REM sleep behavior disorder (RBD) is a parasomnia with dream enactment and presence of REM sleep without atonia (RSWA). RBD diagnosed manually via polysomnography (PSG) scoring, which is time intensive. Isolated RBD (iRBD) is also associated with a high probability of conversion to Parkinson's disease. Diagnosis of iRBD is largely based on clinical evaluation and subjective PSG ratings of REM sleep without atonia. Here we show the first application of a novel spectral vision transformer (SViT) to PSG signals for detection of RBD and compare the results to the more conventional convolutional neural network architecture. The vision-based deep learning models were applied to scalograms (30 or 300 s windows) of the PSG data (EEG, EMG and EOG) and the predictions interpreted. A total of 153 RBD (96 iRBD and 57 RBD with PD) and 190 controls were included in the study and 5-fold bagged ensemble was used. Model outputs were analyzed per-patient (averaged), with regards to sleep stage, and the SViT was interpreted using integrated gradients. Models had a similar per-epoch test F1 score. However, the vision transformer had the best per-patient performance, with an F1 score 0.87. Training the SViT on channel subsets, it achieved an F1 score of 0.93 on a combination of EEG and EOG. EMG is thought to have the highest diagnostic yield, but interpretation of our model showed that high relevance was placed on EEG and EOG, indicating these channels could be included for diagnosing RBD.

Obstructive sleep apnea (OSA) is characterized by decreased breathing events that occur through the night, with severity reported as the apnea-hypopnea index (AHI), which is associated with certain craniofacial features. In this study, we used data from 1366 patients collected as part of Stanford Technology Analytics and Genomics in Sleep (STAGES) across 11 US and Canadian sleep clinics and analyzed 3D craniofacial scans with the goal of predicting AHI, as measured using gold standard nocturnal polysomnography (PSG). First, the algorithm detects pre-specified landmarks on mesh objects and aligns scans in 3D space. Subsequently, 2D images and depth maps are generated by rendering and rotating scans by 45-degree increments. Resulting images were stacked as channels and used as input to multi-view convolutional neural networks, which were trained and validated in a supervised manner to predict AHI values derived from PSGs. The proposed model achieved a mean absolute error of 11.38 events/hour, a Pearson correlation coefficient of 0.4, and accuracy for predicting OSA of 67% using 10-fold cross-validation. The model improved further by adding patient demographics and variables from questionnaires. We also show that the model performed at the level of three sleep medicine specialists, who used clinical experience to predict AHI based on 3D scan displays. Finally, we created topographic displays of the most important facial features used by the model to predict AHI, showing importance of the neck and chin area. The proposed algorithm has potential to serve as an inexpensive and efficient screening tool for individuals with suspected OSA.

Neural networks are becoming more and more popular for the analysis of physiological time-series. The most successful deep learning systems in this domain combine convolutional and recurrent layers to extract useful features to model temporal relations. Unfortunately, these recurrent models are difficult to tune and optimize. In our experience, they often require task-specific modifications, which makes them challenging to use for non-experts. We propose U-Time, a fully feed-forward deep learning approach to physiological time series segmentation developed for the analysis of sleep data. U-Time is a temporal fully convolutional network based on the U-Net architecture that was originally proposed for image segmentation. U-Time maps sequential inputs of arbitrary length to sequences of class labels on a freely chosen temporal scale. This is done by implicitly classifying every individual time-point of the input signal and aggregating these classifications over fixed intervals to form the final predictions. We evaluated U-Time for sleep stage classification on a large collection of sleep electroencephalography (EEG) datasets. In all cases, we found that U-Time reaches or outperforms current state-of-the-art deep learning models while being much more robust in the training process and without requiring architecture or hyperparameter adaptation across tasks.