Seizure detection with deep neural networks for review of two-channel electroencephalogram.

Ultra-long-term electroencephalographic (EEG) registration using minimally invasive low-channel devices is an emerging technology to assess sporadic seizure events. Highly sensitive automatic seizure detection algorithms are needed for semiautomatic evaluation of these prolonged recordings. We describe the design and validation of a deep neural network for two-channel seizure detection. The model is trained using EEG recordings from 590 patients in a publicly available seizure database. These recordings are based on the full 10-20 electrode system and include seizure annotations created by reviews of the full set of EEG channels. Validation was performed using 48 scalp EEG recordings from an independent epilepsy center and consensus seizure annotations from three neurologists. For each patient, a three-electrode subgroup (two channels with a common reference) of the full montage was selected for validation of the two-channel model. Mean sensitivity across patients of 88.8% and false positive rate across patients of 12.9/day were achieved. The proposed training approach is of great practical relevance, because true recordings from low-channel devices are currently available only in small numbers, and the generation of gold standard seizure annotations in two EEG channels is often difficult. The study demonstrates that automatic seizure detection based on two-channel EEG data is feasible and review of ultra-long-term recordings can be made efficient and effective.

Accurate identification of EEG recordings with interictal epileptiform discharges using a hybrid approach: Artificial intelligence supervised by human experts.

OBJECTIVE: To evaluate the diagnostic performance of artificial intelligence (AI)-based algorithms for identifying the presence of interictal epileptiform discharges (IEDs) in routine (20-min) electroencephalography (EEG) recordings. METHODS: We evaluated two approaches: a fully automated one and a hybrid approach, where three human raters applied an operational IED definition to assess the automated detections grouped into clusters by the algorithms. We used three previously developed AI algorithms: Encevis, SpikeNet, and Persyst. The diagnostic gold standard (epilepsy or not) was derived from video-EEG recordings of patients' habitual clinical episodes. We compared the algorithms with the gold standard at the recording level (epileptic or not). The independent validation data set (not used for training) consisted of 20-min EEG recordings containing sharp transients (epileptiform or not) from 60 patients: 30 with epilepsy (with a total of 340 IEDs) and 30 with nonepileptic paroxysmal events. We compared sensitivity, specificity, overall accuracy, and the review time-burden of the fully automated and hybrid approaches, with the conventional visual assessment of the whole recordings, based solely on unrestricted expert opinion. RESULTS: For all three AI algorithms, the specificity of the fully automated approach was too low for clinical implementation (16.67%; 63.33%; 3.33%), despite the high sensitivity (96.67%; 66.67%; 100.00%). Using the hybrid approach significantly increased the specificity (93.33%; 96.67%; 96.67%) with good sensitivity (93.33%; 56.67%; 76.67%). The overall accuracy of the hybrid methods (93.33%; 76.67%; 86.67%) was similar to the conventional visual assessment of the whole recordings (83.33%; 95% confidence interval [CI]: 71.48-91.70%; p > .5), yet the time-burden of review was significantly lower (p < .001). SIGNIFICANCE: The hybrid approach, where human raters apply the operational IED criteria to automated detections of AI-based algorithms, has high specificity, good sensitivity, and overall accuracy similar to conventional EEG reading, with a significantly lower time-burden. The hybrid approach is accurate and suitable for clinical implementation.

Early Epileptiform Discharges and Clinical Signs Predict Nonconvulsive Status Epilepticus on Continuous EEG.

BACKGROUND: Critical care continuous electroencephalography (CCEEG) represents the gold standard for detection of nonconvulsive status epilepticus (NCSE) in neurological critical care patients. It is unclear which findings on short-term routine EEG and which clinical parameters predict NCSE during subsequent CCEEG reliably. The aim of the present study was to assess the prognostic significance of changes within the first 30 min of EEG as well as of clinical parameters for the occurrence of NCSE during subsequent CCEEG. METHODS: Systematic analysis of the first 30 min and the remaining segments of prospective CCEEG recordings according to the ACNS Standardized Critical Care EEG Terminology and according to recently proposed NCSE criteria as well as review of clinical parameters of 85 consecutive neurological critical care patients. Logistic regression and binary classification tests were used to determine the most useful parameters within the first 30 min of EEG predicting subsequent NCSE. RESULTS: The presence of early sporadic epileptiform discharges (SED) and early rhythmic or periodic EEG patterns of "ictal-interictal uncertainty" (RPPIIIU) (OR 15.51, 95% CI 2.83-84.84, p = 0.002) and clinical signs of NCS (OR 18.43, 95% CI 2.06-164.62, p = 0.009) predicted NCSE on subsequent CCEEG. Various combinations of early SED, early RPPIIIU, and clinical signs of NCS showed sensitivities of 79-100%, specificities of 49-89%, and negative predictive values of 95-100% regarding the incidence of subsequent NCSE (p < 0.001). CONCLUSIONS: Early SED and early RPPIIIU within the first 30 min of EEG as well as clinical signs of NCS predict the occurrence of NCSE during subsequent CCEEG with high sensitivity and high negative predictive value and may be useful to select patients who should undergo CCEEG.

Relational local electroencephalography representations for sleep scoring.

Computational sleep scoring from multimodal neurophysiological time-series (polysomnography PSG) has achieved impressive clinical success. Models that use only a single electroencephalographic (EEG) channel from PSG have not yet received the same clinical recognition, since they lack Rapid Eye Movement (REM) scoring quality. The question whether this lack can be remedied at all remains an important one. We conjecture that predominant Long Short-Term Memory (LSTM) models do not adequately represent distant REM EEG segments (termed epochs), since LSTMs compress these to a fixed-size vector from separate past and future sequences. To this end, we introduce the EEG representation model ENGELBERT (electroEncephaloGraphic Epoch Local Bidirectional Encoder Representations from Transformer). It jointly attends to multiple EEG epochs from both past and future. Compared to typical token sequences in language, for which attention models have originally been conceived, overnight EEG sequences easily span more than 1000 30 s epochs. Local attention on overlapping windows reduces the critical quadratic computational complexity to linear, enabling versatile sub-one-hour to all-day scoring. ENGELBERT is at least one order of magnitude smaller than established LSTM models and is easy to train from scratch in a single phase. It surpassed state-of-the-art macro F1-scores in 3 single-EEG sleep scoring experiments. REM F1-scores were pushed to at least 86%. ENGELBERT virtually closed the gap to PSG-based methods from 4-5 percentage points (pp) to less than 1 pp F1-score.

Detecting temporal lobe seizures in ultra long-term subcutaneous EEG using algorithm-based data reduction.

OBJECTIVE: Ultra long-term monitoring with subcutaneous EEG (sqEEG) offers objective outpatient recording of electrographic seizures as an alternative to self-reported epileptic seizure diaries. This methodology requires an algorithm-based automatic seizure detection to indicate periods of potential seizure activity to reduce the time spent on visual review. The objective of this study was to evaluate the performance of a sqEEG-based automatic seizure detection algorithm. METHODS: A multicenter cohort of subjects using sqEEG were analyzed, including nine people with epilepsy (PWE) and 12 healthy subjects, recording a total of 965 days. The automatic seizure detections of a deep-neural-network algorithm were compared to annotations from three human experts. RESULTS: Data reduction ratios were 99.6% in PWE and 99.9% in the control group. The cross-PWE sensitivity was 86% (median 80%, range 69-100% when PWE were evaluated individually), and the corresponding median false detection rate was 2.4 detections per 24 hours (range: 2.0-13.0). CONCLUSIONS: Our findings demonstrated that step one in a sqEEG-based semi-automatic seizure detection/review process can be performed with high sensitivity and clinically applicable specificity. SIGNIFICANCE: Ultra long-term sqEEG bears the potential of improving objective seizure quantification.

Activation patterns of interictal epileptiform discharges in relation to sleep and seizures: An artificial intelligence driven data analysis.

OBJECTIVE: To quantify effects of sleep and seizures on the rate of interictal epileptiform discharges (IED) and to classify patients with epilepsy based on IED activation patterns. METHODS: We analyzed long-term EEGs from 76 patients with at least one recorded epileptic seizure during monitoring. IEDs were detected with an AI-based algorithm and validated by visual inspection. We then used unsupervised clustering to characterize patient sub-cohorts with similar IED activation patterns regarding circadian rhythms, deep sleep activation, and seizure occurrence. RESULTS: Five sub-cohorts with similar IED activation patterns were found: "Sporadic" (14%, n = 10) without or few IEDs, "Continuous" (32%, n = 23) with weak circadian/deep sleep or seizure modulation, "Nighttime & seizure activation" (23%, n = 17) with high IED rates during normal sleep times and after seizures but without deep sleep modulation, "Deep sleep" (19%, n = 14) with strong IED modulation during deep sleep, and "Seizure deactivation" (12%, n = 9) with deactivation of IEDs after seizures. Patients showing "Deep sleep" IED pattern were diagnosed with temporal lobe epilepsy in 86%, while 80% of the "Sporadic" cluster were extratemporal. CONCLUSIONS: Patients with epilepsy can be characterized by using temporal relationships between rates of IEDs, circadian rhythms, deep sleep and seizures. SIGNIFICANCE: This work presents the first approach to data-driven classification of epilepsy patients based on their fully validated temporal pattern of IEDs.

Standardization of neurophysiology signal data into the DICOM® standard.

A standard format for neurophysiology data is urgently needed to improve clinical care and promote research data exchange. Previous neurophysiology format standardization projects have provided valuable insights into how to accomplish the project. In medical imaging, the Digital Imaging and Communication in Medicine (DICOM) standard is widely adopted. DICOM offers a unique environment to accomplish neurophysiology format standardization because neurophysiology data can be easily integrated with existing DICOM-supported elements such as video, ECG, and images and also because it provides easy integration into hospital Picture Archiving and Communication Systems (PACS) long-term storage systems. Through the support of the International Federation of Clinical Neurophysiology (IFCN) and partners in industry, DICOM Working Group 32 (WG-32) has created an initial set of standards for routine electroencephalography (EEG), polysomnography (PSG), electromyography (EMG), and electrooculography (EOG). Longer and more complex neurophysiology data types such as high-definition EEG, long-term monitoring EEG, intracranial EEG, magnetoencephalography, advanced EMG, and evoked potentials will be added later. In order to provide for efficient data compression, a DICOM neurophysiology codec design competition will be held by the IFCN and this is currently being planned. We look forward to a future when a common DICOM neurophysiology data format makes data sharing and storage much simpler and more efficient.

An artificial intelligence-based EEG algorithm for detection of epileptiform EEG discharges: Validation against the diagnostic gold standard.

OBJECTIVE: To validate an artificial intelligence-based computer algorithm for detection of epileptiform EEG discharges (EDs) and subsequent identification of patients with epilepsy. METHODS: We developed an algorithm for automatic detection of EDs, based on a novel deep learning method that requires a low amount of labeled EEG data for training. Detected EDs are automatically grouped into clusters, consisting of the same type of EDs, for rapid visual inspection. We validated the algorithm on an independent dataset of 100 patients with sharp transients in their EEG recordings (54 with epilepsy and 46 with non-epileptic paroxysmal events). The diagnostic gold standard was derived from the video-EEG recordings of the patients' habitual events. RESULTS: The algorithm had a sensitivity of 89% for identifying EEGs with EDs recorded from patients with epilepsy, a specificity of 70%, and an overall accuracy of 80%. CONCLUSIONS: Automated detection of EDs using an artificial intelligence-based computer algorithm had a high sensitivity. Human (expert) supervision is still necessary for confirming the clusters of detected EDs and for describing clinical correlations. Further studies on different patient populations will be needed to confirm our results. SIGNIFICANCE: The automated algorithm we describe here is a useful tool, assisting neurophysiologist in rapid assessment of EEG recordings.

Deep learning for estimation of functional brain maturation from EEG of premature neonates.

EEG monitoring of early brain function and development in neonatal intensive care units may help to identify infants with high risk of serious neurological impairment and to assess brain maturation for evaluation of neurodevelopmental progress. Automated analysis of EEG data makes continuous evaluation of brain activity fast and accessible. A convolutional neural network (CNN) for regression of EEG maturational age of premature neonates from marginally preprocessed serial EEG recordings is proposed. The CNN was trained and validated using 141 EEG recordings from 43 preterm neonates born below 28 weeks of gestation with normal neurodevelop-mental outcome at 12 months of corrected age. The estimated functional brain maturation between the first and last EEG recording increased in each patient. On average over 96% of repeated measures within an infant had an increasing EEG maturational age according to the post menstrual age at EEG recording time. Our algorithm has potential to be deployed to support neonatologists for accurate estimation of functional brain maturity in premature neonates.

Automated Long-Term EEG Review: Fast and Precise Analysis in Critical Care Patients.

Background: Ongoing or recurrent seizure activity without prominent motor features is a common burden in neurological critical care patients and people with epilepsy during ICU stays. Continuous EEG (CEEG) is the gold standard for detecting ongoing ictal EEG patterns and monitoring functional brain activity. However CEEG review is very demanding and time consuming. The purpose of the present multirater, EEG expert reviewer study, is to test and assess the clinical feasibility of an automatic EEG pattern detection method (Neurotrend). Methods: Four board certified EEG reviewers used Neurotrend to annotate 76 CEEG datasets à 6 h (in total 456 h of EEG) for rhythmic and periodic EEG patterns (RPP), unequivocal ictal EEG patterns and burst suppression. All reviewers had a predefined time limit of 5 min (± 2 min) per CEEG dataset and were compared to a predefined gold standard (conventional EEG review with unlimited time). Subanalysis of specific features of RPP was conducted as well. We used Gwet's AC1 and AC2 coefficients to calculate interrater agreement (IRA) and multirater agreement (MRA). Also, we determined individual performance measures for unequivocal ictal EEG patterns and burst suppression. Bonferroni-Holmes correction for multiple testing was applied to all statistical tests. Results: Mean review time was 3.3 min (± 1.9 min) per CEEG dataset. We found substantial IRA for unequivocal ictal EEG patterns (0.61-0.79; mean sensitivity 86.8%; mean specificity 82.2%, p < 0.001) and burst suppression (0.68-0.71; mean sensitivity 96.7%; mean specificity 76.9% p < 0.001). Two reviewers showed substantial IRA for RPP (0.68-0.72), whereas the other two showed moderate agreement (0.45-0.54), compared to the gold standard (p < 0.001). MRA showed almost perfect agreement for burst suppression (0.86) and moderate agreement for RPP (0.54) and unequivocal ictal EEG patterns (0.57). Conclusions: We demonstrated the clinical feasibility of an automatic critical care EEG pattern detection method on two levels: (1) reasonable high agreement compared to the gold standard, (2) reasonable short review times compared to previously reported EEG review times with conventional EEG analysis.

Rhythmic and periodic EEG patterns of 'ictal-interictal uncertainty' in critically ill neurological patients.

OBJECTIVES: To study periodic and rhythmic EEG patterns classified according to Standardized Critical Care EEG Terminology (SCCET) of the American Clinical Neurophysiology Society and their relationship to electrographic seizures. METHODS: We classified 655 routine EEGs in 371 consecutive critically ill neurological patients into (1) normal EEGs or EEGs with non-specific abnormalities or interictal epileptiform discharges, (2) EEGs containing unequivocal ictal EEG patterns, and (3) EEGs showing rhythmic and periodic EEG patterns of 'ictal-interictal uncertainty' (RPPIIIU) according to SCCET. RESULTS: 313 patients (84.4%) showed normal EEGs, non-specific or interictal abnormalities, 14 patients (3.8%) had EEGs with at least one electrographic seizure, and 44 patients (11.8%) at least one EEG containing RPPIIIU, but no EEG with electrographic seizures. Electrographic seizures occurred in 11 of 55 patients (20%) with RPPIIIU, but only in 3 of 316 patients (0.9%) without RPPIIIU (p⩽0.001). Conversely, we observed RPPIIIU in 11 of 14 patients (78.6%) with electrographic seizures, but only in 44 of 357 patients (12.3%) without electrographic seizures (p⩽0.001). CONCLUSIONS: On routine-EEG in critically ill neurological patients RPPIIIU occur 3 times more frequently than electrographic seizures and are highly predictive for electrographic seizures. SIGNIFICANCE: RPPIIIU can serve as an indication for continuous EEG recordings.

Monitoring burst suppression in critically ill patients: Multi-centric evaluation of a novel method.

OBJECTIVE: To develop a computational method to detect and quantify burst suppression patterns (BSP) in the EEGs of critical care patients. A multi-center validation study was performed to assess the detection performance of the method. METHODS: The fully automatic method scans the EEG for discontinuous patterns and shows detected BSP and quantitative information on a trending display in real-time. The method is designed to work without setting any patient specific parameters and to be insensitive to EEG artifacts and periodic patterns. For validation a total of 3982 h of EEG from 88 patients were analyzed from three centers. Each EEG was annotated by two reviewers to assess the detection performance and the inter-rater agreement. RESULTS: Average inter-rater agreement between pairs of reviewers was κ=0.69. On average 22% of the review segments included BSP. An average sensitivity of 90% and a specificity of 84% were measured on the consensus annotations of two reviewers. More than 95% of the periodic patterns in the EEGs were correctly suppressed. CONCLUSION: A fully automatic method to detect burst suppression patterns was assessed in a multi-center study. The method showed high sensitivity and specificity. SIGNIFICANCE: Clinically applicable burst suppression detection method validated in a large multi-center study.

Spatial correlation based artifact detection for automatic seizure detection in EEG.

Automatic EEG-processing systems such as seizure detection systems are more and more in use to cope with the large amount of data that arises from long-term EEG-monitorings. Since artifacts occur very often during the recordings and disturb the EEG-processing, it is crucial for these systems to have a good automatic artifact detection. We present a novel, computationally inexpensive automatic artifact detection system that uses the spatial distribution of the EEG-signal and the location of the electrodes to detect artifacts on electrodes. The algorithm was evaluated by including it into the automatic seizure detection system EpiScan and applying it to a very large amount of data including a large variety of EEGs and artifacts.

EpiScan: online seizure detection for epilepsy monitoring units.

An online seizure detection algorithm for long-term EEG monitoring is presented, which is based on a periodic waveform analysis detecting rhythmic EEG patterns and an adaptation module automatically adjusting the algorithm to patient-specific EEG properties. The algorithm was evaluated using 4.300 hours of unselected EEG recordings from 48 patients with temporal lobe epilepsy. For 66% of the patients the algorithm detected 100% of the seizures. A mean sensitivity of 83% was achieved. An average of 7.2 false alarms within 24 hours for unselected EEG makes the algorithm attractive for epilepsy monitoring units.