Neonatal seizure detection using atomic decomposition with a novel dictionary

XAI4EEG: spectral and spatio-temporal explanation of deep learning-based seizure detection in EEG time series

In clinical practice, algorithmic predictions may seriously jeopardise patients’ health and thus are required to be validated by medical experts before a final clinical decision is met. Towards that aim, there is need to incorporate explainable artificial intelligence techniques into medical research. In the specific field of epileptic seizure detection there are several machine learning algorithms but less methods on explaining them in an interpretable way. Therefore, we introduce XAI4EEG: an application-aware approach for an explainable and hybrid deep learning-based detection of seizures in multivariate EEG time series. In XAI4EEG, we combine deep learning models and domain knowledge on seizure detection, namely (a) frequency bands, (b) location of EEG leads and (c) temporal characteristics. XAI4EEG encompasses EEG data preparation, two deep learning models and our proposed explanation module visualizing feature contributions that are obtained by two SHAP explainers, each explaining the predictions of one of the two models. The resulting visual explanations provide an intuitive identification of decision-relevant regions in the spectral, spatial and temporal EEG dimensions. To evaluate XAI4EEG, we conducted a user study, where users were asked to assess the outputs of XAI4EEG, while working under time constraints, in order to emulate the fact that clinical diagnosis is done - more often than not - under time pressure. We found that the visualizations of our explanation module (1) lead to a substantially lower time for validating the predictions and (2) leverage an increase in interpretability, trust and confidence compared to selected SHAP feature contribution plots.

Neonatal EEG seizure detection using a new signal structural complexity measure based on matching pursuit decomposition with nonstationary dictionary

BACKGROUND AND OBJECTIVE

In newborns, it is often difficult to accurately differentiate between seizure and non-seizure based solely on clinical manifestations. This highlights the importance of electroencephalogram (EEG) in the recognition and management of neonatal seizures. This paper proposes an effective algorithm for the detection of neonatal seizure using multichannel EEG.

METHODS

Neonatal EEG changes morphology as it alternates between seizure and non-seizure states. A new signal complexity measure based on matching pursuit (MP) decomposition is proposed and used to detect transitions between these two states. The new measure, referred to as weighted structural complexity (WSC), was used for the detection of seizures in 30 newborn EEG records. Multiple IIR filters and an MP-based filter were designed and used to remove artifacts from the EEG data. Geometrical correlation between the EEG data channels was applied to reduce the number of false detections caused by remnant artifacts. The seizure detector's performance was assessed using several epoch-based (e.g., accuracy) and event-based (GDR = good detection rate and FD/h = false detections per hour) metrics.

RESULTS

Compared to the neurologist marking, the proposed detector was able to detect EEG seizures with 94% accuracy, 90.9% GDR, and 0.14 FD/h (95% CI: [0.06, 0.34]).

CONCLUSIONS

The high performance of the MP-based detector may have significant implications for the accurate diagnosis of neonatal seizures and the appropriate use of anticonvulsants and ongoing clinical assessment and care of the newborn.

Comparison of Empirical Mode Decomposition, Wavelets, and Different Machine Learning Approaches for Patient-Specific Seizure Detection Using Signal-Derived Empirical Dictionary Approach

Analysis of long-term multichannel EEG signals for automatic seizure detection is an active area of research that has seen application of methods from different domains of signal processing and machine learning. The majority of approaches developed in this context consist of extraction of hand-crafted features that are used to train a classifier for eventual seizure detection. Approaches that are data-driven, do not use hand-crafted features, and use small amounts of patients' historical EEG data for classifier training are few in number. The approach presented in this paper falls in the latter category, and is based on a signal-derived empirical dictionary approach, which utilizes empirical mode decomposition (EMD) and discrete wavelet transform (DWT) based dictionaries learned using a framework inspired by traditional methods of dictionary learning. Three features associated with traditional dictionary learning approaches, namely projection coefficients, coefficient vector and reconstruction error, are extracted from both EMD and DWT based dictionaries for automated seizure detection. This is the first time these features have been applied for automatic seizure detection using an empirical dictionary approach. Small amounts of patients' historical multi-channel EEG data are used for classifier training, and multiple classifiers are used for seizure detection using newer data. In addition, the seizure detection results are validated using 5-fold cross-validation to rule out any bias in the results. The CHB-MIT benchmark database containing long-term EEG recordings of pediatric patients is used for validation of the approach, and seizure detection performance comparable to the state-of-the-art is obtained. Seizure detection is performed using five classifiers, thereby allowing a comparison of the dictionary approaches, features extracted, and classifiers used. The best seizure detection performance is obtained using EMD based dictionary and reconstruction error feature and support vector machine classifier, with accuracy, sensitivity and specificity values of 88.2, 90.3, and 88.1%, respectively. Comparison is also made with other recent studies using the same database. The methodology presented in this paper is shown to be computationally efficient and robust for patient-specific automatic seizure detection. A data-driven methodology utilizing a small amount of patients' historical data is hence demonstrated as a practical solution for automatic seizure detection.

Seizure Detection Algorithms in Critically Ill Children: A Comparative Evaluation

OBJECTIVES

To evaluate the performance of commercially available seizure detection algorithms in critically ill children.

DESIGN

Diagnostic accuracy comparison between commercially available seizure detection algorithms referenced to electroencephalography experts using quantitative electroencephalography trends.

SETTING

Multispecialty quaternary children's hospital in Canada.

SUBJECTS

Critically ill children undergoing electroencephalography monitoring.

INTERVENTIONS

Continuous raw electroencephalography recordings (n = 19) were analyzed by a neurophysiologist to identify seizures. Those recordings were then converted to quantitative electroencephalography displays (amplitude-integrated electroencephalography and color density spectral array) and evaluated by six independent electroencephalography experts to determine the sensitivity and specificity of the amplitude-integrated electroencephalography and color density spectral array displays for seizure identification in comparison to expert interpretation of raw electroencephalography data. Those evaluations were then compared with four commercial seizure detection algorithms: ICTA-S (Stellate Harmonie Version 7; Natus Medical, San Carlos, CA), NB (Stellate Harmonie Version 7; Natus Medical), Persyst 11 (Persyst Development, Prescott, AZ), and Persyst 13 (Persyst Development) to determine sensitivity and specificity in comparison to amplitude-integrated electroencephalography and color density spectral array.

MEASUREMENTS AND MAIN RESULTS

Of the 379 seizures identified on raw electroencephalography, ICTA-S detected 36.9%, NB detected 92.3%, Persyst 11 detected 75.9%, and Persyst 13 detected 74.4%, whereas electroencephalography experts identified 76.5% of seizures using color density spectral array and 73.7% using amplitude-integrated electroencephalography. Daily false-positive rates averaged across all recordings were 4.7 with ICTA-S, 126.3 with NB, 5.1 with Persyst 11, 15.5 with Persyst 13, 1.7 with color density spectral array, and 1.5 with amplitude-integrated electroencephalography. Both Persyst 11 and Persyst 13 had sensitivity comparable to that of electroencephalography experts using amplitude-integrated electroencephalography and color density spectral array. Although Persyst 13 displayed the highest sensitivity for seizure count and seizure burden detected, Persyst 11 exhibited the best trade-off between sensitivity and false-positive rate among all seizure detection algorithms.

CONCLUSIONS

Some commercially available seizure detection algorithms demonstrate performance for seizure detection that is comparable to that of electroencephalography experts using quantitative electroencephalography displays. These algorithms may have utility as early warning systems that prompt review of quantitative electroencephalography or raw electroencephalography tracings, potentially leading to more timely seizure identification in critically ill patients.

Automatic Seizure Detection using Fully Convolutional Nested LSTM

The automatic seizure detection system can effectively help doctors to monitor and diagnose epilepsy thus reducing their workload. Many outstanding studies have given good results in the two-class seizure detection problems, but most of them are based on hand-wrought feature extraction. This study proposes an end-to-end automatic seizure detection system based on deep learning, which does not require heavy preprocessing on the EEG data or feature engineering. The fully convolutional network with three convolution blocks is first used to learn the expressive seizure characteristics from EEG data. Then these robust EEG features pertinent to seizures are presented as an input to the Nested Long Short-Term Memory (NLSTM) model to explore the inherent temporal dependencies in EEG signals. Lastly, the high-level features obtained from the NLSTM model are fed into the softmax layer to output predicted labels. The proposed method yields an accuracy range of 98.44–100% in 10 different experiments based on the Bonn University database. A larger EEG database is then used to evaluate the performance of the proposed method in real-life situations. The average sensitivity of 97.47%, specificity of 96.17%, and false detection rate of 0.487 per hour are yielded. For CHB–MIT Scalp EEG database, the proposed model also achieves a segment-level sensitivity of 94.07% with a false detection rate of 0.66 per hour. The excellent results obtained on three different EEG databases demonstrate that the proposed method has good robustness and generalization power under ideal and real-life conditions.

Time-Varying EEG Correlations Improve Automated Neonatal Seizure Detection

The aim of this study was to develop methods for detecting the nonstationary periodic characteristics of neonatal electroencephalographic (EEG) seizures by adapting estimates of the correlation both in the time (spike correlation; SC) and time–frequency domain (time–frequency correlation; TFC). These measures were incorporated into a seizure detection algorithm (SDA) based on a support vector machine to detect periods of seizure and nonseizure. The performance of these nonstationary correlation measures was evaluated using EEG recordings from 79 term neonates annotated by three human experts. The proposed measures were highly discriminative for seizure detection (median [Formula: see text]: 0.933 IQR: 0.821–0.975, median [Formula: see text]: 0.883 IQR: 0.707–0.931). The resultant SDA applied to multi-channel recordings had a median AUC of 0.988 (IQR: 0.931–0.998) when compared to consensus annotations, outperformed two state-of-the-art SDAs [Formula: see text] and was noninferior to the human expert for 73/79 of neonates.

Neonatal Seizure Detection Using Deep Convolutional Neural Networks

Identifying a core set of features is one of the most important steps in the development of an automated seizure detector. In most of the published studies describing features and seizure classifiers, the features were hand-engineered, which may not be optimal. The main goal of the present paper is using deep convolutional neural networks (CNNs) and random forest to automatically optimize feature selection and classification. The input of the proposed classifier is raw multi-channel EEG and the output is the class label: seizure/nonseizure. By training this network, the required features are optimized, while fitting a nonlinear classifier on the features. After training the network with EEG recordings of 26 neonates, five end layers performing the classification were replaced with a random forest classifier in order to improve the performance. This resulted in a false alarm rate of 0.9 per hour and seizure detection rate of 77% using a test set of EEG recordings of 22 neonates that also included dubious seizures. The newly proposed CNN classifier outperformed three data-driven feature-based approaches and performed similar to a previously developed heuristic method.

A dataset of neonatal EEG recordings with seizure annotations

Neonatal seizures are a common emergency in the neonatal intensive care unit (NICU). There are many questions yet to be answered regarding the temporal/spatial characteristics of seizures from different pathologies, response to medication, effects on neurodevelopment and optimal detection. The dataset presented in this descriptor contains EEG recordings from human neonates, the visual interpretation of the EEG by the human experts, supporting clinical data and codes to assist access. Multi-channel EEG was recorded from 79 term neonates admitted to the NICU at the Helsinki University Hospital. The median recording duration was 74 min (IQR: 64 to 96 min). The presence of seizures in the EEGs was annotated independently by three experts. An average of 460 seizures were annotated per expert in the dataset; 39 neonates had seizures and 22 were seizure free, by consensus. The dataset can be used as a reference set of neonatal seizures, in studies of inter-observer agreement and for the development of automated methods of seizure detection and other EEG analyses.

Patient-specific seizure detection in long-term EEG using signal-derived empirical mode decomposition (EMD)-based dictionary approach

OBJECTIVE

The objective of the work described in this paper is the development of a computationally efficient methodology for patient-specific automatic seizure detection in long-term multi-channel EEG recordings.

APPROACH

A novel patient-specific seizure detection approach based on a signal-derived empirical mode decomposition (EMD)-based dictionary approach is proposed. For this purpose, we use an empirical framework for EMD-based dictionary creation and learning, inspired by traditional dictionary learning methods, in which the EMD-based dictionary is learned from the multi-channel EEG data being analyzed for automatic seizure detection. We present the algorithm for dictionary creation and learning, whose purpose is to learn dictionaries with a small number of atoms. Using training signals belonging to seizure and non-seizure classes, an initial dictionary, termed as the raw dictionary, is formed. The atoms of the raw dictionary are composed of intrinsic mode functions obtained after decomposition of the training signals using the empirical mode decomposition algorithm. The raw dictionary is then trained using a learning algorithm, resulting in a substantial decrease in the number of atoms in the trained dictionary. The trained dictionary is then used for automatic seizure detection, such that coefficients of orthogonal projections of test signals against the trained dictionary form the features used for classification of test signals into seizure and non-seizure classes. Thus no hand-engineered features have to be extracted from the data as in traditional seizure detection approaches.

MAIN RESULTS

The performance of the proposed approach is validated using the CHB-MIT benchmark database, and averaged accuracy, sensitivity and specificity values of 92.9%, 94.3% and 91.5%, respectively, are obtained using support vector machine classifier and five-fold cross-validation method. These results are compared with other approaches using the same database, and the suitability of the approach for seizure detection in long-term multi-channel EEG recordings is discussed.

SIGNIFICANCE

The proposed approach describes a computationally efficient method for automatic seizure detection in long-term multi-channel EEG recordings. The method does not rely on hand-engineered features, as are required in traditional approaches. Furthermore, the approach is suitable for scenarios where the dictionary once formed and trained can be used for automatic seizure detection of newly recorded data, making the approach suitable for long-term multi-channel EEG recordings.

Bio-Signal Complexity Analysis in Epileptic Seizure Monitoring: A Topic Review

Complexity science has provided new perspectives and opportunities for understanding a variety of complex natural or social phenomena, including brain dysfunctions like epilepsy. By delving into the complexity in electrophysiological signals and neuroimaging, new insights have emerged. These discoveries have revealed that complexity is a fundamental aspect of physiological processes. The inherent nonlinearity and non-stationarity of physiological processes limits the methods based on simpler underlying assumptions to point out the pathway to a more comprehensive understanding of their behavior and relation with certain diseases. The perspective of complexity may benefit both the research and clinical practice through providing novel data analytics tools devoted for the understanding of and the intervention about epilepsies. This review aims to provide a sketchy overview of the methods derived from different disciplines lucubrating to the complexity of bio-signals in the field of epilepsy monitoring. Although the complexity of bio-signals is still not fully understood, bundles of new insights have been already obtained. Despite the promising results about epileptic seizure detection and prediction through offline analysis, we are still lacking robust, tried-and-true real-time applications. Multidisciplinary collaborations and more high-quality data accessible to the whole community are needed for reproducible research and the development of such applications.

Improved neonatal seizure detection using adaptive learning

In neonatal intensive care units performing continuous EEG monitoring, there is an unmet need for around-the-clock interpretation of EEG, especially for recognizing seizures. In recent years, a few automated seizure detection algorithms have been proposed. However, these are suboptimal in detecting brief-duration seizures (<; 30s), which frequently occur in neonates with severe neurological problems. Recently, a multi-stage neonatal seizure detector, composed of a heuristic and a data-driven classifier was proposed by our group and showed improved detection of brief seizures. In the present work, we propose to add a third stage to the detector in order to use feedback of the Clinical Neurophysiologist and adaptively retune a threshold of the second stage to improve the performance of detection of brief seizures. As a result, the false alarm rate (FAR) of the brief seizure detections decreased by 50% and the positive predictive value (PPV) increased by 18%. At the same time, for all detections, the FAR decreased by 35% and PPV increased by 5% while the good detection rate remained unchanged.

Epileptic Seizure Detection with Log-Euclidean Gaussian Kernel-Based Sparse Representation

Epileptic seizure detection plays an important role in the diagnosis of epilepsy and reducing the massive workload of reviewing electroencephalography (EEG) recordings. In this work, a novel algorithm is developed to detect seizures employing log-Euclidean Gaussian kernel-based sparse representation (SR) in long-term EEG recordings. Unlike the traditional SR for vector data in Euclidean space, the log-Euclidean Gaussian kernel-based SR framework is proposed for seizure detection in the space of the symmetric positive definite (SPD) matrices, which form a Riemannian manifold. Since the Riemannian manifold is nonlinear, the log-Euclidean Gaussian kernel function is applied to embed it into a reproducing kernel Hilbert space (RKHS) for performing SR. The EEG signals of all channels are divided into epochs and the SPD matrices representing EEG epochs are generated by covariance descriptors. Then, the testing samples are sparsely coded over the dictionary composed by training samples utilizing log-Euclidean Gaussian kernel-based SR. The classification of testing samples is achieved by computing the minimal reconstructed residuals. The proposed method is evaluated on the Freiburg EEG dataset of 21 patients and shows its notable performance on both epoch-based and event-based assessments. Moreover, this method handles multiple channels of EEG recordings synchronously which is more speedy and efficient than traditional seizure detection methods.