StageGuard Improves Sleep Staging by Enforcing Physiological Constraints

Researchers have developed StageGuard, an innovative structured-inference framework designed to improve the accuracy and physiological plausibility of automated sleep staging. While deep learning models have achieved high epoch-level accuracy in sleep staging, they often generate hypnograms that violate known physiological invariants, such as rare transitions between sleep stages or excessive fragmentation. These violations can introduce biases into critical sleep architecture metrics, regardless of overall accuracy. StageGuard acts as a plug-and-play wrapper for any neural sleep-staging backbone, infusing it with physiology-informed priors. It employs two main mechanisms: a differentiable soft transition penalty during training to discourage physiologically rare transitions, and a semi-Markov constrained decoder during inference. This decoder uses a duration-augmented state space to jointly enforce transition penalties and minimum bout durations. Crucially, StageGuard allows for rare transitions when strong emission evidence exists, preserving the ability to detect informative pathological events rather than strictly prohibiting them. Evaluations across six different deep learning backbones and four datasets demonstrate StageGuard's effectiveness. It significantly reduces the transition-violation rate (TVR) to physiologically realistic levels and lowers the fragmentation index (FI) by 56-62%, all while maintaining or slightly improving classification accuracy. The framework's ability to satisfy physiological constraints also translates into a 59-79% reduction in error for derived sleep-architecture statistics and more faithfully recovers expert-defined subgroup differences, such as those related to OSA severity and age.