Wavelet analysis for neonatal electroencephalographic seizures

Analysis of the impact of deep learning know-how and data in modelling neonatal EEG

The performance gains achieved by deep learning models nowadays are mainly attributed to the usage of ever larger datasets. In this study, we present and contrast the performance gains that can be achieved via accessing larger high-quality datasets versus the gains that can be achieved from harnessing the latest deep learning architectural and training advances. Modelling neonatal EEG is particularly affected by the lack of publicly available large datasets. It is shown that greater performance gains can be achieved from harnessing the latest deep learning advances than using a larger training dataset when adopting AUC as a metric, whereas using AUC90 or AUC-PR as metrics greater performance gains are achieved from using a larger dataset than harnessing the latest deep learning advances. In all scenarios the best performance is obtained by combining both deep learning advances and larger datasets. A novel developed architecture is presented that outperforms the current state-of-the-art model for the task of neonatal seizure detection. A novel method to fine-tune the presented model towards site-specific settings based on pseudo labelling is also outlined. The code and the weights of the model are made publicly available for benchmarking future model performances for neonatal seizure detection.

EEG in neonatal seizures: where to look and what to see

INTRODUCTION

Neonatal seizures are frequent and carry a detrimental prognostic outlook. Diagnosis is based on EEG confirmation. Classification has recently changed.

AREAS COVERED

We consulted original papers, book chapters, atlases, and reviews to provide a narrative overview on EEG characteristics of neonatal seizures. We searched PubMed, without time restrictions (last visited: 31 May 2022). Additional papers were extracted from the references list of selected papers. We describe the typical neonatal ictal EEG discharges morphology, location, and propagation, together with age-dependent features. Etiology-dependent electroclinical features, when identifiable, are presented for both acute symptomatic neonatal seizures and neonatal-onset epilepsies and developmental/epileptic encephalopathies. The few ictal variables known to predict long-term outcome have been discussed.

EXPERT OPINION

Multimodal neuromonitoring in critically ill newborns, high-density EEG, and functional neuroimaging might increase our insight into the neurophysiological bases of seizures in newborns. Increasing availability of long-term monitoring with conventional video-EEG and automated detection methods will allow clinicians and researchers to gather an ever expanding bulk of clinical and neurophysiological data to enhance accuracy with deep phenotyping. The latest classification proposal represents an input for critically revising our diagnostic abilities with respect to seizure definition, duration, and semiology, possibly further promoting clinical research.

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.

A method for AI assisted human interpretation of neonatal EEG

The study proposes a novel method to empower healthcare professionals to interact and leverage AI decision support in an intuitive manner using auditory senses. The method’s suitability is assessed through acoustic detection of the presence of neonatal seizures in electroencephalography (EEG). Neurophysiologists use EEG recordings to identify seizures visually. However, neurophysiological expertise is expensive and not available 24/7, even in tertiary hospitals. Other neonatal and pediatric medical professionals (nurses, doctors, etc.) can make erroneous interpretations of highly complex EEG signals. While artificial intelligence (AI) has been widely used to provide objective decision support for EEG analysis, AI decisions are not always explainable. This work developed a solution to combine AI algorithms with a human-centric intuitive EEG interpretation method. Specifically, EEG is converted to sound using an AI-driven attention mechanism. The perceptual characteristics of seizure events can be heard using this method, and an hour of EEG can be analysed in five seconds. A survey that has been conducted among targeted end-users on a publicly available dataset has demonstrated that not only does it drastically reduce the burden of reviewing the EEG data, but also the obtained accuracy is on par with experienced neurophysiologists trained to interpret neonatal EEG. It is also shown that the proposed communion of a medical professional and AI outperforms AI alone by empowering the human with little or no experience to leverage AI attention mechanisms to enhance the perceptual characteristics of seizure events.

Variability analysis of epileptic EEG using the maximal overlap discrete wavelet transform

Purpose

To determine if there is a difference in the wavelet variances of seizure and non-seizure channels in the EEG of an epileptic subject.

Methods

A six-level decomposition was applied using the Maximal Overlap Discrete Wavelet Transform (MODWT). The wavelet variance and 95% CIs were calculated for each level of the decomposition. The number of changes in variance for each level were found using a change-point detection method of Whitcher. The Kruskal-Wallis test was used to determine if there were differences in the median number of change points within channels and across frequency bands (levels).

Results

No distinctive pattern was found for the wavelet variances to differentiate the seizure and non-seizure channels. The seizure channels tended to have lower variances for each level and overall, but this pattern only held for one of the three seizure channels (RAST4). The median number of change points did not differ between the seizure and non-seizure channels either within each channel or across the frequency bands.

Conclusion

The use of the MODWT in examining the variances and changes in variance did not show specific patterns which differentiate between seizure and non-seizure channels.

Applications of advanced signal processing and machine learning in the neonatal hypoxic-ischemic electroencephalogram

Perinatal hypoxic-ischemic-encephalopathy significantly contributes to neonatal death and life-long disability such as cerebral palsy. Advances in signal processing and machine learning have provided the research community with an opportunity to develop automated real-time identification techniques to detect the signs of hypoxic-ischemic-encephalopathy in larger electroencephalography/amplitude-integrated electroencephalography data sets more easily. This review details the recent achievements, performed by a number of prominent research groups across the world, in the automatic identification and classification of hypoxic-ischemic epileptiform neonatal seizures using advanced signal processing and machine learning techniques. This review also addresses the clinical challenges that current automated techniques face in order to be fully utilized by clinicians, and highlights the importance of upgrading the current clinical bedside sampling frequencies to higher sampling rates in order to provide better hypoxic-ischemic biomarker detection frameworks. Additionally, the article highlights that current clinical automated epileptiform detection strategies for human neonates have been only concerned with seizure detection after the therapeutic latent phase of injury. Whereas recent animal studies have demonstrated that the latent phase of opportunity is critically important for early diagnosis of hypoxic-ischemic-encephalopathy electroencephalography biomarkers and although difficult, detection strategies could utilize biomarkers in the latent phase to also predict the onset of future seizures.

Diagnosis of seizures and encephalopathy using conventional EEG and amplitude integrated EEG

Seizures are more common in the neonatal period than at any other time of life, partly due to the relative hyperexcitability of the neonatal brain. Brain monitoring of sick neonates in the NICU using either conventional electroencephalography or amplitude integrated EEG is essential to accurately detect seizures. Treatment of seizures is important, as evidence increasingly indicates that seizures damage the brain in addition to that caused by the underlying etiology. Prompt treatment has been shown to reduce seizure burden with the potential to ameliorate seizure-mediated damage. Neonatal encephalopathy most commonly caused by a hypoxia-ischemia results in an alteration of mental status and problems such as seizures, hypotonia, apnea, and feeding difficulties. Confirmation of encephalopathy with EEG monitoring can act as an important adjunct to other investigations and the clinical examination, particularly when considering treatment strategies such as therapeutic hypothermia. Brain monitoring also provides useful early prognostic indicators to clinicians. Recent use of machine learning in algorithms to continuously monitor the neonatal EEG, detect seizures, and grade encephalopathy offers the exciting prospect of real-time decision support in the NICU in the very near future.

Characterisation of ictal and interictal states of epilepsy: A system dynamic approach of principal dynamic modes analysis

Epilepsy is a brain disorder characterised by the recurrent and unpredictable interruptions of normal brain function, called epileptic seizures. The present study attempts to derive new diagnostic indices which may delineate between ictal and interictal states of epilepsy. To achieve this, the nonlinear modeling approach of global principal dynamic modes (PDMs) is adopted to examine the functional connectivity of the temporal and frontal lobes with the occipital brain segment using an ensemble of paediatric EEGs having the presence of epileptic seizure. The distinct spectral characteristics of global PDMs are found to be in line with the neural rhythms of brain dynamics. Moreover, we find that the linear trends of associated nonlinear functions (ANFs) associated with the 2nd and 4th global PDMs (representing delta, theta and alpha bands) of Fp1–F3 may differentiate between ictal and interictal states of epilepsy. These findings suggest that global PDMs and their associated ANFs may offer potential utility as diagnostic neural measures for ictal and interictal states of epilepsy.

Clinical implementation of a neonatal seizure detection algorithm

Technologies for automated detection of neonatal seizures are gradually moving towards cot-side implementation. The aim of this paper is to present different ways to visualize the output of a neonatal seizure detection system and analyse their influence on performance in a clinical environment. Three different ways to visualize the detector output are considered: a binary output, a probabilistic trace, and a spatio-temporal colormap of seizure observability. As an alternative to visual aids, audified neonatal EEG is also considered. Additionally, a survey on the usefulness and accuracy of the presented methods has been performed among clinical personnel. The main advantages and disadvantages of the presented methods are discussed. The connection between information visualization and different methods to compute conventional metrics is established. The results of the visualization methods along with the system validation results indicate that the developed neonatal seizure detector with its current level of performance would unambiguously be of benefit to clinicians as a decision support system. The results of the survey suggest that a suitable way to visualize the output of neonatal seizure detection systems in a clinical environment is a combination of a binary output and a probabilistic trace. The main healthcare benefits of the tool are outlined. The decision support system with the chosen visualization interface is currently undergoing pre-market European multi-centre clinical investigation to support its regulatory approval and clinical adoption.

Robust neonatal EEG seizure detection through adaptive background modeling

Adaptive probabilistic modeling of the EEG background is proposed for seizure detection in neonates with hypoxic ischemic encephalopathy. The decision is made based on the temporal derivative of the seizure probability with respect to the adaptively modeled level of background activity. The robustness of the system to long duration "seizure-like" artifacts, in particular those due to respiration, is improved. The system was developed using statistical leave-one-patient-out performance assessment, on a large clinical dataset, comprising 38 patients of 1479 h total duration. The developed technique was then validated by a single test on a separate totally unseen randomized prospective dataset of 51 neonates totaling 2540 h of duration. By exploiting the proposed adaptation, the ROC area is increased from 93.4% to 96.1% (41% relative improvement). The number of false detections per hour is decreased from 0.42 to 0.24, while maintaining the correct detection of seizure burden at 70%. These results on the unseen data were predicted from the rigorous leave-one-patient-out validation and confirm the validity of our algorithm development process.

Behavioral and EEG effects of GABAergic manipulation of the nigro-tectal pathway in the Wistar audiogenic rat (WAR) strain II: an EEG wavelet analysis and retrograde neuronal tracer approach

The role of the substantia nigra pars reticulata (SNPr) and superior colliculus (SC) network in rat strains susceptible to audiogenic seizures still remain underexplored in epileptology. In a previous study from our laboratory, the GABAergic drugs bicuculline (BIC) and muscimol (MUS) were microinjected into the deep layers of either the anterior SC (aSC) or the posterior SC (pSC) in animals of the Wistar audiogenic rat (WAR) strain submitted to acoustic stimulation, in which simultaneous electroencephalographic (EEG) recording of the aSC, pSC, SNPr and striatum was performed. Only MUS microinjected into the pSC blocked audiogenic seizures. In the present study, we expanded upon these previous results using the retrograde tracer Fluorogold (FG) microinjected into the aSC and pSC in conjunction with quantitative EEG analysis (wavelet transform), in the search for mechanisms associated with the susceptibility of this inbred strain to acoustic stimulation. Our hypothesis was that the WAR strain would have different connectivity between specific subareas of the superior colliculus and the SNPr when compared with resistant Wistar animals and that these connections would lead to altered behavior of this network during audiogenic seizures. Wavelet analysis showed that the only treatment with an anticonvulsant effect was MUS microinjected into the pSC region, and this treatment induced a sustained oscillation in the theta band only in the SNPr and in the pSC. These data suggest that in WAR animals, there are at least two subcortical loops and that the one involved in audiogenic seizure susceptibility appears to be the pSC-SNPr circuit. We also found that WARs presented an increase in the number of FG+ projections from the posterior SNPr to both the aSC and pSC (primarily to the pSC), with both acting as proconvulsant nuclei when compared with Wistar rats. We concluded that these two different subcortical loops within the basal ganglia are probably a consequence of the WAR genetic background.

Inclusion of temporal priors for automated neonatal EEG classification

The aim of this paper is to use recent advances in the clinical understanding of the temporal evolution of seizure burden in neonates with hypoxic ischemic encephalopathy to improve the performance of automated detection algorithms. Probabilistic weights are designed from temporal locations of neonatal seizure events relative to time of birth. These weights are obtained by fitting a skew-normal distribution to the temporal seizure density and introduced into the probabilistic framework of the previously developed neonatal seizure detector. The results are validated on the largest available clinical dataset, comprising 816.7 h. By exploiting these priors, the receiver operating characteristic area is increased by 23% (relative) reaching 96.74%. The number of false detections per hour is decreased from 0.45 to 0.25, while maintaining the correct detection of seizure burden at 70%.

EEG signal description with spectral-envelope-based speech recognition features for detection of neonatal seizures

In this paper, features which are usually employed in automatic speech recognition (ASR) are used for the detection of seizures in newborn EEG. In particular, spectral envelope-based features, composed of spectral powers and their spectral derivatives are compared to the established feature set which has been previously developed for EEG analysis. The results indicate that the ASR features which model the spectral derivatives, either full-band or localized in frequency, yielded a performance improvement, in comparison to spectral-power-based features. Indeed it is shown here that they perform reasonably well in comparison with the conventional EEG feature set. The contribution of the ASR features was analyzed here using the support vector machines (SVM) recursive feature elimination technique. It is shown that the spectral derivative features consistently appear among the top-rank features. The study shows that the ASR features should be given a high priority when dealing with the description of the EEG signal.

EEG-based neonatal seizure detection with Support Vector Machines

OBJECTIVE

The study presents a multi-channel patient-independent neonatal seizure detection system based on the Support Vector Machine (SVM) classifier.

METHODS

A machine learning algorithm (SVM) is used as a classifier to discriminate between seizure and non-seizure EEG epochs. Two post-processing steps are proposed to increase both the temporal precision and the robustness of the system. The resulting system is validated on a large clinical dataset of 267 h of EEG data from 17 full-term newborns with seizures.

RESULTS

The performance of the system using event-based metrics is reported. The system showed the best up-to-date performance of a neonatal seizure detection system. The system was able to achieve an average good detection rate of ~89% with one false seizure detection per hour, ~96% with two false detections per hour, or ~100% with four false detections per hour. An analysis of errors revealed sources of misclassification in terms of both missed seizures and false detections.

CONCLUSIONS

The results obtained with the proposed SVM-based seizure detection system allow for its practical application in neonatal intensive care units.

SIGNIFICANCE

The proposed SVM-based seizure detection system can greatly assist clinical staff, in a neonatal intensive care unit, to interpret the EEG. The system allows control of the final decision by choosing different confidence levels which makes it flexible for clinical needs. The obtained results may provide a reference for future seizure detection systems.

Gaussian mixture models for classification of neonatal seizures using EEG

A real-time neonatal seizure detection system is proposed based on a Gaussian mixture model classifier. The system includes feature transformation techniques and classifier output postprocessing. The detector was evaluated on a database of 20 patients with 330 h of recordings. A detailed analysis of the choice of parameters for the detector is provided. A mean good detection rate of 79% was obtained with only 0.5 false detections per hour. A thorough review of all misclassified events was performed, from which a number of patterns causing false detections were identified.

Rapid oscillatory activity in delta brushes of premature and term neonatal EEG

We compared frequency and power of neonatal EEG delta brush rapid oscillatory activity (ROA) using multiple band frequency analysis (MBFA) in three groups; pre-term (PT, post-conceptional age 33-35.6 weeks, n=5); full-term (FT, 39.4-40.6 weeks, n=5) and pre-term or full-term with phenobarbital exposure (PB, n=5). Mean number of delta brushes analyzed was 29.4 (range 26-47) in PT, 20.8 (14-33) in FT and 20 (7-37) in PB. Mean frequency+/-standard deviation (s.d.) was 16.9+/-2.1 Hz (range 15-20 Hz) in PT, 17.3+/-1.9 Hz (15-20 Hz) in FT and 16.1+/-1.6 Hz (14-19 Hz) in PB. Mean power+/-s.d. was 22.9+/-6.2 microV(2) (range 16-39 microV(2)) in PT, 11.9+/-4.1 microV(2) (7-19 microV(2)) in FT and 17.1+/-6.2 microV(2) (9-26 microV(2)) in PB. Power was significantly higher in PT than FT (p<0.005). Power after merging PB into respective PT (PT', n=8) and FT (FT', n=7) groups, remained significantly higher in PT' (mean+/-s.d. 21.8+/-7.4 microV(2)) than FT' (11.4+/-3.6 microV(2)) (p<0.05). We characterise ROA in delta brushes in maturing neonates using MBFA, which may provide additional information for assessing future seizure recurrence and epilepsy risk.

A comparison of quantitative EEG features for neonatal seizure detection

OBJECTIVE

This study was undertaken to identify the best performing quantitative EEG features for neonatal seizures detection from a test set of 21.

METHODS

Each feature was evaluated on 1-min, artefact-free segments of seizure and non-seizure neonatal EEG recordings. The potential utility of each feature for neonatal seizure detection was determined using receiver operating characteristic analysis and repeated measures t-tests. A performance estimate of the feature set was obtained using a cross-fold validation and combining all features together into a linear discriminant classifier model.

RESULTS

Significant differences between seizure and non-seizure segments were found in 19 features for 17 patients. The best performing features for this application were the RMS amplitude, the line length and the number of local maxima and minima. An estimate of the patient independent classifier performance yielded a sensitivity of 81.08% and specificity of 82.23%.

CONCLUSIONS

The individual performances of 21 quantitative EEG features in detecting electrographic seizure in the neonate were compared and numerically quantified. Combining all features together into a classifier model led to superior performance than that provided by any individual feature taken alone.

SIGNIFICANCE

The results documented in this study may provide a reference for the optimum quantitative EEG features to use in developing and enhancing neonatal seizure detection algorithms.

An algorithm for the automatic detection of seizures in neonatal amplitude-integrated EEG

AIM

To develop and evaluate an algorithm for the automatic screening of electrographic neonatal seizures (ENS) in amplitude-integrated electroencephalography (aEEG) signals.

METHODS

CFM recordings were recorded in asphyxiated (near)term newborns. ENS of at least 60 sec were detected based on their characteristic pattern in the aEEG signal, an increase of its lower boundary. The algorithm was trained using five CFM recordings (training set) annotated by a neurophysiologist, observer1. The evaluation of the algorithm was based on eight different CFM recordings annotated by observer1 (test set observer 1) and an independent neurophysiologist, observer2 (test set observer 2).

RESULTS

The interobserver agreement between observer1 and 2 in interpreting ENS from the CFM recordings was high (G coefficient: 0.82). After dividing the eight CFM recordings into 1-min segments and classification in ENS or non-ENS, the intraclass correlation coefficient showed high correlations of the algorithm with both test sets (respectively, 0.95 and 0.85 with observer1 and 2). The algorithm showed in five recordings a sensitivity > or = 90% and approximately 1 false positive ENS per hour. However, the algorithm showed in three recordings much lower sensitivities: one recording showed ENSs of extremely high amplitude that were incorrectly classified by the algorithm as artefacts and two recordings suffered from low interobserver agreement.

CONCLUSION

This study shows the feasibility of automatic ENS screening based on aEEG signals and may facilitate in the bed-side interpretation of aEEG signals in clinical practice.

EEG wavelet analyses of the striatum–substantia nigra pars reticulata–superior colliculus circuitry: Audiogenic seizures and anticonvulsant drug administration in Wistar audiogenic rats (War strain)

The importance of the substantia nigra pars reticulata (SNPr), striatum (STR) and superior colicullus (SC) in the blockade of experimental seizures is well known. But, in audiogenic seizures (brainstem tonic-clonic seizures), the anticonvulsant activity of these nuclei is still controversial. In the present study we aimed to analyze the STR-SNPr-CS circuitry in the audiogenic seizures of Wistar audiogenic rat (WAR). Behavioral and electroencephalographic (EEG) data were collected from WARs under no treatment or injection with systemic (phenobarbital) or intracerebral (intranigral) drugs (muscimol and phenobarbital). The main EEG frequency oscillation of STR, SNPr and SC seen before, during and after audiogenic seizures or during seizure protection, was determinated with wavelet spectral analyses. This method allows the association between behavior and EEG (video-EEG). Audiogenic seizures last only for half a minute in average, suggesting that the interruptions of seizures are probably not due to exhaustion. Systemic phenobarbital caused an acute and dose-dependent behavioral and EEGraphic anticonvulsant effect both in WARs. The dose of phenobarbital 15mg/kg protected animals almost completely, without side effects such as ataxia and sedation. In our data, this endogenous "natural" seizure blockade (or termination) seems to be similar to the "forced" seizure abolition, like the one caused by a systemic non-ataxic phenobarbital dose, because in both cases an intense decrease in the EEG main frequency oscillation can be seen in SNPr and SC. Intranigral phenobarbital or muscimol did not protect animals, and actually induced an increase in the main EEG frequency oscillation in SC. The main finding of the present study is that, in contrast to what is well believed about the incapacity to control audiogenic seizures by the striato-nigro-tectal circuitry, we collected here evidences that these nuclei are involved in the ability to block these seizures. However, the striato-nigro-tectal circuitry in WARs, a genetically developed strain, seems to have different functional mechanisms when compared with normal rats.

Wavelet-Crosscorrelation Analysis: Non-Stationary Analysis of Neurophysiological Signals

OBJECTIVE

Wavelet-crosscorrelation analysis is a new application of wavelet analysis used to show the propagation of epileptiform discharges and to localize the corresponding lesions. We have shown previously that this analysis can help predict brain conditions statistically (Mizuno-Matsumoto et al. 2002). Our objective was to assess whether wavelet-crosscorrelation analysis reveals the initiation and propagation of epileptiform activity in human patients.

METHODS

The data obtained from three patients with simple partial seizures (SPS) using whole-head magnetoencephalography (MEG) were analyzed by the wavelet-crosscorrelation method. Wavelet-crosscorrelation coefficients (WCC), the coherent structure of each possible pair of signals from 64 MEG channels forvarious periods, and the time lag (TL) in two related signals, were ascertained.

RESULTS

We clearly demonstrated both localization of the irritative zone and propagation of the epileptiform discharges.

CONCLUSIONS

Wavelet-crosscorrelation analysis can help reveal and visualize the dynamic changes of brain conditions. The method of this analysis can compensate for other existing methods for the analysis of MEG, electroencephalography (EEG) or Elecotrocorticography (ECoG).

SIGNIFICANCE

Our proposed method suggests that revealing and visualizing the dynamic changes of brain conditions can help clinicians and even patients themselves better understand such conditions.