Interest in sleep has been growing in the last decades, considering its benefits for well-being, but also to diagnose sleep troubles. The gold standard to monitor sleep consists of recording the course of many physiological parameters during a whole night. The human interpretation of resulting curves is time consuming. We propose an automatic knowledge-based decision system to support sleep staging. This system handles temporal data, such as events, to combine and aggregate atomic data, so as to obtain high-abstraction-levels contextual decisions. The proposed system relies on a semantic reprentation of observations, and on contextual knowledge base obtained by formalizing clinical practice guidelines. Evaluated on a dataset composed of 131 full night polysomnographies, results are encouraging, but point out that further knowledge need to be integrated.