Detecção e classificação de potenciais evocados auditivos baseadas em filtros casados

• Ano de defesa: 2020
• Autor(a) principal: Spirandeli, Amanda Franco
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Engenharia Biomédica
• Palavras-chave em Português: Filtros casados; Matched Filters; Interface cérebro máquina; Brain Machine Interface; Potenciais relacionados à eventos; Event related potential; Eletroencefalografia; Electroencephalography; CNPQ::CIENCIAS EXATAS E DA TERRA
• Link de acesso: https://repositorio.ufu.br/handle/123456789/29177; http://doi.org/10.14393/ufu.di.2020.272

A Brain Machine Interface (BMI) is a direct communication system between the brain and an external device. For a special group of subjects suffering, for example, from locked in syndrome, exogenous BMIs based on oddball paradigms have been presented in the literature. Such systems exploit event-related potentials (ERPs) that are evoked by specific stimuli and make use of its components such as the N200 and P300 as input signals for communication / interface control. However, current models for feature extraction and classification of cortical signals for the control of BMIs depend on complex machine learning algorithms with high computational cost, such as neural networks and support vector machines. Therefore, the rate of commands that can be decoded by such BMIs is still below the desired. In this context, the development of novel decoding methods that can improve the real-time performance of exogenous BMIs is highly desired. Matched filters have long been used as a real-time sorting tool in telecommunication systems and is a valid approach for detecting a known signal under Gaussian noise. In general, it is an easy to implement system that maximizes the output signal-to-noise ratio at any given time without requiring high computational costs. In this work, a novel framework is proposed based on a matched filter bank, associated with an electroencephalographic signal acquisition (EEG) system, feature extraction and peak detection for classification of neural information associated with ERPs. The system was validated using both synthetic EEG signals and real EEG data obtained from BMI databases. The results demonstrate the viability of the proposed framework. In the experiments performed, we sought to discriminate between two target classes to allow the detection of ‘yes’ and ‘no’ answers. From simulated data the method obtained on average 90.27% of performance and 7.70 bit / min. Validation with real EEG data resulted in an average accuracy of up to 77.50%, with information transfer rate of up to 3.19 bit / min. The validation of the system demonstrated satisfactory results, allowing the decoding of PREs with the dependence of few signal characteristics, without fine adjustments, paving the way for the use of BMI control algorithms in embedded systems assisting individuals affected by the incarceration syndrome.