Automatic Sleep Stage Classification using Convolutional Neural Networks with Long Short-Term Memory
The division of sleep into different stages using EEG signals is a commonplace practice in sleep laboratories and an indispensable tool for clinicians and researchers. Despite the advances in artificial intelligence, the sleep stage scoring process is in most cases still performed manually. As the scoring process is tedious and time-consuming, its automatization is desirable. In this study a convolutional neural network is trained to automatically extract features from raw 30-second epochs of the EEG, EMG and EOG. An extension of the network using long short-term memory is used to make a sleep stage prediction given the six preceding epochs. To validate the automatic feature extraction a comparison with a hand-crafted feature extraction approach using 37 features is made. The networks were trained using the first 50 records of the public CCSHS dataset and further validated on the public Sleep-EDFx dataset, the public UCD dataset, the private EMSA dataset, as well as records 50 to 100 of the CCSHS. Results show that the network is able to achieve state-of-the-art performance on the CCSHS (record 0-50, accuracy 89%, F1 81%). Furthermore, without retraining, the network successfully recognizes sleep stages on unseen data of a similar cohort (CCSHS record 50-100: accuracy 91%, F1 84%, Sleep-EDFx accuracy 81%, F1 72%, EMSA 83% F1 72%). Performance were consistently higher compared to the hand-crafted feature approach. The results demonstrate that automatic feature extraction on sleep data is possible and learns features similar to the ones described in the sleep scoring manual of the American Academy of Sleep Medicine.
Faculteit der Sociale Wetenschappen