Titulo Estágio
ESP-RC: Epileptic Seizure Prediction Based on Reservoir Computing
Áreas de especialidade
Sistemas Inteligentes
Local do Estágio
DEI-FCTUC
Enquadramento
Despite available drug and surgical treatment options, more than 30% of patients with epilepsy continue to experience seizures. An option to improve the quality of life of these patients encompasses the development of algorithms that could warn the patient of an impending seizure or trigger an antiepileptic device to prevent seizure occurrence. Classification algorithms, based on electroencephalogram (EEG) features, have been applied to discriminate the different brain states, namely by computational intelligence methods (feed-forward artificial neural networks (FF-ANN) and support vector machines (SVM)). The results obtained with FF-ANN and SVM presented some success on predicting seizures, but far from the performance required for a real clinical application. Thus, new approaches are needed to achieve the quality levels required for the clinical use of seizure prediction algorithms.
Objetivo
This proposal aims to develop a new methodology based on recurrent artificial neural networks for seizure prediction (RNN). Unlike FF-ANN, RNN are defined by presenting internal loops between neurons. This propriety leads to the network to exhibit dynamic temporal behaviour, i.e., memory of their past activity. The dynamical temporal behaviour enables RNN to be more powerful than FF-ANN in classification of time-series, such as the EEG features available for seizure prediction. The problem with the application of RNN to real world problems is the complexity of their training by the usual gradient-descent methods. The recently reservoir computing (RC) approach is referred to solve the training problem of RNN and will be considered in this master thesis [1]. The RC approach is defined by using a large and random RNN as an excitable medium where each neuron generates its non-linear transform of the input, and where the overall output is a linear combination of some outputs from the neurons inside the network.
[1] Lukoševičius, M., Jaeger, H., Reservoir computing approaches to recurrent neural network training, Computer Science Review, 3(3):27-149 (2009).
Plano de Trabalhos - Semestre 1
1. Literature review
• Objective: In the first phase the student should understand the problem of seizure prediction by computational intelligence methods, and understand the main methods based on reservoir computing.
• Start: September 2015
• End: November 2015
• Expected Result: A report describing the actual formulations of seizure prediction by computational intelligence methods, and the main methods based on reservoir computing that are available.
2a. Implementation of RC methods
• Objective: Development of a toolbox for seizure prediction by RC.
• Start: October 2015
• End: Dezember 2015
• Expected Result: A toolbox that should enable the development of predictors based in RC.
Plano de Trabalhos - Semestre 2
2b. Implementation of RC methods
• Objective: Development of a toolbox for seizure prediction by RC.
• Start: January 2016
• End: March 2016
• Expected Result: A toolbox that should enable the development of predictors based in RC.
3. Analysis of the algorithms in a large set of patients
• Objective: Using the developed toolbox, predictors for 100 patients from the EPILEPSIAE database will be developed.
• Start: February 2016
• End: June 2016
• Expected Result: A report describing the results obtained in the 100 patients considered.
4. Thesis writing
• Start: May 2016
• End: July 2016
Condições
The work is supposed to be developed on the Adaptive Computation Group Laboratories-CISUC, under normal logistic conditions.
Observações
Depending on the work quality, publications in international conferences may occur.
Orientador
César Alexandre Domingues Teixeira
cteixei@dei.uc.pt 📩