Mixture of experts for automated detection of phasic arousals in sleep signals

摘要

Detector design from training data generally relies on the fact that data labels are provided unambiguously by experts. However this optimistic situation is not always realistic, particularly when the phenomena to be detected are complex and poorly understood. The aim of this work is to propose a method for the mixture of experts which authorizes a posteriori analysis of each expert relevance. Our method consists in optimizing the normalized mutual information between the observations and a function of the labels provided by the pool of experts. The method is successfully applied to transient detection in physiological sleep signals.