Personalized Sleep Monitoring Using Smartphones and Semi-supervised Learning

Sleep is a critical aspect of an individual’s physical and mental well-being. Hence a large body of sleep monitoring solutions has been gaining popularity, including data-driven AI techniques with mHealth adaptations of wearable, smartphone, and contactless-sensing modalities. Regardless, proposed solutions by prior works, by and large, require gathering sufficient ground truth data to develop personalize and highly accurate sleep prediction models. This requirement inherently presents a challenge of such models underperforming when inferring sleep on new users without labeled data. However, unlabeled data is often more abundantly available in a real-world application than gathering labeled data. In this paper, we propose SleepLess, which uses semi-supervised learning over unlabeled data sensed from the user’s smartphone network activity to develop personalized models and detect their sleep duration for the night. Specifically, it uses a pre-trained model on an existing set of users to produce pseudo labels for unlabeled data of a new user and achieves personalization by fine-tuning over selectively picking the pseudo labels. Our IRB-approved user study among 23 users found SleepLess model yielding around 96% accuracy, between 12–27 min of sleep time error and 18–25 min of wake time error. Comparison against other approaches that sought to predict with fewer labeled data found SleepLess, similarly yielding best performance. Our study demonstrates the feasibility of achieving personalization in sleep prediction models by utilizing unlabeled data extracted from network activity of users’ smartphones, through the application of a semi-supervised transfer learning approach.