SET Net A Hybrid Deep Learning Framework For EEG Based Attention Relaxation Classification In Brain Computer Interface Applications

Summary

This paper presents the squeeze-and-excitation transformer network (SET-Net) which is a hybrid CNN-Transformer network that recognizes attention and relaxation states using spectrogram features to identify EEG-based attention and relaxation states with 93.7% accuracy and high generalization to be used in Brain Computer Interface (BCI) applications.

Abstract

Brain Computer Interface (BCI) technology is used as a tool to provide a direct communication channel between the human brain and the external environment. This has applications in assistive and interactive systems. The non-invasive acquisition of neural activity using Electroencephalography (EEG) is standard with BCI systems. However, it is difficult to effectively classify cognitive states, including attention and relaxation, because the EEG signal is non-stationary and noisy. To overcome these shortcomings, this study proposes a new hybrid deep learning architecture namely, squeeze-and-excitation (SE) transformer network (SET-Net). In this method, EEG signals are preprocessed and divided into temporal windows to analyze the significant spectrogram representations. The proposed model achieves a classification accuracy of 93.7%, with F1-score of 0.93 and ROC-AUC of 0.98. This demonstrates the effectiveness of the hybrid deep learning model with enhanced discrimination of EEG-based attention-relaxation state. This offers a scalable system for real-time BCI applications as assistive technology in human-computer interaction.

Introduction

Brain–Computer Interfaces (BCIs) have become a disruptive technology that allows the human brain and external devices to communicate directly without using traditional neuromuscular connections. Electroencephalography (EEG) is the most common and widely used technique owing to non-invasive nature, portability, and cost-effectiveness¹. EEG-based BCIs have been used in healthcare as well as in neurorehabilitation, gaming, and assistive technologies for persons with neurological disabilities by examining brain activity². One of the most significant applications is the identification of cognitive states, including atte....

Protocol

The study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Ethics Committee of SR University, Warangal, Telangana, India (Approval No. 007/2026). Informed consent was obtained in writing from all participants before data collection.

Participant Selection

EEG data were collected from 11 healthy participants (age range: 25–45 years; mean age: 28 years) with no history of neurological or psychiatric disorders. In this study, a total of 1175 EEG segments were considered, of which 527 belong to the relaxation class and 648 belong to the attention class. Pa....

Results

Spectrogram-Based Feature Representation

The main goal of this study was to classify the states of attention and relaxation properly with the help of EEG signals. The conversion of raw EEG signals into spectrograms has made it possible to extract discriminative features effectively. The spectrograms indicate an augmentation of alpha band activity (8–13 Hz) during relaxation and an augmentation of beta activity (14–25 Hz) during attention. This distinct split in frequency-domai.......

Discussion

The current research suggests a hybrid deep learning model, SET-Net, for EEG-based attention-relaxation classification. To learn the spatial, spectral, and temporal features of EEG signals, the architecture combines convolutional neural networks, squeeze-and-excitation attention, and transformer encoders, as shown in Figure 5. The findings indicate that the proposed method can meet the study objective. The accuracy, F1-score, and ROC-AUC values (Table 6) are high, proving th.......

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Computer / Laptop	Generic	Any standard system	≥8 GB RAM recommended
Data acquisition script	Custom	collect.py	Python-based acquisition script
Disposable gel electrodes	Generic	Pre-gelled Ag/AgCl electrodes	Single-use, adhesive
EEG acquisition module	Upside Down Labs	BioAmp EXG Pill	Analog front-end for biopotential signals
Electrode gel (if reusable electrodes used)	Generic	Conductive gel	Improves signal conductivity
Interactive application (game)	Generic	Any simple game (e.g., racing)	Keyboard-controlled interface
Jumper wires	Generic	Male–Female connectors	Standard breadboard wires
Matplotlib / Seaborn	Open-source	Latest versions	Visualization libraries
Microcontroller board	Arduino	Uno / Maker Uno	10-bit ADC, USB interface
Python programming environment	Python Software Foundation	Python ≥3.8	Open-source programming language
PyTorch library	Meta AI	Version ≥1.12	Deep learning framework
Scikit-learn	Open-source	Latest stable version	ML utilities
Signal processing libraries	SciPy	Latest version	Filtering, STFT
USB cable	Generic	USB Type-A to B	Data transfer cable