The seizure prediction characteristic: a general framework to assess and compare seizure prediction methods
Introduction
The recurrent and sudden incidence of seizures can lead to dangerous and possibly life-threatening situations [1]. Since disturbance of consciousness and sudden loss of motor control often occur without any warning, the ability to predict epileptic seizures would reduce patients’ anxiety, thus improving quality of life and safety considerably [2]. Constraints in everyday life would be alleviated, and secondary behavioral disturbances might be avoided. Knowing in advance that a seizure will occur could widen therapeutic options dramatically. For example, long-term treatment with antiepileptic drugs, which may cause cognitive or other neurological side effects, could be reduced to a targeted and short-acting intervention [3].
During the last decade, several methods have been suggested for prediction of epileptic seizures, based on intracranial or scalp EEG recordings, that use concepts of linear and nonlinear time series analysis [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16]. It has been claimed that seizures can be predicted at least 20 minutes beforehand, maybe up to 1.5 hours prior to onset for temporal lobe epilepsy. However, there has been so far no evaluation of the performance of seizure prediction methods based on long-term high-quality data [17]. Furthermore, recognized performance standards for assessing and comparing seizure prediction methods are lacking [18]. Up to now, most seizure prediction methods have been evaluated by analyzing few and brief preseizure data sets to obtain their sensitivity. Moreover, no or insufficient interictal data have been investigated to determine their specificity.
In 1998, Osorio et al. proposed that both seizure detection and prediction methods should be evaluated with respect to sensitivity and false prediction rate [19], [20]. We have extended this approach and suggest the “seizure prediction characteristic” to evaluate and compare the performance of seizure prediction methods. This assessment criterion is based on clinical and statistical considerations.
In the following, we focus on the properties and basic requirements of a clinically applicable seizure prediction method, which determine its assessment criterion in a straightforward way. Our approach is illustrated by its application to the “dynamical similarity index,” a seizure prediction method introduced by Le van Quyen et al. [7]. For this purpose, we have used intracranial EEG data from 21 patients with pharmacorefractory focal epilepsy. The examined data pool comprises 582 hours of EEG data and 88 seizures.
Section snippets
Seizure prediction methods and intervention systems
A clinical application controlling seizures consists of a seizure prediction method that raises an alarm in case of an upcoming seizure and an intervention system that is able to control a seizure (Fig. 1). For a successful application the properties and interdependencies of these two components have to be considered.
A seizure prediction method has to forecast an upcoming epileptic seizure by raising an alarm in advance of seizure onset. A perfect seizure prediction method would indicate the
Sensitivity and false prediction rate
A seizure prediction method should forecast a high percentage of seizures. This “sensitivity” is calculated as the fraction of correct predictions to all seizures. In a realistic setting, false predictions cannot be prevented and have to be permitted if they appear scarcely. They are quantified by the number of false predictions in a given time interval, the false prediction rate (FPR), which is the appropriate measure for specificity in the present context.
To increase sensitivity, the
The maximum false prediction rate FPRmax
It may not be possible to circumvent false alarms completely, but their negative impact leads to the question of how many of them can be tolerated per time unit. The negative effects of false predictions depend on the chosen intervention system. In the case of a simple warning, the patient prepares himself or herself during the seizure prediction horizon and expects a seizure at any moment during the seizure occurrence period. Since in the case of a false prediction, the seizure will not arise
Minimum seizure prediction horizon (SPHmin) and maximum seizure occurrence period (SOPmax)
All intervention systems require a certain period to become effective. Whereas implanted devices may need only a few seconds to control an upcoming seizure, a warning system has to predict the seizure at least tens of seconds before onset, providing enough time to prevent dangerous situations. This intervention period determines the minimum seizure prediction horizon (SPHmin) for a successful clinical application.
Similarly, the chosen intervention system determines the maximum seizure
Unspecific seizure prediction methods
Seizure prediction methods should have a significantly higher sensitivity than unspecific ones like the random and periodical prediction methods.
Assessing seizure prediction methods
The parameter set of a seizure prediction method is adjusted until the method is most sensitive without producing false predictions exceeding the upper bound FPRmax. Therefore, interictal data of at least 1/FPRmax duration are required to verify the condition for the false prediction rate. For example, to verify FPRmax corresponding to one false alarm per day, at least 24 hours of interictal data are necessary. During this time interval only one false prediction is permitted. Even more EEG data
An application: the dynamical similarity index method
Le van Quyen et al. introduced a seizure prediction method called “dynamical similarity index” [7]. In several studies, they applied their method to EEG data from patients suffering from temporal lobe epilepsy [8], [9] and neocortical epilepsy [15]. We implemented the dynamical similarity index as introduced in [7]; a brief description of the method is given in Appendix A.
The dynamical similarity index was applied to a large data pool of intracranial EEG data from 21 patients suffering from
Conclusion
The unpredictability of upcoming seizures is a central problem for patients with uncontrolled epilepsy and for their relatives [24]. The level of uncertainty and the associated stress of a patient will be reduced dramatically if a correct prediction of seizures is possible [25], leading to a higher degree of perceived self-control [26]. To contribute to a reduction in uncertainty about the imminent occurrence of a seizure, it is necessary to consider both false and correct predictions.
The above
References (27)
- et al.
Anticipation of epileptic seizures from standard EEG recordings
Lancet
(2001) - et al.
Spatio-temporal dynamics of the primary epileptogenic area in temporal lobe epilepsy characterized by neuronal complexity loss
Electroencephalogr. Clin. Neurophysiol.
(1995) - et al.
Epileptic seizures may begin hours in advance of clinical onset: a report of five patients
Neuron
(2001) - et al.
Mean phase coherence as a measure for phase synchronization and its application to the EEG of epilepsy patients
Physica D
(2000) - et al.
EEG analysis with simulated neuronal cell models helps to detect pre-seizure changes
Clin. Neurophysiol.
(2002) - et al.
Prediction of epileptic seizures
Lancet Neurol.
(2002) - et al.
Course of chronic focal epilepsy resistant to anticonvulsant treatment
Seizure
(2001) Coping with the uncertainty of uncontrolled epilepsy
Seizure
(1993)- et al.
Patients’ experiences of injury as a result of epilepsy
Epilepsia
(1997) The brainstorms companion: epilepsy in our view
(1994)