Titulo Estágio
Hybrid Feature Selection over EEG Multiple Electrodes for Therapeutic Epilepsy
Área Tecnológica
Informática Médica
Local do Estágio
DEI
Enquadramento
Epilepsy is one of the most common neurological disorders. Part of the epileptic patients cannot be treated by any drug or even by surgery. The life of these patients is affected by the occurrence of sudden (and apparently) unpredictable seizures. Thus, the seizure prediction technology is extremely important for use in therapeutic epilepsy devices to trigger intervention before a crisis can occur. In this kind of problems, one essential component is that the features are representative enough for accurate (and timely) diagnosis. During the last few years numerous features derived from EEG (electroencephalogram)/ ECG (electrocardiogram) signals have been extracted. Based on these features, appropriate classifying techniques are often used to predict the on-set seizures. Moreover, given the different sources of features available which makes it a multimodal problem by nature, it is highly probable that a hybrid feature selection method ought to be designed. This raises several challenging algorithmic issues.
Objetivo
In this project, the goal is to develop feature selection algorithms for epileptic seizure prediction. A literature review on the most widely used algorithms ranging from Filters, Wrappers, Feature Recursive Elimination, Genetic Algorithms, Mutual Information Criteria etc. should be done for this problem. In particular, the aim is to know whether one can use any of these algorithms for the (successful) final goal of epileptic seizure prediction. To know whether this is the true for the chosen classifiers is highly relevant, since we could use this information to develop more fast algorithms.
The seizure prediction problem can be viewed as a classification problem. i.e. one can assume that the different features samples (patterns) extracted over time can be linearly or non-linearly separated in two or more classes that discriminate the different cerebral states. Computational intelligence methods, such as support vector machines and artificial neural networks, have been tentatively applied to deal with this classification problem.
A problem that must be solved aiming at the development of efficient (and feasible) predictors is the selection of the appropriate features. Among all extracted features, some may not contribute positively or introduce no novelty to the prediction performance. Besides the problem of discarding irrelevant or useless features, another issue is to identify the training patterns that are relevant for the classification. Therefore, the aim is to select the features that are close to the decision surface rejecting those that are far apart. This procedure reduces complexity, improves testing (and recall phase) performance and reduces time.
For this project the prospective student should develop a system to select the appropriate features (and related patterns) among a set of 642 features, extracted from the patient EEG and ECG. The student will develop feature selection algorithms for the epileptic seizure prediction problem by following the background available in the team. Furthermore, he/she will perform a large experimental analysis by taking into account several types of feature selection methods to analyze the performance of several classify methods by following an appropriate methodology according to the literature.
This work will be performed within the FP7 European project EPILEPSIAE- Evolving Platform for Improving Living Expectation of Patients a leaded by ACG/CIUSC, Portugal (see http://www.epilepsiae.eu) . There will be funding available within the research group that is a leading partner of this project, depending on the curriculum and the disponibility of the candidate.
Plano de Trabalhos - Semestre 1
1 Month: Problem identification
1 Months: Literature review: Feature Selection Methods for Designing Robust Classifiers
1 Month: Implementation of chosen Feature Selection Algorithms in a small set of patients available in the database.
1 Month: Writing of the preliminary report
Plano de Trabalhos - Semestre 2
1 Month: Overall Algorithm implementation
1 Month: Experimental analysis in a large set of patients in the database.
1 Month: Results analysis
1 Month: Thesis writing
Condições
The ACGLab and its computing resources are available for performing experiments. The research group owns a computer cluster for the experimental analysis.
Observações
Orientador
Doutor César Teixeira
cteixei@dei.uc.pt 📩