Derivation and modeling of two new features for the characterization of rapid and slow eye movements in electrooculographic sleep recordings

This work presents a method for the derivation of two new features characterizing the occurrence of both, saccadic and slow eye movements (SEM), in electrooculographic (EOG) sleep recordings. Analysis of EOG activity is of fundamental importance for the clinical interpretation of a subject's sleep pattern. The features here presented are derived from purely horizontal EOG recordings, and have been built to be patient-adaptive and relatively robust against a variety of artifacts. Using the two derived features, performance analysis of two derived Bayes classifiers (respectively for the automatic detection of saccades and of SEM) was validated. Experiments were carried out using a database of 21 whole-night recordings. Automatic and human detections were obtained on a 30-s time grid. Two clinical experts were used as the standard reference. Average kappa indexes were obtained to characterize the agreement between this reference and the automatic detector. Automatic-reference and human-human REM agreements were 0.80 and 0.87, respectively, for the detection of saccades. Corresponding SEM agreements were 0.59 and 0.64, respectively. Our results closely match the expected inter-rater agreement and therefore support the robustness of the method and the validity of the implemented features for the automatic analysis of sleep EOG recordings.