This item is available under a Creative Commons License for non-commercial use only
Electroencephalogram (EEG) can be used to record electrical potentials in the brain by attaching electrodes to the scalp. However, these low amplitude recordings are susceptible to noise which originates from several sources including ocular, pulse and muscle artefacts. Their presence has a severe impact on analysis and diagnoses of brain abnormalities. This research assessed the effectiveness of a stacked convolutional-recurrent auto-encoder (CR-AE) for noise reduction of EEG signal. Performance was evaluated using the signal-to-noise ratio (SNR) and peak signal-to-noise ratio (PSNR) in comparison to principal component analysis (PCA), independent component analysis (ICA) and a simple auto-encoder (AE). The Harrell-Davis quantile estimator was used to compare SNR and PSNR distributions of reconstructed and raw signals. It was found that the proposed CR-AE achieved a mean SNR of 5:53 db and signicantly increased the SNR across all quantiles for each channel compared to the state-of-the-art methods. However, though SNR increased PSNR did not and the proposed CR-AE was outperformed by each baseline across the majority of quantiles for all channels. In addition, though reconstruction error was very low none of the proposed CR-AE architectures could generalize to the second dataset.
Keegan, E. (2020). Stacked convolutional recurrent auto-encoder for noise reduction in EEG. Dissertation. Dublin: Technological University Dublin.