Personalized Sleep Spindle Detection in Whole Night Polysomnography

The polysomnography is an overnight exam that monitors many different bio-physiological signals during sleep. A correct evaluation of the polysomnography recordings is essential to identify possible sleep disorders. Still nowadays the evaluation procedure, called sleep scoring, is carried out manually by a sleep physician or by a technician, who visually inspect the polysomnography traces of the whole night. It can be a quite tedious and time-consuming activity. Many automatic sleep scorings systems are available, but the vast majority remain unused. According to the international standards the detection of some well-defined patterns (sleep spindles, k-complexes, vertex waves…) is needed to score the PSG and to assign at each part of the recording the correct sleep phase. In the present work I focus my research on sleep spindle detection. My new personalized sleep spindle detection algorithm highlights the importance of an invidualized approach. After identifying an optimal set of features that characterizes the spindle, a support vector machine is exploited to distinguish between spindle and non-spindle patterns. The algorithm is assessed on the open source DREAMS database , and on whole night polysomnography recordings from the SPASH database. On the former database the personalization can boost sensitivity significantly, from 66.2% to 80.6%, with a slight decrease in specificity, from 98.6 to 94.3%. On a whole night polysomnography instead, the algorithm reached a sensitivity of 97.5% and a specificity of 97,9%. This study project is integrated into European project E!SPAS in collaboration with two European companies and the Inselspital of Bern. My algorithm could be integrated in an automated scoring system, with a mutual benefit. A correct sleep spindle identification could assist a scoring algorithm and the scoring results could optimize the spindle detection, e.g. reducing the false positive rate.