Detecção de crises epilépticas a partir de sinais de eletroencefalograma usando redes neurais autocodificadoras

Abstract

Seizures are abnormal conditions in an individual’s brain that can cause spasms, convulsions, or even more serious complications. One way to identify these anomalies is through an encephalogram , a test responsible for capturing the patient’s brain activity. With this data, pattern recognition models can be used to determine whether the received signals are normal or not. This work proposes the modeling of such signals with autoencoder neural networks, algorithms capable of detecting anomalies in the EEG signal based on the model’s reconstruction error. Additionally, a training approach was used to simulate the real-time acquisition of signals, enabling the extension of the application for this type of use. The following models were evaluated in the anomaly detection task: linear autoencoder (PCA) and autoencoder networks with 1 and 3 hidden layers. Several simulation scenarios were developed, namely: (i) Scenario 1: Offline training and testing using all patient datasets. Training is performed only with negative (i.e., non-seizure) data; (ii) Scenario 2: Similar to Scenario 1, however, training includes some positive instances (Open Set scheme); and (iii) Scenario 3: Offline training with a single realization of the EEG signal. Online testing with the remaining patient EEG signals. Among the trained models, the one that presented the best performance for the metrics used was the 3-hidden-layer autoencoder network, achieving 91% (accuracy), 93% (sensitivity) and 91% (specificity).