Background and Aim: Monitoring physiological signals during sleep can have substantial impact on detecting temporary intrusion of wakefukness referred to as sleep arousals in order to improve the quality of sleep. To overcome the problems associated with the cubersome visual inspection of these events by sleep experts, automated sleep arousal recognition algorithms have been proposed.
Method: This study proposes a deep model ensemble neural network architecture for automatic arousal recognition from multi-modal sensor signals. Separate branches of the neural network extract features from electroencephalography, electrooculography, oxygen saturation level and the heart rate series (extracted from electrocardiography); and a final multi layer perceptron combines features computed from the signal sources to estimate the probability of arousal in each region of interest. Importantly, we investigate the use of shared-parameter Siamese architectures for unsupervised subject-dependent feature calibration. Namely, at each forward and backward pass through the network we concatenate to the input a user-specific template signal that is processed by an identical copy of the network. The latter is precomputed using both past and future information from the user's whole signal, and effectively implement feature-level personalisation in the end-to-end framework. We train the model end-to-end on one minute signal slices with 50% overlap, and heavily rely on time-series data-augmentation techniques to artificially rebalance class labels in order to tackle difficulties that arise in performing stochastic gradient descent in highly unbalanced dataset. Furthermore, we employ state-of-the-art regularisation and drop-out techniques to avoid overfitting and we rely on early-stopping criteria using AUROC and AUPR statistics computed from a validation set.
Result: The proposed multimodal, personalized end-to-end architecure obtains AUROC and AUPRC scores of 0.815 and 0.214, respectively on Phase I submission. Results on 10 fold cross-validation on the training set provided are whereas 0.849 and 0.287 respectively.