A machine-learning algorithm for detecting seizure termination in scalp EEG

Automated seizure activity tracking and onset zone localization from scalp EEG using deep neural networks

We propose a novel neural network architecture, SZTrack, to detect and track the spatio-temporal propagation of seizure activity in multichannel EEG. SZTrack combines a convolutional neural network encoder operating on individual EEG channels with recurrent neural networks to capture the evolution of seizure activity. Our unique training strategy aggregates individual electrode level predictions for patient-level seizure detection and localization. We evaluate SZTrack on a clinical EEG dataset of 201 seizure recordings from 34 epilepsy patients acquired at the Johns Hopkins Hospital. Our network achieves similar seizure detection performance to state-of-the-art methods and provides valuable localization information that has not previously been demonstrated in the literature. We also show the cross-site generalization capabilities of SZTrack on a dataset of 53 seizure recordings from 14 epilepsy patients acquired at the University of Wisconsin Madison. SZTrack is able to determine the lobe and hemisphere of origin in nearly all of these new patients without retraining the network. To our knowledge, SZTrack is the first end-to-end seizure tracking network using scalp EEG.

Raster plots machine learning to predict the seizure liability of drugs and to identify drugs

In vitro microelectrode array (MEA) assessment using human induced pluripotent stem cell (iPSC)-derived neurons holds promise as a method of seizure and toxicity evaluation. However, there are still issues surrounding the analysis methods used to predict seizure and toxicity liability as well as drug mechanisms of action. In the present study, we developed an artificial intelligence (AI) capable of predicting the seizure liability of drugs and identifying drugs using deep learning based on raster plots of neural network activity. The seizure liability prediction AI had a prediction accuracy of 98.4% for the drugs used to train it, classifying them correctly based on their responses as either seizure-causing compounds or seizure-free compounds. The AI also made concentration-dependent judgments of the seizure liability of drugs that it was not trained on. In addition, the drug identification AI implemented using the leave-one-sample-out scheme could distinguish among 13 seizure-causing compounds as well as seizure-free compound responses, with a mean accuracy of 99.9 ± 0.1% for all drugs. These AI prediction models are able to identify seizure liability concentration-dependence, rank the level of seizure liability based on the seizure liability probability, and identify the mechanism of the action of compounds. This holds promise for the future of in vitro MEA assessment as a powerful, high-accuracy new seizure liability prediction method.

Neural Decoding of EEG Signals with Machine Learning: A Systematic Review

Electroencephalography (EEG) is a non-invasive technique used to record the brain's evoked and induced electrical activity from the scalp. Artificial intelligence, particularly machine learning (ML) and deep learning (DL) algorithms, are increasingly being applied to EEG data for pattern analysis, group membership classification, and brain-computer interface purposes. This study aimed to systematically review recent advances in ML and DL supervised models for decoding and classifying EEG signals. Moreover, this article provides a comprehensive review of the state-of-the-art techniques used for EEG signal preprocessing and feature extraction. To this end, several academic databases were searched to explore relevant studies from the year 2000 to the present. Our results showed that the application of ML and DL in both mental workload and motor imagery tasks has received substantial attention in recent years. A total of 75% of DL studies applied convolutional neural networks with various learning algorithms, and 36% of ML studies achieved competitive accuracy by using a support vector machine algorithm. Wavelet transform was found to be the most common feature extraction method used for all types of tasks. We further examined the specific feature extraction methods and end classifier recommendations discovered in this systematic review.

Scalable and energy efficient seizure detection based on direct use of compressively-sensed EEG data on an ultra low power multi-core architecture

Extracting information from dense multi-channel neural sensors for accurate diagnosis of brain disorders necessitates computationally expensive and advanced signal processing approaches to analyze the massive volume of recorded data. Compressive Sensing (CS) is an efficient method for reducing the computational complexity and power consumption in the resource-constrained multi-site neural systems. However, reconstructing the signal from compressed measurements is computationally intensive, making it unsuitable for real-time applications such as seizure detection. In this paper, a seizure detection algorithm is proposed to overcome these limitations by circumventing the reconstruction phase and directly processing the compressively sampled EEG signals. The Lomb-Scargle Periodogram (LSP) is used to extract the spectral energy features of the compressed data. Performance of the seizure detector using non-linear support vector machine (SVM) classifier, tested on 24 patients of the CHB-MIT data-set for compression ratios (CR) of 1-64x, is 96-93%, 92-87%, 0.95-0.91, and <1 s for sensitivity, accuracy, the area under the curve, and latency, respectively. A power-efficient classification method based on the utilization of dual linear SVM classifiers is proposed. The proposed classification method based on the dual linear SVM classification achieved better classification performance compared to commonly used classifiers, such as K-nearest neighbor, random forest, artificial neural network, and linear SVM, while consuming low power in comparison to non-linear SVM kernels. The hardware-optimized implementation of this algorithm is proposed on a low-power multi-core SoC for near-sensor data analytics: Mr. Wolf. Optimized implementation of this algorithm on Mr. Wolf platform leads to detecting a seizure with an energy budget of 18.4 μJ and 3.9 μJ for a compression ratio of 24x using non-linear SVM classifier and the dual linear SVM based classification method, respectively.

Convolutional neural network for detection and classification of seizures in clinical data

Epileptic seizure detection and classification in clinical electroencephalogram data still is a challenge, and only low sensitivity with a high rate of false positives has been achieved with commercially available seizure detection tools, which usually are patient non-specific. Epilepsy patients suffer from severe detrimental effects like physical injury or depression due to unpredictable seizures. However, even in hospitals due to the high rate of false positives, the seizure alert systems are of poor help for patients as tools of seizure detection are mostly trained on unrealistically clean data, containing little noise and obtained under controlled laboratory conditions, where patient groups are homogeneous, e.g. in terms of age or type of seizures. In this study authors present the approach for detection and classification of a seizure using clinical data of electroencephalograms and a convolutional neural network trained on features of brain synchronisation and power spectrum. Various deep learning methods were applied, and the network was trained on a very heterogeneous clinical electroencephalogram dataset. In total, eight different types of seizures were considered, and the patients were of various ages, health conditions and they were observed under clinical conditions. Despite this, the classifier presented in this paper achieved sensitivity and specificity equal to 0.68 and 0.67, accordingly, which is a significant improvement as compared to the known results for clinical data. Graphical abstract.

Machine learning applications for electroencephalograph signals in epilepsy: a quick review

Machine learning (ML) is a fundamental concept in the field of state-of-the-art artificial intelligence (AI). Over the past two decades, it has evolved rapidly and been employed wildly in many fields. In medicine the widespread usage of ML has been observed in recent years. The present review examines various ML approaches for electroencephalograph (EEG) signal procession in epilepsy research, highlighting applications in the aspect of automated seizure detection, prediction and orientation. The present review also presents advantage, challenge and future direction of ML techniques in the analysis of EEG signals in epilepsy.

Using Relevance Feedback to Distinguish the Changes in EEG During Different Absence Seizure Phases

We carried out a series of statistical experiments to explore the utility of using relevance feedback on electroencephalogram (EEG) data to distinguish between different activity states in human absence epilepsy. EEG recordings from 10 patients with absence epilepsy are sampled, filtered, selected, and dissected from seizure-free, preseizure, and seizure phases. A total of 112 two-second 19-channel EEG epochs from 10 patients were selected from each phase. For each epoch, multiscale permutation entropy of the EEG data was calculated. The feature dimensionality was reduced by linear discriminant analysis to obtain a more discriminative and compact representation. Finally, a relevance feedback technique, that is, direct biased discriminant analysis, was applied to 68 randomly selected queries over nine iterations. This study is a first attempt to apply the statistical analysis of relevance feedback to the distinction of different EEG activity states in absence epilepsy. The average precision in the top 10 returned results was 97.5%, and the standard deviation suggested that embedding relevance feedback can effectively distinguish different seizure phases in absence epilepsy. The experimental results indicate that relevance feedback may be an effective tool for the prediction of different activity states in human absence epilepsy. The simultaneous analysis of multichannel EEG signals provides a powerful tool for the exploration of abnormal electrical brain activity in patients with epilepsy.

Band-sensitive seizure onset detection via CSP-enhanced EEG features

This paper presents two novel epileptic seizure onset detectors. The detectors rely on a common spatial pattern (CSP)-based feature enhancement stage that increases the variance between seizure and nonseizure scalp electroencephalography (EEG). The proposed feature enhancement stage enables better discrimination between seizure and nonseizure features. The first detector adopts a conventional classification stage using a support vector machine (SVM) that feeds the energy features extracted from different subbands to an SVM for seizure onset detection. The second detector uses logical operators to pool SVM seizure onset detections made independently across different EEG spectral bands. The proposed detectors exhibit an improved performance, with respect to sensitivity and detection latency, compared with the state-of-the-art detectors. Experimental results have demonstrated that the first detector achieves a sensitivity of 95.2%, detection latency of 6.43s, and false alarm rate of 0.59perhour. The second detector achieves a sensitivity of 100%, detection latency of 7.28s, and false alarm rate of 1.2per hour for the MAJORITY fusion method.

Early Seizure Detection Using Neuronal Potential Similarity: A Generalized Low-Complexity and Robust Measure

A novel approach using neuronal potential similarity (NPS) of two intracranial electroencephalogram (iEEG) electrodes placed over the foci is proposed for automated early seizure detection in patients with refractory partial epilepsy. The NPS measure is obtained from the spectral analysis of space-differential iEEG signals. Ratio between the NPS values obtained from two specific frequency bands is then investigated as a robust generalized measure, and reveals invaluable information about seizure initiation trends. A threshold-based classifier is subsequently applied on the proposed measure to generate alarms. The performance of the method was evaluated using cross-validation on a large clinical dataset, involving 183 seizure onsets in 1785 h of long-term continuous iEEG recordings of 11 patients. On average, the results show a high sensitivity of 86.9% (159 out of 183), a very low false detection rate of 1.4 per day, and a mean detection latency of 13.1 s from electrographic seizure onsets, while in average preceding clinical onsets by 6.3 s. These high performance results, specifically the short detection latency, coupled with the very low computational cost of the proposed method make it adequate for using in implantable closed-loop seizure suppression systems.

Robust and low complexity algorithms for seizure detection

This paper presents two low complexity and yet robust methods for automated seizure detection using a set of 2 intracranial Electroencephalogram (iEEG) recordings. Most current seizure detection methods suffer from high number of false alarms, even when designed to be subject-specific. In this study, the ratios of power between pairs of frequency bands are used as features to detect epileptic seizures. For comparison, these features are calculated from monopolar and bipolar iEEG recordings. Optimal thresholds are individually determined and used for each feature. Alarms are generated when the measure passes the threshold. The detector was applied to long-term continuous invasive recordings from 5 patients with refractory partial epilepsy, containing 54 seizures in 780 hours. On average, the results revealed 88.9% sensitivity, a very low false detection rate of 0.041 per hour (h(-1)) and detection latency of 9.4 seconds.

Performance-power consumption tradeoff in wearable epilepsy monitoring systems

Automated seizure detection methods can be used to reduce time and costs associated with analyzing large volumes of ambulatory EEG recordings. These methods however have to rely on very complex, power hungry algorithms, implemented on the system backend, in order to achieve acceptable levels of accuracy. In size, and therefore power-constrained EEG systems, an alternative approach to the problem of data reduction is online data selection, in which simpler algorithms select potential epileptiform activity for discontinuous recording but accurate analysis is still left to a medical practitioner. Such a diagnostic decision support system would still provide doctors with information relevant for diagnosis while reducing the time taken to analyze the EEG. For wearable systems with limited power budgets, data selection algorithm must be of sufficiently low complexity in order to reduce the amount of data transmitted and the overall power consumption. In this paper, we present a low-power hardware implementation of an online epileptic seizure data selection algorithm with encryption and data transmission and demonstrate the tradeoffs between its accuracy and the overall system power consumption. We demonstrate that overall power savings by data selection can be achieved by transmitting less than 40% of the data. We also show a 29% power reduction when selecting and transmitting 94% of all seizure events and only 10% of background EEG.

Information theoretic measures of network coordination in high-frequency scalp EEG reveal dynamic patterns associated with seizure termination

How a seizure terminates is still under-studied and, despite its clinical importance, remains an obscure phase of seizure evolution. Recent studies of seizure-related scalp EEGs at frequencies >100 Hz suggest that neural activity, in the form of oscillations and/or neuronal network interactions, may play an important role in preictal/ictal seizure evolution (Andrade-Valenca et al., 2011; Stamoulis et al., 2012). However, the role of high-frequency activity in seizure termination, is unknown, if it exists at all. Using information theoretic measures of network coordination, this study investigated ictal and immediate postictal neurodynamic interactions encoded in scalp EEGs from a relatively small sample of 8 patients with focal epilepsy and multiple seizures originating in temporal and/or frontal brain regions, at frequencies ≤ 100 Hz and >100 Hz, respectively. Despite some heterogeneity in the dynamics of these interactions, consistent patterns were also estimated. Specifically, in several seizures, linear or non-linear increase in high-frequency neuronal coordination during ictal intervals, coincided with a corresponding decrease in coordination at frequencies <100 Hz, suggesting a potential interference role of high-frequency activity, to disrupt abnormal ictal synchrony at lower frequencies. These changes in network synchrony started at least 20-30s prior to seizure offset, depending on the seizure duration. Opposite patterns were estimated at frequencies ≤ 100 Hz in several seizures. These results raise the possibility that high-frequency interference may occur in the form of progressive network coordination during the ictal interval, which continues during the postictal interval. This may be one of several possible mechanisms that facilitate seizure termination. In fact, inhibition of pairwise interactions between EEGs by other signals in their spatial neighborhood, quantified by negative interaction information, was estimated at frequencies ≤ 100 Hz, at least in some seizures.

Improving seizure detection performance reporting: analysing the duration needed for a detection

Improving seizure detection performance relies not only on novel signal processing approaches but also on new accurate, reliable and comparable performance reporting to give researchers better and fairer tools for understanding the true algorithm operation. This paper investigates the sensitivity of current performance metrics to the duration of data that must be marked as candidate seizure activity before a seizure detection is made. The results demonstrate that not all metrics are insensitive to this high level choice in the algorithm design, and provide new approaches for comparing between reported algorithm performances in a robust and reliable manner.

Seizure detection methods using a cascade architecture for real-time implantable devices

Implantable high-accuracy, and low-power seizure detection is a challenge. In this paper, we propose a cascade architecture to combine different seizure detection algorithms to optimize power and accuracy of the overall seizure detection system. The proposed architecture consists of a cascade of two seizure detection stages. In the first-stage detector, a lightweight (low-power) algorithm is used to detect seizure candidates with the understanding that there will be a high number of false positives. In the second-stage detector-and only for the seizure candidates detected in the first detector-a high-accuracy algorithm is used to eliminate the false positives. We show that the proposed cascade architecture can reduce power consumption of seizure detection by 80% with high accuracy, offering a suitable option for real-time implantable seizure detectors.

Discriminating between best performing features for seizure detection and data selection

Seizure detection algorithms have been developed to solve specific problems, such as seizure onset detection, occurrence detection, termination detection and data selection. It is thus inherent that each type of seizure detection algorithm would detect a different EEG characteristic (feature). However most feature comparison studies do not specify the seizure detection problem for which their respective features have been evaluated. This paper shows that the best features/algorithm bases are not the same for all types of algorithms but depend on the type of seizure detection algorithm wanted. To demonstrate this, 65 features previously evaluated for online seizure data selection are re-evaluated here for seizure occurrence detection, using performance metrics pertinent to each seizure detection type whilst keeping the testing methodology the same. The results show that the best performing features/algorithm bases for data selection and occurrence detection algorithms are different and that it is more challenging to achieve high detection accuracy for the former seizure detection type. This paper also provides a comprehensive evaluation of the performance of 65 features for seizure occurrence detection to aid future researchers in choosing the best performing feature(s) to improve seizure detection accuracy.

Compression-ratio-based seizure detection

For wireless seizure monitoring devices seizure detection and data compression are two critical tasks that need to be carefully designed against a very tight power budget to maximize the battery life. These two tasks are usually considered separately and algorithms for each are developed separately. In this paper, we consider having a single low-power algorithm for implementing both seizure detection and data compression. Towards that end, we investigated compression ratio (CR) as a seizure marker and show that the seizure detection can be achieved as a by-product of compression with no additional cost, and thus overall system power can be reduced. We show that the proposed method, the CR-based seizure detection has promising performance with 88% seizure detection accuracy, and 5.5 false positives per hour (FPh) without any computation overhead.