Visual Evoked Potential Classification Support with Convolutional Neural Network and Recurrent Neural Network - A comparative study

• Published in: 2022 International Conference on Decision Aid Sciences and Applications (DASA) pp. 1612 - 1617
• Main Authors: Cheker, Zineb, Chakkor, Saad, El Oualkadi, Ahmed
• Format: Conference Proceeding
• Language: English
• Published: IEEE 23.03.2022
• Subjects: Automatic diagnostic; Biomedical signals; Convolutional neural network; Convolutional neural networks; Latency P100; Optical filters; Pathology; Recurrent neural network; Recurrent neural networks; Stimulated emission; Training; Visual Evoked Potential; Visualization
• DOI: 10.1109/DASA54658.2022.9765034

The analysis of biomedical signals is an interdisciplinary subject, used to develop automatic diagnostic systems for decision support. Among these biomedical signals, we find the VEP: Visual Evoked Potential signals, which are used to analyze the appropriate functioning of the optical pathways. It is noted that the visual reading, as well as the measurement of the amplitude and latency P100, are used to interpret and analyze this signal in medical diagnostics. However, this technique is not very reliable, because the latency is sensitive to different factors such as stimulation conditions, age, sex, etc. On the other side, we find different cases when some pathological VEPs show a normal latency. In this work, we suggest an automated and accurate classification of Visual Evoked Potential signals, in order to make objective decision support, by proposing an optimized deep learning approach based on the exclusion test using recurrent as well as convolutional neural networks, without any features selection/extraction procedure. To differentiate between the effectiveness of the two architectures examined in our study, we applied training and testing procedures in Matlab. The comparison study findings show that Convolutional Neural Network CNN-1D allows a good accuracy reaching 96% compared to Recurrent Neural Network RNN equal to 88%. This difference is justified by the fact that RNN uses signal values as inputs when CNN has the specificity of calculating features using convolution filters.