MLAR-Net: A Multilevel Attention-Based ResNet Module for the Automated Recognition of Emotions Using Single-Channel EEG Signals

• Published in: IEEE access Vol. 13; pp. 99122 - 99144
• Main Authors: Maithri, M., Raghavendra, U., Gudigar, Anjan, Kumar Praharaj, Samir, Sriram, Karthikeyan, Salvi, Massimo, Hong Yeong, Chai, Molinari, Filippo, Rajendra Acharya, U.
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; attention-based ResNet; Availability; Biomedical imaging; Brain; Brain modeling; Channels; Computational modeling; Datasets; Electrodes; Electroencephalography; Emotion recognition; Emotions; explainable AI; Feature extraction; Modules; Multilevel; Real time; Real-time systems; single-channel EEG; Spectrogram; Spectrograms; Support vector machines
• DOI: 10.1109/ACCESS.2025.3576059
• ISSN: 2169-3536

Human emotion recognition is important as it finds applications in multiple domains such as medicine, entertainment, and military. However, accurately identifying emotions remains challenging due to humans' ability to hide or suppress their emotional expressions. Hence it becomes important to recognize emotions by using brain signals as they provide more reliable data. Brain signals can be captured using Electroencephalograms (EEG) electrodes. Most used EEG devices come with multiple channels. However, not all channel information is important for emotion recognition. Another issue with the existing dataset is the availability of a small quantity of samples. To address these challenges, we propose MLAR-Net, a novel multilevel attention module for emotion recognition using single-channel EEG signals. Our approach converts EEG signals into spectrograms using multiple parameters to generate a large set of images. This data is then processed through our proposed MLAR-Net, which integrates a multilevel attention module with ResNet18 architecture. Our study identifies channel number 24 (T7) as the most effective for emotion classification, achieving an average accuracy of 98.06% using a cubic support vector machine and a maximum accuracy of 99.51% using fine K-Nearest Neighbors. The study was conducted using the SEED dataset. It is a publicly available dataset developed by capturing EEG signals from fifteen subjects for three classes of emotions, namely positive, negative, and neutral. The results achieved by the proposed study show an improvement of around 4 to 5% compared to state-of-the-art studies using the same channel. This performance surpasses existing state-of-the-art methods for single-channel EEG-based emotion recognition. Furthermore, we highlight the top-performing channels that can be used for real-time implementation of the system with a minimum number of channels.