IMAGE GENERATION USING BRAIN WAVE ACTIVITY (ELECTROENCEPHALOGRAM) IN MOTOR IMAGERY CASES WITH AUTOENCODER AND TRANSFORMER- BASED MODELS

• Main Author: Tohada Nainggolan, Ghebyon
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/86427
• Institution: Institut Teknologi Bandung

Electroencephalogram (EEG) has advanced in clinical diagnosis, Brain-Computer Interface (BCI) development, and cognitive behavioral research. Motor Imagery (MI) is a mental process where movement is imagined without physical execution, generating EEG signals with significant potential for BCI applications, such as image generation. Generating images based on EEG signals is an approach to better understand the meaning of EEG signals. However, the complex nature of EEG signals complicates analysis and implementation, necessitating precise data acquisition techniques. Currently, image generation models based on MI-EEG have not been extensively developed, prompting this research to address this need and explore the potential. The proposed solution to address this issue involves using a Transformer model for classification and a Variational Autoencoder (VAE) for image generation. The classification model was trained and tested using two variants and one type of training for evaluation: sliding window variant, channel selection variant, and error- proofing per subject. The sliding window variant with a 5 ms window achieved the highest performance with an accuracy of 58.45%, precision of 57.82%, recall of 57.89%, and an F1-Score of 57.57%. The error-proofing model showed the highest performance for subject J at 85.86%, with significant variation in classification performance across subjects, indicating potential issues with dataset quality. Based on this classification model, a VAE was developed for image generation, with performance results a ligning with the classification model's metrics. The image generation model successfully represents movements of the right hand, left hand, right foot, left foot, and tongue. Although the model successfully generated reconstructed images representing motor imagery signals, reconstruction accuracy is still highly dependent on the classification model's performance, variations in individual conditions, and factors in experimental design, such as subject concentration, fatigue, data acquisition configuration, and other aspects.