Low-Complex TD-RBF and TD-SVM Seizures Predictors Based on EEG Energy and ECG Entropy
Authors
Abstract
Purpose:Low complexity classifiers are faced in this work for seizure prediction based on energy and entropy features requiring low computational costs. Time-Delay Radial Basis Function Neural Network (TD-RBFs) and Time-Delay Support Vector Machines (TD-SVMs) are used to study their potential since they have low computational complexity and a good generalization capability.
Methods:
The EEG long-term energy (EEG-LTE) and ECG entropy (ECG-ENT) were computed in data from the 4th IWSP seizure contest. Up to eight data sets, corresponding to 8 recording hours, from Patient 1 were picked for this study. A qualified technician identified the focal channels and then EEG-LTE and ECG-ENT were computed. Each EEG-LTE and ECG-ENT sample is based on a window of 300 seconds, being the inter-window overlap of 295 seconds.
Once features were computed, a classifier was applied to differentiate among the possible two or four cerebral states (pre-ictal, ictal, pos-ictal and inter-ictal). A TD-RBF is a three-layer network, being the first layer a set of inputs related with the actual and past features and/or output information. A second layer formed by a set of neurons, and the last layer is a linear combiner. TD-SVM has similar structure to TD-RBF with the consideration of a threshold at the output.
Results and Discussion:
EEG-LTE based TD-RBF presented appropriate results for training in four hours and validation in one hour. Sensitivities above 60% and specificities between 48% and 95% were obtained. On the other hand, TD-SVM presented a slightly superior performance. ECG-ENT based TD-SVM were trained in seven hours and then tested in the remaining 1 hour, presenting 83% and 98% of sensitivity and specificity, respectively. These preliminary results show that further efforts should be devoted to these low-complex EEG/ECG features and non-linear classifiers since, for the used data, the obtained performance is very promising.