Enhancing Sleep Staging's interpretability via Frequency-Dropout

Automated sleep staging is crucial for efficient sleep disorder diagnosis, but existing deep learning models lack interpretability and struggle to leverage domain-specific knowledge. We propose Frequency-Dropout, a novel regularization technique that enhances model interpretability and performance by controlling useful information flow in the frequency domain. Frequency-Dropout integrates sleep priors into deep networks by applying Discrete Cosine Transform (DCT) to feature maps, followed by targeted dropout of high-frequency noise components. Evaluated on the SleepEDF-20 dataset, Frequency-Dropout boosted single branch CNN network's accuracy by 2.01% (to 81.42%).