A somnologist’s guide to explainable deep neural networks for sleep scoring

Abstract

Background: During the past decade, deep neural networks (DNNs)—a subgroup of machine learning (ML) methods—became popular in many facets of daily life and also found their way into somnology. Although they can support tasks such as sleep scoring, they have not yet found their way into clinical practice. One of the fundamental challenges lies in the fact that many of these models cannot explain the underlying reasons for their decisions, thus weakening users’ trust. Objective: As we expect DNNs for routine sleep medicine to become available soon, this guide introduces somnologists to the fundamentals of DNNs and to the current state of sleep scoring. In addition, we focus on how DNNs can be made more interpretable. Our aim is to foster understanding of this novel technology. Materials and methods: We used the publicly available Sleep Heart Health Study (SHHS) polysomnography (PSG) dataset for training a DNN for sleep scoring. Subsequently, we introduced the integrated gradients (IG) interpretability method and applied it to gain a better understanding of the DNN’s decision making. Results: The results of IG are visualized as pseudo-colors on the PSG data. We show exemplary findings for each sleep stage and compare them to the American Academy of Sleep Medicine (AASM) guideline for sleep scoring. Conclusion: Our model showed partial alignments with the AASM recommendations, e.g., highlighting K‑complexes during N2 sleep. However, interpretation is not straightforward, as the IG values do not follow standardized rules and only express a tendency of the decision-making process. Currently, only few studies focus on interpretability in DNN sleep scoring; however, our results underscore the need for efforts in this direction.