Classificação de sinais de EEG: uma abordagem da visão computacional com redes neurais convolucionais para predição de crises epilépticas

Abstract

Electroencephalography (EEG) is a non-invasive technique for recording and monitoring brain electrical activity. The EEG signals obtained provide valuable information about cognitive processes, mental states, and neurological conditions such as epilepsy. EEG plays a fundamental role as a diagnostic tool in medicine and extends to various fields, including neuroscience and biomedical engineering. One crucial aspect of EEG study is related to the detection and prediction of epileptic seizures. According to the World Health Organization (WHO), epilepsy affects approximately 50 million people worldwide. The traditional diagnosis of epilepsy involves visual analysis, conducted by a neurologist, of hours of EEG recordings to identify electroencephalographic patterns associated with seizures. However, this process can be costly and is subject to human errors. Faced with this challenge, researchers have been dedicated to seeking alternatives that can reduce analysis time and thus assist in more efficient diagnosis. As an alternative to this scenario, the application of artificial intelligence (AI) techniques can aid in the diagnostic process. The objective of our work is to propose a methodology for EEG signal classification to detect epileptic seizures. We used the open-source EEG dataset from Boston Children's Hospital, consisting of recordings of 198 seizures in 24 patients aged between 1.5 and 22 years. The data were pre-processed to remove noise and artifacts using the MNE Python library, specifically designed for neurophysiological analyses. The data were transformed into images for input into a Convolutional Neural Network (CNN). We implemented a 19-layer CNN in the Python language using the free Google Colaboratory platform. We selected 6 patients for model training using cross-validation and used the remaining 18 patients for individual tests to assess the model's generalization capability. We conducted four training sessions using different techniques to address data imbalance. Due to class imbalance that biases accuracy metrics, we considered Recall for class 1 as the most important metric for model selection for test data application. We achieved an average Recall for class 1 of 73% with unbalanced data and 83% with balanced data in the training set. In the test set, the averages were 38% and 74%, respectively. Computational resources were a limiting factor, influencing the size of the training set and the variability of patterns to be learned by the network. Nevertheless, we consider that our classification achieved good performance, especially given the challenge of dealing with extreme class imbalance.