A time-frequency block structure approach to denoising sleep EEG

Automated sleep staging is an extensively researched problem with a spread of hand-crafted feature representations. Currently available representations, however, fail to produce sufficiently accurate results. Previously used features tend to struggle with artifacts and inter-patient variability. To address these issues, an aligned time-frequency block structure model was created. This model can be learned by building upon a combination of existing denoising and consensus clustering techniques. Across multiple datasets, this model significantly reduced error rates from raw spectral features and outperformed bandpower features commonly used in commercial tools. For the DREAMS dataset, classic band power features yielded a 30% error rate; raw spectral features had a higher error rate of 37%; and the novel Dense Denoised Spectral (DDS) features resulted in a 17% error rate.