MOTOR-IMAGERY BASED BCI USING CONVOLUTIONAL NEURAL NETWORK AND ANALYSIS OF LOW-COMPLEXITY HILBERT TRANSFORM

Abstract

Motor imagery (MI) based brain-computer interfaces (BCIs) enable direct communication between the brain and external devices, offering significant potential for individuals with motor disabilities. This thesis explores the development of a Motor-Imagery based BCI using Convolutional Neural Networks (CNNs) and evaluates the effectiveness of the Low Complexity Hilbert Transform for feature extraction. EEG signals are preprocessed using a notch filter to remove power line interference, followed by bandpass filtering to isolate the mu (8-12 Hz) and beta (12-30 Hz) frequency bands. The Hilbert Transform is applied to compute analytic signals, from which instantaneous power is derived. Event-Related Patterns (EPs) are calculated to quantify changes in brain activity during motor imagery tasks. Statistical features such as mean, standard deviation, skewness, and kurtosis are extracted from the EPs to enrich the feature set for classification. Both CNN and Long Short-Term Memory (LSTM) networks are implemented and evaluated. Contrary to common expectations, the CNN model outperformed the LSTM model, achieving higher classification accuracy on both training and test datasets. The CNN demonstrated superior capability in learning and generalizing spatial patterns within the EEG data. A detailed analysis using confusion matrices highlights the CNN's effectiveness in capturing intricate spatial features crucial for accurate MI classification. This thesis underscores the importance of spatial feature extraction and suggests that CNNs hold significant promise for enhancing motor imagery-based BCIs. The findings contribute to the field by demonstrating the efficacy of CNNs in MI classification and providing insights into the application of the Low-Complexity Hilbert Transform for feature extraction.