Empowering Contactless Sleep Health Monitoring with Multi-task Learning

Contactless sensing enables the simultaneous monitoring of multiple sleep indicators without disturbing the individual’s sleep. However, complex relationships among multiple tasks pose challenges for joint monitoring. In the paper, we develop a contactless sleep health monitoring system named SHMIU, which analyzes IR-UWB signals reflected off the human body and predicts sleep stages, sleep apnea state, and blood oxygen saturation during sleep. For enhancing the performance of three tasks, we design a mixture of experts based multi-task learning model and dynamically reweight tasks via Nash bargaining to optimize multi-task learning. We conduct experiments on a 100-subject dataset, and the experimental results show that SHMIU achieves balanced performance across three tasks, with its performance comparable to or superior to that of existing two-task methods.