DenoiseEEG

I worked on a semester-long research project on Signal Processing and DL under Prof. Lalan Kumar in his Multichannel Signal Processing(MSP) lab. In raw Electroencephalography (EEG) signals, unwanted and extraneous electrical signals pose significant challenges in any downstream task or research in clinical diagnosis, neuroscience, and human-computer interaction. To address this issue, I developed a novel, simple, end-to-end DL framework to effectively map noisy EEG signals to clean EEG signals. Traditional methods to denoise EEG signals, like Independent Component Analysis (ICA), are time-consuming and offline, making them infeasible for real-time applications.

I initially created a basic MLP architecture and gradually enhanced its complexity by incorporating 1-D Convolution layers, LSTM layers, and ResNet-inspired skip connections. Our final model achieved an impressive Pearson Correlation Coefficient (PCC) of 0.933. Unlike some existing DL approaches, I used actual EEG data from lift and grasp tasks, avoiding synthetic noise and ensuring authentic signal representation. I utilized ICA using the EEGLAB toolbox in MATLAB to generate the necessary noisy EEG-clean EEG pairs for training.