Knowledge-based decision system for automatic sleep staging using symbolic fusion in a turing machine-like decision process formalizing the sleep medicine guidelines

Automatic sleep staging is challenging since several issues need to be addressed. Traditional approaches from literature do not satisfy medical experts since they do not reflect the cognitive process they perform when scoring polysomnographic curves. We propose a new approach that is based on the implementation of medical knowledge by symbolic fusion. Medical knowledge coming from the international clinical practice guidelines for sleep medicine is formalized as a five-layer framework dedicated to data abstraction in order to deliver local and global propositions and support the interpretation of polysomnographic curves. Firstly, features are extracted from raw curves. Then these features are combined to recognize sleep events in accordance with guidelines. Sleep events are then fused into the criteria required to recognize the different sleep stages. Sleep is not homogeneous through the night. The physiological events observed during the night follow a dynamic that needs to be included into an automatic sleep staging system. In order to take this into account, decision rules are selected and applied to recognize a sleep stage according to the current context. Thereby, transitions are considered with interest. In this paper, we propose to use a Turing Machine-like decision process to handle transitions. To interpret the local observations and properly score a given state, the previous state which has been stored in a specific register is used as a context. One of the advantages of following the principles of symbolic fusion is to benefit from the full traceability of the decision. Hence, it makes possible to discuss each final - or intermediate - decision with an expert and check for relevance. The preliminary results are encouraging since agreement rates provided between decisions taken by our automatic approach and human experts are similar to those measured between human experts (average agreement rate = 54.60% / average Cohen's kappa = 0.40) on a dataset of 131 full polysomnographic recordings.