Human sleep can be described as a series of transitions between distinct states. This makes automatic sleep analysis (scoring) suitable for an automatic implementation using machine learning. However, the task becomes harder when data is sampled using more light-weight or mobile equipment, often chosen due to greater comfort for the patient. In this study we investigate the improvement in sleep scoring when multiple state-of-the-art neural networks are joined into an ensemble, and subsequently distilled into a single model of identical network architecture, but with improved predictive performance. In this study we investigate ensembles of up to 10 networks, and show that, on the same data, ensembles of neural networks perform better than each single subject model (improvement: 2.4%) and that this improvement can be transferred back into a single network using a combination of patient specific data and knowledge distillation.

The study demonstrates both a way to further improve automatic sleep scoring from mobile devices, which in itself is interesting, but also highlights the great potential of the vast amounts of unlabeled personal data which will become available from personal recording devices.