Fig 1.
Illustration of the proposed sequence to sequence deep learning network architecture for automated sleep stage scoring.
The input signal is a sequence of 30-s EEG epochs and the outputs are their corresponding stages (or classes) generated by our proposed method.
Fig 2.
Detailed sketch of the utilized CNN model in the proposed work.
Fig 3.
A schematic diagram of the bidirectional recurrent neural network.
Table 1.
Details of number of sleep stages in each version of Sleep-EDF dataset.
Table 2.
Confusion matrix and per-class performance achieved by the proposed method using Fpz-Cz EEG channel of the EDF-Sleep-2013 database.
Table 3.
Confusion matrix and per-class performance achieved by the proposed method using Pz-Oz EEG channel of the EDF-Sleep-2013 database.
Table 4.
Comparison of performance obtained by our approach with other state-of-the-art algorithms.
Fig 4.
Graphs of the performance of the accuracy (a) and the loss function (b) of the proposed model in each epoch for a randomly selected fold (i.e., the fold 4).
Fig 5.
A example of hypnograms generated by the machine (i.e., the proposed method) and a sleep expert of a subject from the Sleep-EDF-13 dataset; approximately 85% coverage.
Fig 6.
Attention maps of two sequence inputs (EEG epochs) and their corresponding sleep stage scores provided by our proposed method.