Semi-Supervised Approach For Sleep Scoring Using Wavelet Transformed EEG Data

Sammanfattning

Historically, sleep stage classification has relied on manual scoring performed by experienced professionals. In recent years, the emergence of machine learning has enabled the automation of this process, thus saving both time and resources. Many of these sleep staging models depend on data which already has been annotated by a professional, but a few try to exploit the vast amount of unlabelled data that is readily available. This project focuses on just that, to utilize unlabelled electroencephalography (EEG) data in conjunction with labelled data to create a machine learning model which is able to classify sleep stages. Our proposed model consist of three steps: pretraining, fine-tuning and classification. During pretraining, an autoencoder was trained on 450 unlabelled EEG recordings, provided by Kvikna, to learn general representation of the input signals through reconstruction. In the fine-tuning phase, the model shifted from reconstructing the signal to extracting meaningful features, using labelled data from Sleep-EDF to guide this transition. Lastly, a classifier was trained on top of those features to learn how to score sleep stages. The final model reached an accuracy of 81.2%, which is in line with the performance of experienced scorers based on previous studies. This shows that the model could realistically be used as a substitute for manual scoring, saving both time and resources, and with potential applications in both research and clinical settings.