Automatic Sleep Arousal Identification From Physiological Waveforms Using Deep Learning

The 2018 PhysioNet Computing in Cardiology Challenge focused on diagnosing sleep disorders, motivated by enabling treatment to alleviate the associated mental and physical health consequences. The dataset consists of 1,985 patients monitored at an MGH sleep laboratory where vital signs were recorded, and arousal regions were annotated by experts. This work presents a deep-learning method to identify sleep arousals. In traditional machine learning, feature extraction is one of the most time-intensive considerations, requiring a great deal of domain expertise and experimentation. In contrast, deep learning techniques automatically learn variable interactions between pairs or groups of signals, and any relevant temporal dependencies. This allows such algorithms to automatically extract sleep patterns from rich physiological time series. The model presented here integrates ideas from several successful deep learning models to construct a multi-channel time-series convolutional-deconvolutional neural network. This network was trained using cross-entropy loss, and evaluated on a 20% held-out validation set. Hyper-parameters were selected on the AUPRC metric, and training utilized early stopping to prevent over-fitting. The resultant model achieved an AUPRC of 0.369 and an AUROC of 0.855 on the final competition test set.