Towards Evaluating Robustness of EEG-based Sleep Trackers

Recent advancements in mobile health sensing coupled with the availability of large datasets have given rise to large scale adoption of wearable sleep trackers. Wearable sleep trackers enable continuous measurement of sleep in home settings, of which EEG-based trackers are more reliable since they directly measure brain activity. However, EEG-based trackers face challenges in deployment settings when exposed to real-world data which tends to be noisy or out-of-distribution (OOD). In such situations, prediction models may become overconfident, leading to unreliable and inaccurate results. In this study, we explore various scenarios in the deployment settings and measure their impact on the prediction performance of a single-channel based EEG model.