AI techniques for K-complex detection in human sleep EEGs

Two knowledge-based approaches for the detection of K-complexes in human sleep EEGs (electroencephalograms) are compared. In the first approach the waveforms to be recognized are represented in the form of frames that capture the morphological and spatiotemporal characteristics of each object. This system has been equipped with an interface that guides inexperienced users with the definition of the frames and provides output to a signal analysis expert. It allows full flexibility to the user, including the definition of objects to be detected and features to be determined. The second approach is a stand-alone system for the (real-time) analysis of polysomnographic recordings. Data acquisition and preprocessing takes place in real time. In the preprocessing stage, the signal is divided into short, consecutive intervals, and a set of predefined features is extracted from each interval. The features describe so-called pattern candidates relevant to sleep classification. A multilevel approach is followed to weed out the false alarms. The user can modify certain of the thresholds used in the system and has some control over the reasoning strategy followed