizures localized to one or two epileptogenic foci. The use of RNS is limited by its accuracy and efficacy of seizure detection. False-positive detection and triggers affect the battery life, cost, patient comfort, and may result in frequent procedures to replace the battery.
Getting regulatory approval for ML devices is not easy. US FDA has increased the scope to approve and regulate ML software. FDA has already approved a multimodal WD using accelerometry and electrodermal activity; the multimodal WD detected GTCS with high sensitivity (92%–100%) and low false alarm rate (0.2–1/day).
| Conclusion|| |
Increasingly, AI and ML techniques are being leveraged to support decision-making at all levels in the management of PWE. Wider use of these techniques will facilitate a more data-driven approach to understanding epilepsy and its management. However, despite the growing popularity of AI in medicine, and epilepsy in particular, it should be remembered that these approaches are not without drawbacks. Machine learning is constrained by its training set, such that the larger and more comprehensive the training data, the better will be the performance of the algorithm. Thus, such approaches are data expensive and require large computational capacity. This also has a bearing on the time taken for the development of these algorithms. Further, when using unsupervised learning and deep learning methods the features learnt and extracted by the algorithm may be inappropriate, often not easily understood, and impossible to work backward to understand. This is akin to a black box. In short, AI is here to stay and influence all aspects of care for the people with epilepsy and it is necessary to equip ourselves to interface with these smart systems. This balance will help to provide the best possible care to people with epilepsy.
Financial support and sponsorship
Conflicts of interest
There are no conflicts of interest.
| References|| |
Amudhan S, Gururaj G, Satishchandra P. Epilepsy in India I: Epidemiology and public health. Ann Indian Acad Neurol 2015;18:263-77.
] [Full text]
Rupali M, Amit P. A review paper on general concepts of “artificial intelligence and machine learning.” IARJSET 2017;4:79-82.
Rokach L. Ensemble-based classifiers. Artif Intell Rev 2010;33:1-39.
Tukey JW. Exploratory Data Analysis. Reading, Mass.: Addison-Wesley Publishing Co.; 1977.
Jukic S, Saracevic M, Subasi A, Kevric J. Comparison of ensemble machine learning methods for automated classification of focal and non-focal epileptic EEG signals. Mathematics 2020;8:1481-97.
Tuyisenge V, Trebaul L, Bhattacharjee M, Chanteloup-Forêt B, Saubat-Guigui C, Mîndruţă I, et al.
Automatic bad channel detection in intracranial electroencephalographic recordings using ensemble machine learning. Clin Neurophysiol 2018;129:548-54.
Valliani AA, Ranti D, Oermann EK. Deep learning and neurology: A systematic review. Neurol Ther 2019;8:351-65.
Abbasi B, Goldenholz DM. Machine learning applications in epilepsy. Epilepsia 2019;60:2037-47.
Richards BA, Lillicrap TP, Beaudoin P, Bengio Y, Bogacz R, Christensen A, et al.
A deep learning framework for neuroscience. Nat Neurosci 2019;22:1761-70.
Esteva A, Robicquet A, Ramsundar B, Kuleshov V, DePristo M, Chou K, et al.
A guide to deep learning in healthcare. Nat Med 2019;25:24-9.
Chartrand G, Cheng PM, Vorontsov E, Drozdzal M, Turcotte S, Pal CJ, et al.
Deep learning: A primer for radiologists. Radiographics 2017;37:2113-31.
Bigdely-Shamlo N, Mullen T, Kothe C, Su KM, Robbins KA. The PREP pipeline: Standardized preprocessing for large-scale EEG analysis. Front Neuroinform 2015;9:16.
Usman SM, Khalid S, Akhtar R, Bortolotto Z, Bashir Z, Qiu H. Using scalp EEG and intracranial EEG signals for predicting epileptic seizures: Review of available methodologies. Seizure 2019;71:258-69.
Adeli H, Ghosh-Dastidar S, Dadmehr N. A wavelet-chaos methodology for analysis of EEGs and EEG subbands to detect seizure and epilepsy. IEEE Trans Biomed Eng 2007;54:205-11.
Usman SM, Usman M, Fong S. Epileptic seizures prediction using machine learning methods. Comput Math Methods Med 2017;2017:9074759.
Bandarabadi M, Teixeira CA, Rasekhi J, Dourado A. Epileptic seizure prediction using relative spectral power features. Clin Neurophysiol 2015;126:237-48.
Rasekhi J, Mollaei MR, Bandarabadi M, Teixeira CA, Dourado A. Preprocessing effects of 22 linear univariate features on the performance of seizure prediction methods. J Neurosci Methods 2013;217:9-16.
Mahjoub C, Le Bouquin Jeannès R, Lajnef T, Kachouri A. Epileptic seizure detection on EEG signals using machine learning techniques and advanced preprocessing methods. Biomed Tech (Berl) 2020;65:33-50.
Hussain L. Detecting epileptic seizure with different feature extracting strategies using robust machine learning classification techniques by applying advance parameter optimization approach. Cogn Neurodyn 2018;12:271-94.
Zisheng Zhang, Parhi KK. Seizure detection using wavelet decomposition of the prediction error signal from a single channel of intra-cranial EEG. Annu Int Conf IEEE Eng Med Biol Soc 2014;2014:4443-6.
Siuly S, Kabir E, Wang H, Zhang Y. Exploring sampling in the detection of multicategory EEG signals. Comput Math Methods Med 2015;2015:576437.
Kang JH, Chung YG, Kim SP. An efficient detection of epileptic seizure by differentiation and spectral analysis of electroencephalograms. Comput Biol Med 2015;66:352-6.
Daoud H, Bayoumi MA. Efficient epileptic seizure prediction based on deep learning. IEEE Trans Biomed Circuits Syst 2019;13:804-13.
Karayiannis NB, Tao G, Xiong Y, Sami A, Varughese B, Frost JD Jr., et al.
Computerized motion analysis of videotaped neonatal seizures of epileptic origin. Epilepsia 2005;46:901-17.
Ogura Y, Hayashi H, Nakashima S, Zu Soh, Shibanoki T, Shimatani K, et al.
A neural network based infant monitoring system to facilitate diagnosis of epileptic seizures. Annu Int Conf IEEE Eng Med Biol Soc 2015;2015:5614-7.
Achilles F, Tombari F, Belagiannis V, Loesch AM, Noachtar S, Navab N. Convolutional neural networks for real-time epileptic seizure detection. Comput Methods Biocmech Biomed Eng Imaging Vis 2018;6:264-9.
Onorati F, Regalia G, Caborni C, Migliorini M, Bender D, Poh MZ, et al.
Multicenter clinical assessment of improved wearable multimodal convulsive seizure detectors. Epilepsia 2017;58:1870-9.
Milosevic M, Van de Vel A, Bonroy B, Ceulemans B, Lagae L, Vanrumste B, et al.
Automated detection of tonic-clonic seizures using 3-D accelerometry and surface electromyography in pediatric patients. IEEE J Biomed Health Inform 2016;20:1333-41.
Larsen SN, Conradsen I, Beniczky S, Sorensen HB. Detection of tonic epileptic seizures based on surface electromyography. Annu Int Conf IEEE Eng Med Biol Soc 2014;2014:942-5.
Borujeny GT, Yazdi M, Keshavarz-Haddad A, Borujeny AR. Detection of epileptic seizure using wireless sensor networks. J Med Signals Sens 2013;3:63-8.
] [Full text]
Emami A, Kunii N, Matsuo T, Shinozaki T, Kawai K, Takahashi H. Seizure detection by convolutional neural network-based analysis of scalp electroencephalography plot images. Neuroimage Clin 2019;22:101684.
Lian Q, Qi Y, Pan G, Wang Y. Learning graph in graph convolutional neural networks for robust seizure prediction. J Neural Eng 2020;17:035004.
Kiral-Kornek I, Roy S, Nurse E, Mashford B, Karoly P, Carroll T, et al.
Epileptic seizure prediction using big data and deep learning: Toward a mobile system. EBioMedicine 2018;27:103-11.
Liu T, Truong ND, Nikpour A, Zhou L, Kavehei O. Epileptic seizure classification with symmetric and hybrid bilinear models. IEEE J Biomed Health Inform 2020;24:2844-51.
Ahmedt-Aristizabal D, Nguyen K, Denman S, Sridharan S, Dionisio S, Fookes C. Deep motion analysis for epileptic seizure classification. Annu Int Conf IEEE Eng Med Biol Soc 2018;2018:3578-81.
Raghu S, Sriraam N, Temel Y, Rao SV, Kubben PL. A convolutional neural network based framework for classification of seizure types. Annu Int Conf IEEE Eng Med Biol Soc 2019;2019:2547-50.
Zhang J, Cheng W, Wang Z, Zhang Z, Lu W, Lu G, et al.
Pattern classification of large-scale functional brain networks: Identification of informative neuroimaging markers for epilepsy. PLoS One 2012;7:e36733.
Amarreh I, Meyerand ME, Stafstrom C, Hermann BP, Birn RM. Individual classification of children with epilepsy using support vector machine with multiple indices of diffusion tensor imaging. Neuroimage Clin 2014;4:757-64.
Soriano MC, Niso G, Clements J, Ortín S, Carrasco S, Gudín M, et al.
Automated detection of epileptic biomarkers in resting-state interictal MEG data. Front Neuroinform 2017;11:43.
Goker I, Osman O, Ozekes S, Baslo MB, Ertas M, Ulgen Y. Classification of juvenile myoclonic epilepsy data acquired through scanning electromyography with machine learning algorithms. J Med Syst 2012;36:2705-11.
Kassahun Y, Perrone R, De Momi E, Berghöfer E, Tassi L, Canevini MP, et al.
Automatic classification of epilepsy types using ontology-based and genetics-based machine learning. Artif Intell Med 2014;61:79-88.
Won D, Kim W, Chaovalitwongse WA, Tsai JJ. Altered visual contrast gain control is sensitive for idiopathic generalized epilepsies. Clin Neurophysiol 2017;128:340-8.
An S, Kang C, Lee HW. Artificial intelligence and computational approaches for epilepsy. J Epilepsy Res 2020;10:8-17.
Jirsa VK, Proix T, Perdikis D, Woodman MM, Wang H, Gonzalez-Martinez J, et al.
The virtual epileptic patient: Individualized whole-brain models of epilepsy spread. Neuroimage 2017;145:377-88.
Aslan K, Bozdemir H, Sahin C, Noyan Ogulata S. Can neural network able to estimate the prognosis of epilepsy patients according to risk factors? J Med Syst 2010;34:541-50.
Kimiskidis VK, Tsimpiris A, Ryvlin P, Kalviainen R, Koutroumanidis M, Valentin A, et al.
TMS combined with EEG in genetic generalized epilepsy: A phase II diagnostic accuracy study. Clin Neurophysiol 2017;128:367-81.
An S, Malhotra K, Dilley C, Han-Burgess E, Valdez JN, Robertson J, et al.
Predicting drug-resistant epilepsy – A machine learning approach based on administrative claims data. Epilepsy Behav 2018;89:118-25.
Devinsky O, Dilley C, Ozery-Flato M, Aharonov R, Goldschmidt Y, Rosen-Zvi M, et al.
Changing the approach to treatment choice in epilepsy using big data. Epilepsy Behav 2016;56:32-7.
Petrovski S, Szoeke CE, Sheffield LJ, D'souza W, Huggins RM, O'brien TJ. Multi-SNP pharmacogenomic classifier is superior to single-SNP models for predicting drug outcome in complex diseases. Pharmacogenet Genomics 2009;19:147-52.
Grigsby J, Kramer RE, Schneiders JL, Gates JR, Brewster Smith W. Predicting outcome of anterior temporal lobectomy using simulated neural networks. Epilepsia 1998;39:61-6.
Armañanzas R, Alonso-Nanclares L, Defelipe-Oroquieta J, Kastanauskaite A, de Sola RG, Defelipe J, et al.
Machine learning approach for the outcome prediction of temporal lobe epilepsy surgery. PLoS One 2013;8:e62819.
Memarian N, Kim S, Dewar S, Engel J Jr., Staba RJ. Multimodal data and machine learning for surgery outcome prediction in complicated cases of mesial temporal lobe epilepsy. Comput Biol Med 2015;64:67-78.
Goldenholz DM, Jow A, Khan OI, Bagić A, Sato S, Auh S, et al.
Preoperative prediction of temporal lobe epilepsy surgery outcome. Epilepsy Res 2016;127:331-8.
Yankam Njiwa J, Gray KR, Costes N, Mauguiere F, Ryvlin P, Hammers A. Advanced [(18) F] FDG and [(11) C] flumazenil PET analysis for individual outcome prediction after temporal lobe epilepsy surgery for hippocampal sclerosis. Neuroimage Clin 2015;7:122-31.
Lee JS, Lee DS, Kim SK, Lee SK, Chung JK, Lee MC, et al.
Localization of epileptogenic zones in F-18 FDG brain PET of patients with temporal lobe epilepsy using artificial neural network. IEEE Trans Med Imaging 2000;19:347-55.
Kerr WT, Nguyen ST, Cho AY, Lau EP, Silverman DH, Douglas PK, et al.
Computer-aided diagnosis and localization of lateralized temporal lobe epilepsy using interictal FDG-PET. Front Neurol 2013;4:31.
Lopes R, Steinling M, Szurhaj W, Maouche S, Dubois P, Betrouni N. Fractal features for localization of temporal lobe epileptic foci using SPECT imaging. Comput Biol Med 2010;40:469-77.
Rudie JD, Colby JB, Salamon N. Machine learning classification of mesial temporal sclerosis in epilepsy patients. Epilepsy Res 2015;117:63-9.
Keihaninejad S, Heckemann RA, Gousias IS, Hajnal JV, Duncan JS, Aljabar P, et al.
Classification and lateralization of temporal lobe epilepsies with and without hippocampal atrophy based on whole-brain automatic MRI segmentation. PLoS One 2012;7:e33096.
Del Gaizo J, Mofrad N, Jensen JH, Clark D, Glenn R, Helpern J, et al.
Using machine learning to classify temporal lobe epilepsy based on diffusion MRI. Brain Behav 2017;7:e00801.
Grinenko O, Li J, Mosher JC, Wang IZ, Bulacio JC, Gonzalez-Martinez J, et al.
A fingerprint of the epileptogenic zone in human epilepsies. Brain 2018;141:117-31.
Varatharajah Y, Berry B, Cimbalnik J, Kremen V, Van Gompel J, Stead M, et al.
Integrating artificial intelligence with real-time intracranial EEG monitoring to automate interictal identification of seizure onset zones in focal epilepsy. J Neural Eng 2018;15:046035.
Colombo N, Tassi L, Deleo F, Citterio A, Bramerio M, Mai R, et al.
Focal cortical dysplasia type IIa and IIb: MRI aspects in 118 cases proven by histopathology. Neuroradiology 2012;54:1065-77.
Thesen T, Quinn BT, Carlson C, Devinsky O, DuBois J, McDonald CR, et al.
Detection of epileptogenic cortical malformations with surface-based MRI morphometry. PLoS One 2011;6:e16430.
Thomas GP, Jobst BC. Critical review of the responsive neurostimulator system for epilepsy. Med Devices (Auckl) 2015;8:405-11.
Hartshorn A, Jobst B. Responsive brain stimulation in epilepsy. Ther Adv Chronic Dis 2018;9:135-42.
Paldino MJ, Hedges K, Zhang W. Independent contribution of individual white matter pathways to language function in pediatric epilepsy patients. Neuroimage Clin 2014;6:327-32.
Paldino MJ, Zhang W, Chu ZD, Golriz F. Metrics of brain network architecture capture the impact of disease in children with epilepsy. Neuroimage Clin 2017;13:201-8.
Grinspan ZM, Patel AD, Hafeez B, Abramson EL, Kern LM. Predicting frequent emergency department use among children with epilepsy: A retrospective cohort study using electronic health data from 2 centers. Epilepsia 2018;59:155-69.
Sveinsson O, Andersson T, Mattsson P, Carlsson S, Tomson T. Clinical risk factors in SUDEP: A nationwide population-based case-control study. Neurology 2020;94:e419-29.
Beniczky S, Karoly P, Nurse E, Ryvlin P, Cook M. Machine learning and wearable devices of the future. Epilepsia 2021;62 Suppl 2:S116-24.
Zhao X, Lhatoo SD. Seizure detection: Do current devices work? And when can they be useful? Curr Neurol Neurosci Rep 2018;18:40.
Regalia G, Onorati F, Lai M, Caborni C, Picard RW. Multimodal wrist-worn devices for seizure detection and advancing research: Focus on the Empatica wristbands. Epilepsy Res 2019;153:79-82.
[Figure 1], [Figure 2]