Harnessing Cardio-respiratory Sleep Staging under Uncertainty

Automatic sleep stage classification from cardio-respiratory signals has emerged as a promising alternative to traditional polysomnography, which typically uses an extensive set of sensors including electrodes attached to the scalp. Despite impressive results to date, we argue that to harness the benefits of cardio-respiratory sleep staging, we require a greater focus on building models with calibrated uncertainty quantification. We describe how such models could enable important applications in sleep medicine, without necessarily requiring expert-level accuracy as measured by conventional metrics. Our work motivates further investigation into better-calibrated sleep staging models, to enable these applications.