2. ANÁLISIS SITUACIONAL INTERNO 16
2.1. ASPECTOS GENERALES 16
In this section we present a number of the reported issues within the seizure prediction line of research:
Poor use of Machine learning
When machine learning algorithms made their debut in the field of EEG seizure prediction, they were merely used as feature selection tools to narrow down the pre- engineered features and channels that yielded the highest in-sample performance. The community had yet to use machine learning for the actual task of classification. With papers such as (Shoeb et al. 2004; Costa et al. 2008; Santaniello et al. 2011) more sophisticated machine learning methods were used as classifiers for seizure prediction tasks, particularly Artificial Neural Networks. In the majority of the machine learning enhanced studies, a single EEG feature is used for the classification task in order to showcase the power of the particular machine learning algorithm in use. There is little work that combines the powerful features developed in early, seizure prediction research with state of the art machine learning algorithms.
Out-of- sample algorithms
The main fraction of research in EEG seizure prediction involves using simple tuning of well-know binary classification of in-sample data (Mormann et al. 2007). This entails that the classifier has not been verified on unseen data and is likely to perform poorly on out-of-sample predictions. In order for the findings of a seizure detection study be
extended to other datasets, the performance of the algorithm should only be published for the test data, which is essentially required to be unseen ‘out-of-sample’ data (Mormann et al. 2007).
Confounding variables
One other point worth consideration is the confounding variables during the inter-ictal state, which may influence the characterizing features used for the prediction algorithm. Failing to identify and understand such variables could potentially affect the Specificity and Sensitivity of the algorithm. Therefore, building the algorithm should involve features from both the ictal and inter-ictal stage and work towards a better understanding of confounding variables which may be a result of slow-wave sleep, emotional and cognitive states (Mormann et al. 2007).
Mechanisms of Ictogenesis
Another issue that has been somewhat overlooked in most studies, is the process and mechanism of seizure generation itself (Ictogenesis). Some studies have found the mechanism behind certain types of epilepsy (Kalitzin et al. 2002) and some suggest that there may be different seizure generation mechanisms for different brain structures and pathologies, implying that seizure initiation could vary from person to person. Therefore, using EEG prediction algorithms in understanding these mechanisms, and also using these mechanisms to develop better prediction algorithms could possibly result in better research outcome. Some studies have suggested modeling EEG signals in order to have an insight into the dynamical process of seizure-generation through time (Wendling et al. 2003; Suffczynski et al. 2006).
Seizure prevention
One potentially groundbreaking area of research that is often disregarded, is designing intervention systems, which in addition to warning the patient about a seizure, will also prevent this from happening. (Stein et al. 2000) have looked into the local application of short-acting powerful drugs. In another study, electrical stimulations have been suggested, with major focus on deep brain stimulation intervention (Theodore & Fisher 2004) which in a nutshell, uses electrodes to alter the state of the brain from the ictal state. These forms of intervention could benefit from seizure prediction, but also, from
early seizure detection. Seizure prediction, predicts the time at which the seizure could occur, well in advance of EEG ictal state, whereas early seizure-detection, focuses on detecting the seizure onset before the clinical symptoms occur with little time for intervention. The research in this area is very recent and further studies should be carried out in order to investigate potential real-life applicability of such concepts. Surface EEG vs. Invasive EEG
When it comes to using EEG recordings for experiments, much care has to be taken as to which type of recording is being used. Intracranial recordings are used in the majority of seizure-prediction studies and provide much better signal to noise ratio and less artifact, with the added benefit of being recorded directly from the seizure- generating area of the brain, whereas the surface EEG can provide an overall image of all areas of the brain, which is useful for understanding the effects of the environment on seizure generation. However, if closed-loop interventions were to be used, patients would have to wear the EEG cap the entire time in order to monitor their surface recordings. Therefore there are doubts regarding the usefulness of scalp EEGs for intervention purposes (Morrell 2006).
Data requirements
In order to have a reasonable separation between inter-ictal and pre-ictal stages, it is advised to use EEG recordings, which not only have large number of seizures but also have sufficient time interval between the seizures (Mormann et al. 2007).
3.4 Summary
In this chapter, a rich and comprehensive background on epilepsy and an introduction to EEG and its role in various areas of epilepsy diagnosis and treatment were presented. This information is crucial in understanding the problem of seizure prediction and therefore a great segment of this chapter was dedicated to the description of these two topics.
The development of seizure prediction research in the early days up until the state of the art was also presented along with relevant research literature. Finally, common techniques and challenges in the field of seizure prediction were identified. 4 4