Ταξινόμηση των Σταδίων του Ύπνου από Βιοσήματα Φορητών Συσκευών με Χρήση Βαθιών Νευρωνικών Δικτύων

Abstract

Sleep has always been of great interest regarding the understanding of human nature and its existence. Although it has been studied since ancient times, its deeper understanding has only flourished in the last centuries. Studying an individual's sleep patterns can reveal crucial information for their general health and it can also indicate special physical conditions. Under this perspective, monitoring and analyzing sleep in real-time has been of major scientific interest, and the recent advancements in the field of machine learning have allowed the exploitation of a very powerful scientific area for this purpose. The development of sensor technology has also boosted the use of wearable devices for easy and accessible sleep monitoring, allowing an individual's health self-supervision or prevention of special conditions. The objective of this thesis is to study the problem of Sleep Stage Classification employing neural networks, given two datasets derived from wearable devices, which also contain manually transcribed sleep stage labels. The two datasets differ in their source of monitoring on their size, since the first one is significantly smaller than the second. In the first part of the experiments, a bidirectional-LSTM architecture is tested, trained on a set of already extracted features that are common. Τhe model trained on the first dataset achieves competitive performance, and the one trained on the second dataset, while not reaching as high values, is still good. The generalization of the model trained on the first dataset, utilizing unseen data of the second dataset shows poor performance. In the next part of the experiments, an automated feature extraction method is proposed for training the model end-to-end, by integrating a convolutional module on the bidirectional-LSTM architecture that takes as input the raw data. The performance for the second dataset is very promising, achieving prediction values close to those reached with the bidirectional-LSTM on the first dataset. Thus, the proposed convolutional architecture can sufficiently model the temporal relationships among the raw features. The first dataset, does not perform so well in this setup, indicating that the automated feature extraction method is not adequate for it. For the final part of the experiments a hierarchical architecture named SeqSleepNet is utilized, which was initially designed for the task of sleep stage classification with data derived from polysomnography. For the last dataset, the appropriate spectrograms are extracted for training the SeqSleepNet, however the model performs poorly. This observation suggests that the specific dataset does not align well with deeper architectures, but rather with shallower ones with carefully extracted features, either due to its small size, or due to its internal structure. The second dataset does not allow for spectrogram extraction, thus two modifications of the SeqSleepNet are proposed. Both receive the raw features of the second dataset and apply an automated feature selection with an attention mechanism, before passing the data to the classification module of the network. Their difference is that the second modification includes the concept of sleep epoch, where one sleep stage label is aligned to every 30-second window, adding one more dimension to the training input. The experimental results show that both of the proposed SeqSleepNet modifications achieve good performance, even though the second one handles more complex input data. Based on the experiments conducted in this thesis, we highlight the importance of a carefully chosen architecture and data handling for each dataset. We show that competitive results surpassing other recent work on the topic can be reached, either using meaningful extracted features or an automated method utilized in the shallower architecture of bidirectional-LSTM. The deeper architecture of SeqSleepNet is not appropriate for the two wearable-derived datasets.