MobileSleepNet: A Deployment-Oriented Framework for Sleep Staging on Mobile Devices

Accurate sleep staging is essential for both clinical diagnosis and long-term sleep monitoring. However, most existing deep learning approaches rely on large-scale architectures and server-based inference, which limits their applicability in mobile and privacy-sensitive scenarios. This study evaluates whether clinically competitive sleep staging can be achieved entirely on-device under strict computational and energy constraints, using only single-channel electroencephalogram (EEG). Rather than increasing model complexity, we adopt a deployment-oriented design perspective and present an ultra-lightweight inference framework, MobileSleepNet, developed for mobile hardware. The framework combines a compact convolutional backbone for efficient feature extraction with a neural-parameter-free hybrid temporal refinement strategy, where local probability smoothing and statistical state-transition priors are used to enforce physiological consistency without introducing recurrent or attention-based models. This decoupled design enables effective temporal modeling while a small model size and low computational cost. Performance evaluations on four public Polysomnograph (PSG) datasets (Sleep-EDF-20/78 and ISRUC-S1/S3) show that the proposed approach achieves accuracy comparable to models that are orders of magnitude larger. Hardware-level profiling further demonstrates low inference latency and energy consumption natively on modern mobile processors. Offline mobile evaluation using a custom smartphone application confirms that on-device sleep staging can be executed locally from pre-recorded data, without cloud dependency. These results indicate that laboratory-grade sleep analysis can be feasibly translated to mobile devices, supporting privacy-preserving and energy-efficient sleep analysis at the edge. The source codes and video demonstration are available at https://github.com/dongdongzhou1017/MobileSleepNet.