EEG-Brain Activity Monitoring and Predictive Analysis of Signals Using Artificial Neural Networks

Abstract

Predictive observation and real-time analysis of the values of biomedical signals and automatic detection of epileptic seizures before onset are beneficial for the development of warning systems for patients because the patient, once informed that an epilepsy seizure is about to start, can take safety measures in useful time. In this article, Daubechies discrete wavelet transform (DWT) was used, coupled with analysis of the correlations between biomedical signals that measure the electrical activity in the brain by electroencephalogram (EEG), electrical currents generated in muscles by electromyogram (EMG), and heart rate monitoring by photoplethysmography (PPG). In addition, we used artificial neural networks (ANN) for automatic detection of epileptic seizures before onset. We analyzed 30 EEG recordings 10 min before a seizure and during the seizure for 30 patients with epilepsy. In this work, we investigated the ANN dimensions of 10, 50, 100, and 150 neurons, and we found that using an ANN with 150 neurons generates an excellent performance in comparison to a 10-neuron-based ANN. However, this analyzes requests in an increased amount of time in comparison with an ANN with a lower neuron number. For real-time monitoring, the neurons number should be correlated with the response time and power consumption used in wearable devices.

Keywords: EEG; EMG; PPG; artificial neural network; brain monitoring; epilepsy; predictive analysis; signal processing.

Figure 1

Electroencephalogram (EEG) from a patient with no seizure-signal filtering and decomposition using the discrete wavelet transform (DWT) method.

Figure 2

EEG from a patient with epileptic seizure-signal filtering and decomposition using the DWT method.

Figure 3

Patient before and after the seizure, signal decomposition on four levels using DWT.

Figure 4

Patient with epileptic seizure, signal decomposition on four levels using DWT.

Figure 5

3D spectrogram of EEG signals from 13 channels.

Figure 6

3D spectrogram of signals EEG from 13 channels for patient n1 with no epileptic seizures.

Figure 7

3D spectrogram signals EEG from 13 channels for patient n2 with epileptic seizures.

Figure 8

Artificial neural network (ANN) with n neurons, n ∈ {10, 50, 100, 150}.

Figure 9

Regression (R²) for validation, test, and training—ANN with ten neurons.

Figure 10

Regression (R²) for validation, test, and training—ANN with 50 neurons.

Figure 11

Regression (R²) for validation, test, and training—ANN with 100 neurons.

Figure 12

Regression (R²) for validation, test, and training—ANN with 150 neurons.

Figure 13

Error histograms—ANN with 10, 50, and 100 neurons.

Figure 14

Error histogram—ANN with 150 neurons.

Figure 15

Neural network (10 neurons) best validation performance.

Figure 16

Neural network (150 neurons) best validation performance.

Figure 17

Distribution probabilities (Shapiro–Wilk test, Brainstorm).