A neural network method for automatic and incremental learning applied to patient-dependent seizure detection

Dynamics reconstruction and classification via Koopman features

Knowledge discovery and information extraction of large and complex datasets has attracted great attention in wide-ranging areas from statistics and biology to medicine. Tools from machine learning, data mining, and neurocomputing have been extensively explored and utilized to accomplish such compelling data analytics tasks. However, for time-series data presenting active dynamic characteristics, many of the state-of-the-art techniques may not perform well in capturing the inherited temporal structures in these data. In this paper, integrating the Koopman operator and linear dynamical systems theory with support vector machines, we develop an ovel dynamic data mining framework to construct low-dimensional linear models that approximate the nonlinear flow of high-dimensional time-series data generated by unknown nonlinear dynamical systems. This framework then immediately enables pattern recognition, e.g., classification, of complex time-series data to distinguish their dynamic behaviors by using the trajectories generated by the reduced linear systems. Moreover, we demonstrate the applicability and efficiency of this framework through the problems of time-series classification in bioinformatics and healthcare, including cognitive classification and seizure detection with fMRI and EEG data, respectively. The developed Koopman dynamic learning framework then lays a solid foundation for effective dynamic data mining and promises a mathematically justified method for extracting the dynamics and significant temporal structures of nonlinear dynamical systems.

Nonconvulsive epileptic seizure monitoring with incremental learning

Nonconvulsive epileptic seizures (NCSz) and nonconvulsive status epilepticus (NCSE) are two neurological entities associated with increment in morbidity and mortality in critically ill patients. In a previous work, we introduced a method which accurately detected NCSz in EEG data (referred here as 'Batch method'). However, this approach was less effective when the EEG features identified at the beginning of the recording changed over time. Such pattern drift is an issue that causes failures of automated seizure detection methods. This paper presents a support vector machine (SVM)-based incremental learning method for NCSz detection that for the first time addresses the seizure evolution in EEG records from patients with epileptic disorders and from ICU having NCSz. To implement the incremental learning SVM, three methodologies are tested. These approaches differ in the way they reduce the set of potentially available support vectors that are used to build the decision function of the classifier. To evaluate the suitability of the three incremental learning approaches proposed here for NCSz detection, first, a comparative study between the three methods is performed. Secondly, the incremental learning approach with the best performance is compared with the Batch method and three other batch methods from the literature. From this comparison, the incremental learning method based on maximum relevance minimum redundancy (MRMR_IL) obtained the best results. MRMR_IL method proved to be an effective tool for NCSz detection in a real-time setting, achieving sensitivity and accuracy values above 99%.

Assessing the Feasibility of Providing a Real-Time Response to Seizures Detected With Continuous Long-Term Neonatal Electroencephalography Monitoring

PURPOSE

Continuous video electroencephalography (cEEG) monitoring is the recommended gold standard of care for at-risk neonates but is not available in many Neonatal Intensive Care Units (NICUs). To conduct a randomized treatment trial of levetiracetam for the first-line treatment of neonatal seizures (the NEOLEV2 trial), we developed a monitoring infrastructure at five NICUs, implementing recent technological advancements to provide continuous video EEG monitoring and real-time response to seizure detection. Here, we report on the feasibility of providing this level of care.

METHODS

Twenty-five key informant interviews were conducted with study neurologists, neonatologists, coordinators, and EEG technicians from the commercial EEG monitoring company Corticare. A general inductive approach was used to analyze these qualitative data.

RESULTS

A robust infrastructure for continuous video EEG monitoring, remote review, and real-time seizure detection was established at all sites. At the time of this survey, 260 babies had been recruited and monitored for 2 to 6 days. The EEG technician review by the commercial EEG monitoring company was reassuring to families and neonatologists and led to earlier detection of seizures but did not reduce work load for neurologists. Neurologists found the automated neonatal seizure detector algorithm provided by the EEG software company Persyst useful, but the accuracy of the algorithm was not such that it could be used without review by human expert. Placement of EEG electrodes to initiate monitoring, especially after hours, remains problematic.

CONCLUSIONS

Technological advancements have made it possible to provide at-risk neonates with continuous video EEG monitoring, real-time detection of and response to seizures. However, this standard of care remains unfeasible in usual clinical practice. Chief obstacles remain starting a recording and resourcing the real-time specialist review of suspect seizures.

Automated seizure detection accuracy for ambulatory EEG recordings

Objective

To investigate the accuracy of preselected software automatic seizure files to detect at least one seizure per study in prolonged ambulatory EEG recording.

Methods

All the prolonged ambulatory EEG recordings (>24 hours) read at the Northwestern Memorial Hospital from January 2013 to October 2017 were included. We selected only the first study of each patient. We reviewed the studies entirely, and processed the recordings through 1 of 3 different detection software that are commercially available (Persyst 11, Persyst 12, and Gotman TM Event Detection). The proportion of patients with at least one electrographic seizure (≥10 seconds) correctly identified by a seizure detector was calculated. Finally, we evaluated whether the type of seizure (focal vs generalized) may affect the chances of being automatically detected.

Results

We read 1,478 ambulatory EEG studies entirely (2,323 days of EEG recording; average 1.6 d/study). From the first study of each patient (1,257 studies), we found electrographic seizures in 70 (5.6%) studies. In 37 of 70 patients (53%), the automatic detectors correctly identified at least one seizure. Detections happened slightly more frequently in generalized seizures (14/20, 70%) compared to focal seizures (23/50, 46%) (p= 0.06).

Conclusion

Seizures were found in 5.6% of the studies. Automatic seizure detectors identified at least one electrographic seizure in only 53% of the studies. They performed slightly better detecting generalized than focal seizures. Therefore, the review of only automatically selected segments may be of decreased value to identify seizures, in particular when focal seizures are suspected.

Normative amplitude-integrated EEG measures in preterm infants

OBJECTIVE

Assessing qualitative patterns of amplitude-integrated electroencephalography (aEEG) maturation of preterm infants requires personnel with training in interpretation and an investment of time. Quantitative algorithms provide a method for rapidly and reproducibly assessing an aEEG recording independent of provider skill level. Although there are several qualitative and quantitative normative data sets in the literature, this study provides the broadest array of quantitative aEEG measures in a carefully selected and followed cohort of preterm infants with mild or no visible injury on term-equivalent magnetic resonance imaging (MRI) and subsequently normal neurodevelopment at 2 and 7 years of age.

STUDY DESIGN

A two-channel aEEG recording was obtained on days 4, 7, 14 and 28 of life for infants born ⩽30 weeks estimated gestational age. Measures of amplitude and continuity, spectral edge frequency, percentage of trace in interburst interval (IBI), IBI length and frequency counts of smooth delta waves, delta brushes and theta bursts were obtained. MRI was obtained at term-equivalent age and neurodevelopmental testing was conducted at 2 and 7 years of corrected age.

RESULT

Correlations were found between increasing postmenstrual age (PMA) and decreasing maximum amplitude (R= -0.23, P=0.05), increasing minimum amplitude (R=0.46, P=0.002) and increasing spectral edge frequency (R=0.78, P=4.17 × 10(-14)). Negative correlations were noted between increasing PMA and counts of smooth delta waves (R= -0.39, P=0.001), delta brushes (R= -0.37, P=0.003) and theta bursts (R= -0.61, P=5.66 × 10(-8)). Increasing PMA was also associated with a decreased amount of time spent in the IBI (R= -0.38, P=0.001) and a shorter length of the maximum IBI (R= -0.27, P=0.03).

CONCLUSION

This analysis supports a strong correlation between quantitatively determined aEEG measures and PMA, in a cohort of preterm infants with normal term-equivalent age neuroimaging and neurodevelopmental outcomes at 7 years of age, which is both predictable and reproducible. These 'normative' quantitative values support the pattern of maturation previously identified by qualitative analysis.

Seizure detection method based on fractal dimension and gradient boosting

Automatic seizure detection technology is necessary and crucial for the long-term electroencephalography (EEG) monitoring of patients with epilepsy. This article presents a patient-specific method for the detection of epileptic seizures. The fractal dimensions of preprocessed multichannel EEG were firstly estimated using a k-nearest neighbor algorithm. Then, the feature vector constructed for each epoch was fed into a trained gradient boosting classifier. After a series of postprocessing, including smoothing, threshold processing, collar operation, and union of seizure detections in a short time interval, a binary decision was made to determine whether the epoch belonged to seizure status or not. Both the epoch-based and event-based assessments were used for the performance evaluation of this method on the EEG data of 21 patients from the Freiburg dataset. An average epoch-based sensitivity of 91.01% and a specificity of 95.77% were achieved. For the event-based assessment, this method obtained an average sensitivity of 94.05%, with a false detection rate of 0.27/h.

Seizure detection, seizure prediction, and closed-loop warning systems in epilepsy

Nearly one-third of patients with epilepsy continue to have seizures despite optimal medication management. Systems employed to detect seizures may have the potential to improve outcomes in these patients by allowing more tailored therapies and might, additionally, have a role in accident and SUDEP prevention. Automated seizure detection and prediction require algorithms which employ feature computation and subsequent classification. Over the last few decades, methods have been developed to detect seizures utilizing scalp and intracranial EEG, electrocardiography, accelerometry and motion sensors, electrodermal activity, and audio/video captures. To date, it is unclear which combination of detection technologies yields the best results, and approaches may ultimately need to be individualized. This review presents an overview of seizure detection and related prediction methods and discusses their potential uses in closed-loop warning systems in epilepsy.

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.

Automating the analysis of EEG recordings from prematurely-born infants: a Bayesian approach

OBJECTIVE

To implement an automated analysis of EEG recordings from prematurely-born infants and thus provide objective, reproducible results.

METHODS

Bayesian probability theory is employed to compute the posterior probability for developmental features of interest in EEG recordings. Currently, these features include smooth delta waves (0.5-1.5Hz, >100μV), delta brushes (delta portion: 0.5-1.5Hz, >100μV; "brush" portion: 8-22Hz, <75μV), and interburst intervals (<10μV), though the approach taken can be generalized to identify other EEG features of interest.

RESULTS

When compared with experienced electroencephalographers, the algorithm had a true positive rate between 72% and 79% for the identification of delta waves (smooth or "brush") and interburst intervals, which is comparable to the inter-rater reliability. When distinguishing between smooth delta waves and delta brushes, the algorithm's true positive rate was between 53% and 88%, which is slightly less than the inter-rater reliability.

CONCLUSION

Bayesian probability theory can be employed to consistently identify features of EEG recordings from premature infants.

SIGNIFICANCE

The identification of features in EEG recordings provides a first step towards the automated analysis of EEG recordings from premature infants.

Optimal control-based bayesian detection of clinical and behavioral state transitions

Accurately detecting hidden clinical or behavioral states from sequential measurements is an emerging topic in neuroscience and medicine, which may dramatically impact neural prosthetics, brain-computer interface and drug delivery. For example, early detection of an epileptic seizure from sequential electroencephalographic (EEG) measurements would allow timely administration of anticonvulsant drugs or neurostimulation, thus reducing physical impairment and risks of overtreatment. We develop a Bayesian paradigm for state transition detection that combines optimal control and Markov processes. We define a hidden Markov model of the state evolution and develop a detection policy that minimizes a loss function of both probability of false positives and accuracy (i.e., lag between estimated and actual transition time). Our strategy automatically adapts to each newly acquired measurement based on the state evolution model and the relative loss for false positives and accuracy, thus resulting in a time varying threshold policy. The paradigm was used in two applications: 1) detection of movement onset (behavioral state) from subthalamic single unit recordings in Parkinson's disease patients performing a motor task; 2) early detection of an approaching seizure (clinical state) from multichannel intracranial EEG recordings in rodents treated with pentylenetetrazol chemoconvulsant. Our paradigm performs significantly better than chance and improves over widely used detection algorithms.

Model-based seizure detection for intracranial EEG recordings

This paper presents a novel model-based patient-specific method for automatic detection of seizures in the intracranial EEG recordings. The proposed method overcomes the complexities in the practical implementation of the patient-specific approach of seizure detection. The method builds a seizure model (set of basis functions) for a priori known seizure (the template seizure pattern), and uses the statistically optimal null filters as a building block for the detection of similar seizures. The process of modeling the template seizure is fully automatic. Overall, the detection method involves the segmentation of the template seizure pattern, rejection of the redundant and noisy segments, extraction of features from the segments to generate a set of models, selection of the best seizure model, and training of the classifier. The trained classifier is used to detect similar seizures in the remaining data. The resulting seizure detection method was evaluated on a total of 304 h of single-channel depth EEG recordings from 14 patients. The system performance is further compared to the Qu-Gotman patient-specific system using the same data. A significant improvement in the proposed system, in terms of specificity, is observed over the compared method.

Quickest detection of drug-resistant seizures: an optimal control approach

Epilepsy affects 50 million people worldwide, and seizures in 30% of the cases remain drug resistant. This has increased interest in responsive neurostimulation, which is most effective when administered during seizure onset. We propose a novel framework for seizure onset detection that involves (i) constructing statistics from multichannel intracranial EEG (iEEG) to distinguish nonictal versus ictal states; (ii) modeling the dynamics of these statistics in each state and the state transitions; you can remove this word if there is no room. (iii) developing an optimal control-based "quickest detection" (QD) strategy to estimate the transition times from nonictal to ictal states from sequential iEEG measurements. The QD strategy minimizes a cost function of detection delay and false positive probability. The solution is a threshold that non-monotonically decreases over time and avoids responding to rare events that normally trigger false positives. We applied QD to four drug resistant epileptic patients (168 hour continuous recordings, 26-44 electrodes, 33 seizures) and achieved 100% sensitivity with low false positive rates (0.16 false positive/hour). This article is part of a Supplemental Special Issue entitled The Future of Automated Seizure Detection and Prediction.

SIGNAL REGULARITY-BASED AUTOMATED SEIZURE DETECTION SYSTEM FOR SCALP EEG MONITORING

The purpose of the present study was to build a clinically useful automated seizure detection system for scalp EEG recordings. To achieve this, a computer algorithm was designed to translate complex multichannel scalp EEG signals into several dynamical descriptors, followed by the investigations of their spatiotemporal properties that relate to the ictal (seizure) EEG patterns as well as to normal physiologic and artifact signals. This paper describes in detail this novel seizure detection algorithm and reports its performance in a large clinical dataset.

Patient-specific early seizure detection from scalp electroencephalogram

The objective of this study is to develop a method for automatic detection of seizures before or immediately after clinical onset using features derived from scalp electroencephalogram. This detection method is patient specific. It uses recurrent neural networks and a variety of input features. For each patient, we trained and optimized the detection algorithm for two cases: (1) during the period immediately preceding seizure onset and (2) during the period immediately after seizure onset. Continuous scalp electroencephalogram recordings (duration 15-62 hours, median 25 hours) from 25 patients, including a total of 86 seizures, were used in this study. Preonset detection was successful in 14 of the 25 patients. For these 14 patients, all of the testing seizures were detected before seizure onset with a median preonset time of 51 seconds and false-positive (FP) rate was 0.06/hour. Postonset detection had 100% sensitivity, 0.023/hour FP rate, and median delay of 4 seconds after onset. The unique results of this study relate to preonset detection. Our results suggest that reliable preonset seizure detection may be achievable for a significant subset of patients with epilepsy without use of invasive electrodes.

Comparison of seizure detection algorithms in continuously monitored pediatric patients

SUMMARY

Robust, automated seizure detection has long been an important goal in epilepsy research because of both the possibilities for portable intervention devices and the potential to provide prompter, more efficient treatment while in clinic. The authors present results on how well four seizure detection algorithms (based on principal eigenvalue [EI], total power, Kolmogorov entropy [KE], and correlation dimension) discriminated between ictal and interictal EEG and electrocorticoencephalography (ECoG) from four patients (aged 13 months to 21 years). Test data consisted of 46 to 78 hours of continuously acquired EEG/ECoG for each patient (245 hours total), and the detectors' accuracy was checked against seizures found by a board-certified neurologist and an experienced registered EEG technician. The results were patient-specific: no algorithm performed well on a 13-month-old patient, and no algorithm consistently performed best on the other three patients. One of the metrics (EI) supported the existence of a postictal period of 5 to 15 minutes in the three oldest patients, but no strong evidence of a preictal anticipation was found. Two metrics (EI and KE) cycled continuously with a period of several hours in a 21-year-old patient, highlighting the importance of continuous analysis to differentiate background cycling from anticipation.

Algorithm architectures for patient dependent seizure detection

OBJECTIVE

The goal of this work is to determine whether improved performance (compared to patient independent algorithms) can be achieved by an algorithm, developed on the fly, that requires no user input beyond the identification of the first one or two seizures in the record.

METHODS

The previously developed AutoLearn algorithm, which employs the probabilistic neural network (PNN), is tested on 209 seizures obtained from the epilepsy monitoring unit (EMU) or ambulatory recordings. A construction algorithm is used to compare a variety of algorithm architectures and factors. The Taguchi design of experiments (DoE) method is employed find the significant factors without resorting to a full factorial design.

RESULTS

Architectures that train a single PNN per channel and use segmentation to identify ranges of similar activity are preferred. The two best architectures are insensitive to the levels of any of the other factors tested. The training time for the algorithm is less than 1s, and approximately 2 min are required to find the seizures in an 8 h record.

CONCLUSIONS

The final algorithm, which requires no input from the user other than the marking of the first seizure in a record, performs as well or better than the 3 seizure detectors on EMU and ambulatory records. The algorithm performs nearly as well as human experts on the EMU records.

SIGNIFICANCE

The described method can be used to identify unusual seizures (or other patterns) that will be missed by the current generation of seizure detectors. We expect that the methods developed here will also aid the development of patient independent seizure detectors that can improve their performance over time by incorporating new examples.