Seizure Prediction and Closed-Loop Intervention

Anticipation of epileptic seizures is, among others, the most challenging aspect in epileptology. Recent findings indicate that particularly synchronization measures show a promising performance that exceeds chance level if tested by statistical validation. Estimators for generalized synchronization offer the possibility for detecting driver-responder relationships and are thus highly attractive to identify interacting brain structures involved in ictogenesis. Despite the conceptual simplicity of nonlinear interdependency measures, real-time applications are currently limited by calculations for the large number of electrode combinations. Promising systems for measuring generalized synchronization while minimizing space and energy consumption are cellular nonlinear networks (CNN) as they offer a massive computing power, are capable of universal computation, and are already available as analogue integrated circuits. Studying coupled model system (structurally identical and non-identical) we show the ability of our software CNN to approximate both symmetrical and asymmetrical interdependency measures with a sufficient accuracy (~ 90 %; out-of-sample validation). A subsequent analysis of multi-day, multi-channel intracranial EEG recordings from up to now two patients undergoing presurgical evaluation shows that a long-lasting pre-ictal state (duration: ~ 4 hours) can be detected using our CNN-based approach to measure generalized synchronization. At present, the computation times for a CNN-based estimation and a numerical calculation of interdependency measures are comparable. Hardware implementations will provide a speed-up of orders of magnitudes making complex real-time applications possible. This work was supported by the Deutsche Forschungsgemeinschaft.

We propose a data-driven approach to measure coupling between stochastic processes. Extending a method proposed by Siegert et al. (Phys. Lett. A 243, 275, 1998) we estimate deterministic and stochastic parts of the dynamics and define measures that both detect the direction and quantify the strength of the coupling. We numerically test the method using time series from coupled stochastic model systems. Applying our method to intracranially recorded EEG time series from patients suffering from focal epilepsies, we observe interdependencies between EEG signals from different contacts that allow one to identify the epileptic focus even during the interictal state. We discuss possible applications for the detection of pre-seizure states.

Introduction: Seizure prediction algorithms are often computationally complex and require the use of large multi-patient intracranial EEG data sets. Measures employed to avoid in-sample testing necessitate partitioning data and separating training and testing computations. Rigorous methods are required to characterize algorithm performance. The computational burden of these tasks is a barrier to developing and evaluating seizure prediction algorithms. Methods: A cluster computing system was designed specifically for seizure prediction research. Storage requirements were derived from the desire to study a meaningful population of 50 to 100 patients, with a typical record length of 100 to 200 gigabytes per subject. Computational requirements were derived from the need to execute and assess the results from hundreds of candidate algorithms, comprising signal processing, feature extraction, classification, cross-validation, and performance metric calculation tasks. The resulting system includes 120 computing nodes comprising 675 gigaflops of processing power and 40 terabytes of on-line storage. Custom software supports automated preparation and management of large quantities of clinical data, rapid implementation and execution of algorithms, rigorous performance analysis, and management of cluster hardware. Results: Over an 18 month period, the cluster computing system was used to analyze approximately 500 candidate algorithms and variants, involving the processing of approximately 5 million patient-hours from a data set of 76 patients. Conclusions: Rapid, rigorous, evaluation of candidate seizure prediction algorithms using large data sets and cross-validation methods is beyond the capability of presently available individual personal computers, but may be accomplished using cluster computing techniques.

Introduction: Assembling a database for developing and evaluating seizure prediction algorithms requires data from multiple centers to be converted to a common format and validated. Differences in data collection practices and equipment must be addressed. Methods: In the process of collecting and pre-processing patient data, inconsistencies were identified, catalogued, and corrected using a flexible pre-processing system. Results: Issues that required correction were: Lead changes within subject – due to headbox limitations, investigators sometimes exchanged leads for a given channel within the same recording session. Dropped samples – some EEG manufacturers allow system pausing or account for time drift by assigning clock times to sample times, creating discontinuities within the binary data. Labeling inconsistencies – physicians and technicians frequently used different channel names within subjects, complicating electrode identification. Non-integer sampling frequencies – some EEG systems account for clock/sample drift by assigning fractional sampling frequencies, making resampling difficult. Time zone inconsistencies – it is not uncommon for data within differing portions of a manufacturer’s file set (EEG, video, annotations, patient demographics) to employ differing time standards, potentially causing desynchronization. Anonymization – patient information is embedded across many data file types, complicating deidentification efforts. Calibration factor changes within session – investigators infrequently alter their headbox gain within sessions, requiring recalibration of binary data to provide a uniform amplitude scale. Conclusions: Significant data pre-conditioning is required to facilitate use in algorithm development. The aforementioned system supports mass processing, while allowing a dynamically growing library of corrective measures to be applied to subject files.

Introduction: Seizure prediction studies have focused primarily on patients with well-localized seizures, and electrodes proximate to the ictal onset zone. Whether seizure prediction is more challenging for patients with extra-temporal and/or regional seizure onset remains unexplored. Methods: Continuous subdural EEG recordings of 76 EMU patients were analyzed with a seizure prediction algorithm. The dataset comprises 7452 hours of data, with 301 primary seizures (clusters excluded). Of these patients, seizure onsets were regional neocortical in 7, unilateral mesial temporal in 18, and multi-focal and/or extra-temporal in the remainder. A block-wise (90 minute) statistical model employing k-fold cross-validation was used to compute prediction sensitivity and the false positive rate for each patient. The Wilcoxon signed ranks test was used to examine algorithm performance versus a chance predictor (paired rate of alerts), and the Kruskall-Wallis test for differences by seizure onset pattern. Results: Mean sensitivity was 79% (p < 0.0001 vs. chance), with a false positive rate of 0.08/hr (2.2/seizure). Sensitivity did not differ by onset pattern (p = 0.77) with 80% for mesial temporal, 88% for regional, and 77% for all others. False positive rate was similarly unaffected (p = 0.72) with rates of 0.08/hr for mesial temporal, 0.07/hr for regional, and 0.08/hr for all others (2.6, 2.4, 2.0/seizure, respectively). Conclusions: High sensitivity seizure prediction with clinically acceptable false alert rates has been demonstrated in a population of 76 EMU patients. Performance for patients with mesial temporal or regional neocortical onsets did not differ from the population as a whole.

In a clinical study currently underway at the Beth Israel Deaconness Medical Center in Boston, Massachusetts we are demonstrating the feasibility of automatically triggering the Vagus Nerve Stimulator Therapy System (Cyberonics Inc.) in response to real-time detection of a neurologist-specified (ictal or interictal) abnormality in the scalp EEG. We recently enrolled our first patient, and confirmed that our computerized system reliably and automatically activated the subject's vagus nerve stimulator whenever the patient experienced an inter-ictal epileptiform burst that was pre-specified by a neurologist. Future results form the study will be presented at the meeting. Our computerized system relies on a patient-specific detection algorithm to differentiate the neurologist-specified (ictal or interictal) epileptiform abnormality from the remaining activity in a subject’s EEG. The patient-specific algorithm classifies a subject’s EEG as being consistent or inconsistent with the neurologist-specified abnormality using the support-vector machine learning algorithm and feature vectors derived from spectral analysis of the EEG. Prior to our clinical study of closed-loop, EEG-triggered vagus nerve stimulation, we retrospectively evaluated the sensitivity, specificity, and detection latency of our patient-specific algorithm using 536 hours of continuously recorded scalp EEG from 16 pediatric epilepsy subjects. We noted that our patient-specific algorithm detected 132/143 seizures with a 6.8+/-2.4 second detection latency and 0.2+/-0.7 false alarms per hour. For comparison, the Reveal algorithm (a commercial, patient non-specific detector) evaluated on the same data detected 94/143 seizures with a 17.8+/-10.0 second detection latency and 2.0+/-5.3 false alarms per hour.

The derivation and analysis of EEG signal feature extraction methods has been treated in several investigations. Although, the results of recent studies indicate that a pre-ictal transition to a seizure can be detected, up to now the sensitivity and specificity of the applied methods are not sufficient for an automated seizure prediction. In this contribution we will present recent results obtained by applying a signal prediction algorithm to segmented EEG-signals of long-term recordings obtained in ECoG and SEEG measurements. The prediction performance of autonomous single-layer delay-type discrete time CNN (DTCNN) has been analysed for different cell coupling structures and several feedback delays. Firstly, nonlinear feedbacks have been assumed which have been represented by polynomial weight functions of different order. By taking the assumption that for each EEG data segment a certain DTCNN predictor can be applied, a series of different prediction coefficients and errors results in an EEG analysis. We observed distinct changes of these quantities in our investigations by considering data of different patients. A detailed discussion of all results will be given in the paper.

The time-varying dynamics of epileptic seizures and the high inter-individual variability make their detection difficult. EEG (electroencephalograph) has been an important clinical tool for the evaluation and treatment of epilepsy. Epileptic seizures occur when a massive group of neurons in the cerebral cortex suddenly begin to discharge in a highly organized rhythmic pattern and exotic phenomena happen during and epileptic seizure which make the long-term prediction of seizures very difficult. On the other hand, neural networks and neuro-fuzzy models are well-known mathematical methods for describing the complex and chaotic forms in nature. In this study, a combination of singular spectrum analysis and the neuro-fuzzy interpretation of locally linear models is proposed to make accurate long term prediction of EEG signals for alarming epileptic seizures. The principal components obtained from spectral analysis have narrow band frequency spectra and definite linear or nonlinear trends and periodic patterns, hence they are predictable in large prediction horizon. The incremental learning algorithm initiates a model for each of the components as an optimal linear least squares estimation, and adds the nonlinear neurons if they help to reduce error indices over training and validation sets. Therefore the algorithm automatically constructs the best linear or nonlinear model for each of the principal components to achieve maximum generalization, and the long term prediction of the original time series is obtained by recombining the predicted components. Results depict the power of the proposed method in long term prediction EEG signals during epileptic seizures.

Download the poster as PDF file

Recent approaches to epileptic seizure prediction have utilised nonlinear analysis tools, with many methods such as Dynamical similarity index [1] based on delay reconstruction. Although delay reconstruction is a powerful tool, the EEG is ultimately an unsuitable signal for analysis within its current framework. Delay reconstruction theoretically applies to noiseless time-series data, recorded from an autonomous and low-dimensional system. However, the EEG is a noisy signal, recorded from a non-autonomous system – our brains are both non-stationary and input driven. Furthermore, there is little evidence that the brain is low-dimensional. This work investigates if a signal more suited to delay reconstruction can be generated from EEG data through some adaptions to the DSI method. Firstly, we analyse a time-series that is the difference between two closely spaced intracranial recordings. This has the advantage that both the common electrode noise (e.g. 50Hz) and the effect of far away dynamics are reduced. This attempts to create a time-series representation of the underlying local system by cancelling the common input from far away brain dynamics. Secondly, given the hypothesis that the brain is lower dimensional during the pre-ictal period [2] we propose that the template of reference dynamics used is formed from the pre-ictal (rather than inter-ictal) period. Preliminary findings on the Freiburg competition dataset show sensitivity of 25%-100% across patients, with high false positive rates (FPR) of 1-6.6 FP/hr. Further analysis is required to investigate if the FPR can be reduced and to verify if this can offer any improvement over traditional DSI. The high FPR of preliminary results suggests that perhaps future epilepsy prediction research should concentrate on non-reconstruction based methods. [1] K. Lehnertz and C.E. Elger. Can epileptic seizures be predicted? Evidence from nonlinear time series analysis of brain electrical activity. Physical Review Letters, 80(22):5019–5022, 1998. [2] M. Le Van Quyen, J. Martinerie, M. Baulac, and F. Varela. Anticipating epileptic seizures in real time by a non-linear analysis of similarity between EEG recordings. NeuroReport, 10(10):2149–2155, 1999

Download the poster as PDF file

Purpose: High frequency oscillations (HFOs) have recently proven to be useful in examining epileptic transients. It is thought that these HFOs are specific to epileptic pathologies and may aid diagnosis in terms of localization and provide further understanding of underlying mechanisms. Here we investigate the role of HFOs in terms of seizure prediction. Method: Intracranial EEG was recorded from several patients who were undergoing pre-surgical assessment from subdural ECoG grid electrodes or depth electrodes. The EEG was sampled at 4kHz. Spectral content was examined in the time-frequency domain during interictal, preictal and ictal time periods. Results: Results from the data analysis demonstrated high frequency spectral power changes in relation to epileptic events. These spectral power changes were compared to synchronization/desynchronization event related data that were recorded from non-epileptogenic tissue, that also demonstrate significant HFO components above 250 Hz. Conclusion: HFOs are a possible common factor within a diverse range of epileptic etiologies. If this is confirmed by future work, this may provide a method of non-patient-specific seizure anticipation within a desirable time range for focal therapeutic intervention.

Seizure prediction is a very challenging topic in EEG processing. Same of these difficulties are caused by the fact, that the results of different studies can not be easily compared or reproduced. Partly, there are technological reasons and a lack of standardization. Different data formats, sampling rates and filter settings make it difficult to compare the data. Often, only univariate or bivariate signal processing methods are used, which limit the number of parameters to one or two. Furthermore, no single unambiguous performance measure is available, and the various performance metrics can not be easily compared. BioSig is an open source software library for biomedical signal processing. BioSig supports the reading of over 40 different biomedical signal data formats; BioSig contains various methods for quality control and artifact processing [1], many feature extraction and classification methods are supported, and long list of various evaluation criteria are supported. BioSig is free/libre open source software (FLOSS), licensed with the GNU GPL. BioSig provides a common framework, that supports the comparison of different methods and approaches. Parts of BioSig were developed for the Brain-Computer interface research which requires single trial EEG classification. Many of these tools could be useful for developing and comparing methods on seizure prediction and detection, for investigating the interictal-ictal transition, and for modeling the dynamics of epileptic processes. Recently support for the database of the epileptics seizure prediction contest was added and we are investigating how BioSig can contribute best to the challenge of epileptic seizure prediction.

Download the poster as PDF file

PURPOSE: Automated detection of high-frequency oscillations (HFOs) in intracranial recordings of human epileptic EEG has become an interesting research topic. Those HFOs have been shown to occur localized in space at the seizure onset zone as well as localized in time before an upcoming seizure. METHODS: We analyzed recordings of 2 patients suffering from pharmacoresistant focal epilepsy. They underwent implantation of subdural grid and depth electrodes for presurgical long-term EEG monitoring with a sampling rate of 1024Hz. For both patients we selected 3 intrafocal, 3 extrafocal, and a reference channel. The EEG signals have been bandpass filtered by a dyadic filterbank (wavelet analysis), which also replaces preprocessing like spectral equalization or preemphasis. The frequency bands of interest are 32-64, 64-128, 128-256, and 256-512Hz. For each band, short-time energy and short-time line-length have been computed and compared to a threshold. Also, a duration threshold has been implemented. The data from start of recording to the first seizure have been used as a training set to adjust the threshold for each band. Short HFOs (max. 80ms) which range over several bands are classified as spikes and will be excluded from further analysis. RESULTS: In the poster we show results regarding spatial and temporal location of the detected HFOs for each band and we hypothesize their potential usefulness for seizure prediction. CONCLUSIONS: A novel multiband method for HFO detection, which is able to find HFOs of specified minimal duration, has been developed. The benefit is that HF components belonging to spikes can be detected and treated separately.

Download the poster as PDF file