EEG Based Sentimental Analysis

Abstract

One of the main challenges for biomedical researchers is to recognize emotions from elec troencephalogram (EEG) signals, which are non-invasive measurements of the brain’s electrical activity. EEG signals are complex and dynamic, making it difficult to identify and understand the neural processes related to emotional states. Emotion recognition is an important aspect of human communication and interaction, and it has motivated the development of novel meth ods for accurate detection and analysis of emotional states. In recent years, deep learning has emerged as a powerful technique for emotion recognition using EEG data, mainly using con volutional neural networks (CNNs) as the dominant models. However, most of the existing deep learning models rely on complex feature engineering, which increases their computational cost and time. In this work, we propose a novel convolutional network and also fine-tune pre trained CNN models that can effectively capture and extract emotion-relevant information from simple 2D spectrograms derived from the 1D EEG signals. We use short-time Fourier trans form (STFT) and mel spectrograms to generate simple 2D representations of the EEG signals from the SEED dataset, which contains EEG signals labeled with emotions. We then train our proposed models on these spectrograms, without any complex feature engineering. Our proposed models are a novel simple convolutional network called EEG-based convolutional network (Electroencephalogram based convolutional network (EEG-ConvNet)), which consists of only five convolutional layers with batch normalization and max pooling; and fine-tuned ver sions of GoogLeNet and ResNet-34, which are pre-trained CNN models. On the SEED dataset of EEG signals, our proposed models achieve high accuracies: fine-tuned GoogLeNet achieves 99.97% and 99.95% accuracies on STFT and mel spectrograms respectively; fine-tuned ResNet 34 achieves 99.49% and 99.31% on STFT and mel spectrograms respectively; and our proposed EEG-ConvNet achieves 99.03% accuracy on STFT spectrograms. Our experimental results show that our proposed methods outperform the previous work on the SEED dataset using deep learning