|
|
| Line 2: |
Line 2: |
| {{Epilepsy}} | | {{Epilepsy}} |
| {{CMG}} {{AE}} {{VVS}} | | {{CMG}} {{AE}} {{VVS}} |
| == EEG ==
| |
| === Electrophysiology ===
| |
|
| |
|
| Most epileptics seize without warning, and their seizures can have dangerous or fatal consequences, if they come at a bad time and lead to an accident. In the brain, identifiable electrical changes precede the clinical onset of a seizure by tens of seconds, and these changes can be recorded in an [[electroencephalogram]] ([[EEG]]). Many people have wondered if EEG’s might be used to predict seizures minutes or even hours ahead of time, but as of now, this sort of prediction has not been feasible. Many researchers are working, however, to create a system capable of detecting seizures before they clinically manifest themselves.
| | == Overview == |
|
| |
|
| The early detection of a seizure has many potential benefits. Advanced warning would allow patients to take action to minimize their risk of injury and, in some circumstances, would allow them to summon help. An automatic detection system could also be made to trigger pharmacological intervention in the form of fast-acting drugs or [[electrical stimulation]].
| | == Electroencephalogram == |
| | | An ECG may be helpful in the diagnosis of epilepsy. Findings on an EEG suggestive of epilepsy include: |
| It is relatively easy to place the [[electrodes]] needed to record an EEG, but it has not been so easy to develop an [[algorithm]] to detect the onset of a seizure. For any given patient, assuming his or her seizures originate in one focus, seizure-onset EEG patterns are largely conserved from one seizure episode to the next. Unfortunately, there is great EEG variation between patients, both in terms of baseline and in terms of seizure-onset patterns. This variation has made the development of a generic, “one-size-fits-all” algorithm difficult.
| |
| | |
| Patient-specific algorithms based on [[machine learning]] have shown more promise. Machine learning algorithms compute [[binary decision trees]] from manually labeled training sets of data. EEG data must be translated into a format that the computer can interpret. Important information must be kept while superfluous information must be discarded. Although there are many conceivable ways of performing this “[[feature extraction]],” wavelet decomposition seems to be an effective way of extracting pertinent information from EEG signals.
| |
| | |
| The training set for the machine-learning algorithm must be labeled by hand. For an algorithm being developed by Dr. Schachter and Prof. Guttag of MIT, EEG recordings are split into two-second time windows, and each window is labeled as “seizure onset” or “not seizure onset.”
| |
| | |
| The algorithm then takes the labeled training set and uses it to construct a decision tree capable of classifying unlabeled EEG patterns as “seizure onset” or “not seizure onset.” The training set is unavoidably unbalanced because most time windows do not involve seizures. Certain algorithms, such as the support vector machine algorithm chosen by Schachter and Guttag, are better suited than others to handle this unbalanced training set.
| |
| | |
| In the hospital, the patient-specific algorithm of Schachter and Guttag has worked fairly well. In one trial, it detected 131 out of 139 seizures in 36 patients. In another, it caught 53 out of 58 seizures. The algorithm outperformed generic algorithms.
| |
| | |
| In the future, Dr. Schachter and Prof. Guttag hope to improve their algorithm so that it is less sensitive to electrode placement and so that it functions effectively with input from fewer electrodes and with smaller training sets. Their goal is to create an unobtrusive device that can be worn continually by epileptics to detect impending seizures. Such a device would greatly enhance the ability of these people to safely go about their lives.
| |
|
| |
|
| ==References== | | ==References== |