Abstracts

Rationale: Seizure forecasting algorithms use high dimensionality data to evaluate the likelihood of an impending seizure. Using such feature spaces is constrained by the dimensionality and over fitting problems what would degrade the performance of the classifier and create a major obstacle toward the development of online prediction algorithms. Feature selection is desirable in order to find the most discriminative characteristics and use them as inputs to the classifier. A patient-specific algorithm based on a reduced feature space is proposed in this study.Methods: The proposed algorithm consists of preprocessing, feature extraction, electrode selection, feature selection and classification. Band pass and notch, zero phases, finite impulse response filters were used to remove the artifacts of intracranial electroencephalography (iEEG) recordings. Features of relative spectral power, decorrelation time, Hjorth mobility and complexity, as well as spectral edge frequency and power were extracted from 5 sec-long non overlapping windows. A hybrid double selection method is then proposed for an optimal selection of the electrodes-features combination. The most discriminative electrodes were selected using the minimum redundancy maximum relevance approach while the combination of the most contributing features was found by means of a Genetic Algorithm (GA). A Support Vector Machine (SVM) and an Adaptive Neuro Fuzzy Inference system were used for the classification of interictal and preictal samples. The performance is evaluated in terms of sensitivity and specificity. A hold out validation is used to perform on-sample optimization (selection and training) and out-of-sample testing. The algorithm was tested on iEEG recordings from 5 dogs available at the kaggle seizure prediction challenge repository.Results: The results show that the selected subset of features performs almost equally and sometimes even better than the whole set. The double selection method reduced the number of features form 224 to less than 6 features (mean: 5, STD: 1). Using an SVM, the algorithm has achieved a mean sensitivity of 90.28% and a specificity of 88.53% after selection compared to a mean sensitivity of 85.49% and a specificity of 80.11% when using the whole feature set. The algorithm’s performance was evaluated using the whole set of features, the mRMR selection method only, the GA selection method only and the proposed approach. Figure 1 shows the ROC curves obtained when using different methods for dog1. The area under the curve has increased after the selection of features.Conclusions: It has been demonstrated that the reduction of feature dimensions is possible in seizure prediction studies. The selected subsets of features were different for all five dogs confirming the need for subject specific, individually tailored algorithms. Band power features, mainly the Gamma band, were the most frequently selected (68% of the selected features). These features have already demonstrated their suitability for seizure prediction.