Constrained Seizure Detection
- Kelleher, D., Faul, S., Temko, A., and Marnane W.P., “On the Effect of Reduced Sampling Rate and Bitwidth on Seizure Detection”, IEEE International Symposium on Intelligent Signal Processing (WISP2009), Budapest, Hungry, August 26-28, 2009, pp. 153-156.
- Faul, S., Temko, A. and Marnane, W.P., "Age-Independent Seizure Detection”, 31st Annual International IEEE EMBS Conference, Hilton Minneapolis, Minnesota, USA, September 2-6, 2009.
EEG recording systems are generally implemented on computer systems that are confined to a hospital ward. This setup requires that patients are admitted to hospital for a number of days to have their EEG recorded, which is afterward analysed by clinical staff. One way to free up hospital space and the patient, is to use a portable EEG recorder. In this way a patient can stay at home while their EEG is analysed.
While these portable devices allow the patient more freedom, they are still limited in their functionality. EEG is simply recorded to a hard drive or flash memory, and the patient returns to the clinic to have the data downloaded. This data may or may not contain epileptiform activity, and the patient may have to retain the recorded for a further period.
The aim of this research is to use the groups expertise in signal processing, and in particular seizure detection, to introduce a level of intelligence to these devices. This would have a number of advantages for the patient and clinical staff. Rather than the EEG signals being blindly recorded, requiring a large amount of memory and hence power and bulk, algorithms can be used to compress the EEG and even to decide what data is important to record. Communications functionality can provide the clinician with information about the recorded data, without the need to analyse the whole data recording in detail, saving a large amount of time.
While these advances in functionality and intelligence allow for more efficient diagnosis and analysis, they come with the price of added power consumption. Therefore the main focus of this research is on the implementation of these signal processing and communications techniques in a power constrained manner, allowing for smaller batteries and hence smaller, lightweight devices. This will greatly improve the standard of care for patients requiring EEG analysis.
One method of power reduction is to reduce the amount of data that needs to be analysed by the system while maintaining the accuracy of the classification. Two methods are primarily investigated:
i) down-sampling of the electroencephalogram (EEG) data
ii) quantisation of the EEG
A Support Vector Machine (SVM) classifier is used in the system to differentiate between seizure and non-seizure segments. The following diagram gives an overview of the detection system used:
Figure 1 - Seizure Detection System Overview
The above system is composed of four main parts, which are explained in more detail as follows:
i) Pre-processing: A notch filter is firstly applied to the raw EEG data to get rid of 50 Hz mains interference. Next the data is down-sampled from the initial sampling rate of 256 Hz to 128 Hz, 64 Hz and 32 Hz respectively. This is followed by quantisation of the data from full 32 bit precision to 16 bits, 12 bits and finally, 8 bits, before the EEG signal is passed onto the feature extraction stage.
ii) Feature Extraction: During this step, the EEG signal is split into overlapping epochs. An overlap of 50% is taken as a reference in most of the literature (for example, an epoch length of 10 s with 50% overlap would mean that successive epochs have 5 s of overlap with one another). A set of 47 features for each channel of each epoch of the EEG signal are computed, and these feature sets are then passed onto the final stage of the system.
iii) Classification is performed by means of an SVM. SVMs are binary classification schemes which require the linearly non-separable data to be transformed to a higher dimension feature space where it is separable. Classification consists of two stages: firstly, training of the classifier on a section of the data and secondly, testing of the system on the remainder of the unseen data.
iv) The final step of the overall seizure detection system is a post-processing stage. The output from the classification is a vector representing the classifiers confidence that any given epoch of each channel is either seizure or non-seizure. A moving average filter is first applied to the output to minimise noise (primarily to reduce the false detection rate). This filtered output is then compared to a series of threshold values to obtain a binary result, indicating seizure or non-seizure activity. Finally, a collar is applied to this binary result. This collar essentially extends a detected event by a number of epochs on either side, thus increasing the accuracy of the system.