Patient-specific bivariate-synchrony-based seizure prediction for short prediction horizons
Introduction
Epileptic seizures are manifested by hyper-synchronous rapid neural activity (McCormick and Contreras, 2001, Avanzini and Franceschetti, 2003). This synchrony occurs primarily between the membrane potentials of neurons or their spike trains, but depending on the brain networks within which the seizure occurs, this synchrony at the neuronal scale can give rise to emergent field potentials that can also synchronize across different brain regions. Given the difficulty of experimentally analysing seizure-related synchrony at the neuronal scale in humans, intracranial electroencephalography (iEEG), which involves course sampling of mesoscopic field potentials across the brain, has been the tool of choice for analysing seizure-related synchrony (Arnhold et al., 1999, Lai et al., 2007, Kiss et al., 2008, Sabesan et al., 2009).
A primary application for the analysis of seizure-related synchrony has been epileptic seizure prediction (Litt and Lehnertz, 2002, Chávez et al., 2003, Lehnertz et al., 2003, Mormann et al., 2003a, Mormann et al., 2003b, Jouny et al., 2005, Mormann et al., 2005, Schelter et al., 2006b, Osterhage et al., 2007, Schad et al., 2008, Mirowski et al., 2009). It is thought that as brain activity moves along a trajectory towards a synchronized seizure state, there will be noticeable changes in synchrony in iEEG recordings that can be used to predict the seizure state. Automated seizure prediction algorithms can be useful for giving a patient a warning before an oncoming seizure, or for activating a deep brain stimulator to prevent or abort seizures – see Litt and Lehnertz (2002), Lehnertz et al. (2003), Mormann et al. (2007) and Hughes (2008) for reviews.
Mormann et al., 2003a, Mormann et al., 2003b, Mormann et al., 2005 were one of the first groups to analyse the bivariate-synchrony of iEEG signals for the purposes of seizure prediction. Mormann et al. (2003b) performed an analysis of synchrony on data taken from 10 temporal lobe epilepsy patients by tracking mean phase coherence (MPC) and maximum linear cross-correlation between iEEG channel pairs. For both measures of synchrony, they observed pre-ictal decreases in synchrony for 12 out of 14 seizures. Mormann et al. (2003a) also performed a similar analysis on data from 18 focal epilepsy patients and they observed decreases in synchrony prior to 26 out of 32 seizures. Mormann et al. (2005) investigated the predictability of seizures by comparing pre-ictal and interictal distributions for 5 patients. For relevant phase synchrony measures it was found that statistically significant prediction performance could be obtained with prediction horizons of 240 and 5 min for a constant baseline and dynamic baseline, respectively. Chávez et al. (2003) analysed synchrony-based seizure prediction using phase synchrony and non-linear regression analysis in 2 patients with focal epilepsy. For the 10–25 Hz frequency band it was observed that decreases in synchrony occurred within 30 min before seizures in both patients. Jouny et al. (2005) observed no pre-ictal changes when they tracked a univariate autoregressive measure of synchrony in 2 patients. Schelter et al. (2006b) also used MPC for the purposes of prediction and obtained sensitivities (i.e. the proportion of seizures correctly predicted) in the range of 0.4–1 for a maximum false prediction rate (FPR – the number of false predictions per hour) of 0.15 h−1. However, performance was only better than a random predictor for 2 out of the 4 patients and for certain prediction thresholds. Similar conditional results were obtained by Winterhalder et al. (2006) and Schelter et al. (2007). Using a synchrony measure based on multivariate coincidence detection of iEEG signals, Schad et al. (2008) obtained sensitivity of 0.5 with a maximum FPR of 0.15 h−1 for 26 seizures recorded from 6 patients. This performance was better than that of a random predictor for certain thresholds. Mirowski et al. (2009) combined non-linear classifiers with phase-synchrony measures for 21 patients with dis-continuous data and found perfect prediction performance for 71% of the patients.
This paper addresses five major aspects of proper evaluation of bivariate-synchrony-based seizure prediction applied to individuals (Mormann et al., 2005) by (1) analysing synchrony between all iEEG channel pairs to find the channel pairs that provide the best synchrony-based seizure prediction performance for a given patient; (2) performing the analysis on long-term continuous iEEG data, instead of discontinuous chunks of data; (3) analysing the different times over which pre-ictal changes in synchrony could take place; (4) determining whether or not increases, as opposed to decreases, in synchrony are also relevant to seizure prediction; and (5) comparing the performance of a synchrony-based predictor with a random predictor. The following paragraphs summarise each of these aspects.
Chávez et al. (2003) observed focal decreases in synchrony preceding seizures in 2 patients. This focal decrease in synchrony is to some degree contrary to the results of Mormann et al., 2003a, Mormann et al., 2003b who observed that pre-ictal desynchronization was not necessarily confined to the focus, but could instead be observed in more distant, even contralateral areas of the brain. These seemingly conflicting results highlight one of the problems of bivariate-synchrony analysis for seizure prediction; namely, it is difficult to analyse the synchrony between all possible pairs of iEEG channels without using a significant amount of computer power and time. As a result, many studies choose to analyse a subset of channels to look for pre-ictal changes in synchrony. Given that the brain is highly complex and everyone's brain is different, for a given individual it is not clear which brain regions are connected and which are not. Hence it is hard to know which iEEG channel pairs to use to compute the synchrony in order to detect pre-ictal changes. The work presented in this paper approaches this problem by doing ‘brute force’ synchrony analysis of all pairs of iEEG recording channels. While this analysis cannot be done in real-time or on-line, it can be performed off-line to select which channels perform best at synchrony-based seizure prediction. Then these best performing channels could be analysed for the purposes of real time seizure prediction.
Another drawback of previous studies is that many involved discontinuous data sets. For example, in Mormann et al. (2003b), the data for each patient included waking state segments of at least one seizure with a minimum of 10 min recording time before seizure onset and at least one interictal recording of at least 15 min. Using discontinuous data does not provide a good idea of the true performance of the seizure prediction method, since the brain is in a non-seizure state most of the time, and the FPR or specificity of the predictor needs to be evaluated over long, preferably continuous EEG segments to get a better feel for the quality of the predictor. The work presented in this paper presents an analysis of synchrony-based seizure prediction applied to long-term continuous recordings greater than 30 h per patient. Mormann et al. (2005) investigated bivariate-synchrony-based prediction for 311 h of continuous data from 5 patients. However, their analysis was retrospective and allowed for longer prediction horizons than are investigated in this study.
Mormann et al. (2003a) state that pre-ictal decreases in synchrony occur over the scale of minutes to several hours. This indicates that progress towards a seizure is different for different patients, especially in terms of synchrony of the iEEG. As a result, a synchrony-based seizure predictor should be evaluated on an individual basis and be flexible with respect to the variability in prediction times observed across patients. This paper presents a pseudo-prospective offline evaluation of synchrony using a variety of seizure prediction horizons preceding a seizure, which determine whether or not a prediction is true or false. The seizure prediction horizon giving the best performance for that patient would then be used for real time seizure prediction. While long prediction horizons may be informative, this paper focuses on prediction horizons under 15 min because we are primarily interested in seizure prediction for an implantable seizure control device and under 15 min should be adequate time for drug delivery or electrical stimulation to intervene.
Given that seizures are thought to involve increased levels of synchrony between brain regions, it seems intuitive that a seizure emerges as a result of increasing levels of synchrony between brain regions during a pre-ictal period. This increase in synchrony may only be in the region of the seizure focus, but one might expect that it could occur between larger scale brain areas such that pre-ictal increases in synchrony of brain activity could be detected by measuring synchrony between iEEG signals. It should also be expected that brain areas that are distant from the focus might initially be synchronized to some degree with the focus through global background activity in the brain. Then when a seizure starts to emerge at the focus, one might expect to see a decrease in synchrony between the focus and the distant region because the focus would be on its path to seizure while the distant region would still be following the global background activity of the brain. As a result one should expect to see both pre-ictal increases and/or decreases in synchrony depending on which brain regions one analyses (Winterhalder et al., 2006). The present paper evaluates both increases and decreases in synchrony as a means of seizure prediction. Mormann et al. (2005) also investigated whether or not increases or decreases in synchrony were predictive of seizures and found that both increases and decreases in synchrony in different channel pairs performed similarly for cases that showed statistically significant performance.
Previous work has compared the performance of a bivariate-synchrony-based seizure predictor with a Poisson process-based random predictor (Schelter et al., 2006b). However, the analysis by Schelter et al. (2006b) involved only a subset of recorded iEEG channels. This paper involves a comparison with a random predictor in which all channels are analysed. Moreover, two forms of random predictor are investigated: the analytical Poisson predictor (Schelter et al., 2006b, Schad et al., 2008) and alarm time surrogates (Andrzejak et al., 2009).
Section snippets
EEG data and patient characteristics
This study involves continuous long-term iEEG recordings from 3 patients recorded at the Epilepsy Center of the University Medical Center, Freiburg, Germany, and 3 patients recorded at the Epilepsy Clinic, St. Vincent's Hospital Melbourne, Australia. The Freiburg data was provided to participants in the seizure prediction contest of the 3rd and 4th International Workshops on Epileptic Seizure Prediction (http://www.iwsp4.org), and will likely be incorporated into the seizure prediction contest
Best performing channel pairs
The performance of the best performing channel pair and method for each patient is given in Table 2 for both decreases and increases in synchrony. Channel pairs and methods were ranked based on the R measure, and the optimal threshold was picked based on visual inspection of sensitivity versus FPR curves of the best pair and the random predictor. In particular, the optimal threshold was taken to be the threshold corresponding to the greatest difference between the sensitivities obtained with
Difficulties with synchrony-based prediction
While the whole notion of synchrony-based seizure prediction seems intuitively appealing with respect to how the epileptic brain transitions to the highly synchronous seizure state, there are significant limitations in the iEEG signal that make it difficult to detect relevant changes in synchrony. The main difficulties include electrode scale, electrode location, and noise. This study, which investigated whether or not there are pre-ictal decreases or increases in synchrony within 15 min of a
Acknowledgements
This work was supported by an Australian Research Council Linkage Project grant (LP0560684), The Bionic Ear Institute, and St. Vincent's Hospital Melbourne. We are grateful for the EEG data provided by the patients and to the St. Vincent's Hospital Melbourne Neurophysiology Clinic for collecting the data. We are also grateful to the Freiburg Seizure Prediction Group at the Epilepsy Center of the University Medical Center, Freiburg, Germany, for providing data from 3 patients. This data is
References (44)
- et al.
Seizure prediction: any better than chance?
Clin. Neurophysiol.
(2009) - et al.
A robust method for detecting interdependences: application to intracranially recorded EEG
Phys. D
(1999) - et al.
Cellular biology of epileptogenesis
Lancet Neurol.
(2003) Progress in predicting seizure episodes with nonlinear methods
Epilepsy Behav.
(2008)- et al.
Signal complexity and synchrony of epileptic seizures: is there an identifiable preictal period?
Clin. Neurophysiol.
(2005) - et al.
Characterization of synchronization in interacting groups of oscillators: application to seizures
Biophys. J.
(2008) - et al.
Comparison of Hilbert transform and wavelet methods for the analysis of neuronal synchrony
J. Neurosci. Methods
(2001) - et al.
Comparison of three nonlinear seizure prediction methods by means of the seizure prediction characteristic
Phys. D
(2004) - et al.
Classification of patterns of EEG synchronization for seizure prediction
Clin. Neurophysiol.
(2009) - et al.
Mean phase coherence as a measure for phase synchronization and its application to the EEG of epilepsy patients
Phys. D
(2000)