PerSleep: A Visual Analytics Approach for Performance Assessment of Sleep Staging Models

Machine learning is becoming increasingly popular in the medical domain. In the near future, clinicians expect predictive models to support daily tasks such as diagnosis and prognostic analysis. For this reason, it is utterly important to evaluate and compare the performance of such models so that clinicians can safely rely on them. In this paper, we focus on sleep staging wherein machine learning models can be used to automate or support sleep scoring. Evaluation of these models is complex because sleep is a natural process, which varies among patients. For adoption in clinical routine, it is important to understand how the models perform for different groups of patients. Moreover, models can be trained to recognize different characteristics in the data, and model developers need to understand why and how performance of the different models varies. To address these challenges, we present a visual analytics approach to evaluate the performance of predictive models on sleep staging and to help experts better understand these models with respect to patient data (e.g., conditions, medication, etc.). We illustrate the effectiveness of our approach by comparing multiple models trained on real-world sleep staging data with experts.