The Feasibility Analysis Of Spindle Detection Based On Crowdsourcing

Sleep spindles is a sign of non-rapid eye movement sleep stage N2, its frequency range is from 11 HZ to 16 Hz, the generated time is longer than 0.5 second, the amplitude increases at first and then decreases, its shape is similar to the shuttle. In the research of the human memory and intelligence and in the clinical diagnosis of some mental diseases, sleep spindles play an important role. So far, despite the endless stream of spindles automatic detection algorithm, but manual detection based on the macroscopic observation has been the highest accuracy of the method, and the expert eye manual delineation can called the gold standard of the spindles, but generally it is difficult to find experts. Therefore, the paper win a large number of non-expert labeled spindle data set through crowdsourcing, then gets the standard of non-expert group, then compares the standard of non-expert group with the gold standard of the expert group, and see whether the standard of the non-expert group can replace on the gold standard of the expert group.Before we carry out this experiment, we need to collect and select some sleep EEG data used in the experiment. Before collecting the data, we need subjects to fill in a number of questionnaires in order to make a brief assessment of the mood, sleep and other situations of objects in the near future, so we develop the online sleep research questionnaire survey system. This system replaces the traditional paper questionnaire survey. On the one hand, it makes the scope of the function is wider and faster. On the other hand, it can save a lot of manpower, financial resources, material resources and events. Moreover, the result can be stored in the computer, so that it can be used at any time.Because the amount of a night of sleep EEG data is very large, and this study is to mark the sleep spindles through crowdsourcing, so the amount of objects that label spindles, has a lot, so eventually we need to deal with a great deal of experimental data. If marking the spindles by the offline availability of MATLAB, so data statistics and data processing will more complicated, so we develop online spindles marking system based WEB. The object can online mark the sleep spindles if he has a computer, and the result of marked sleep spindles will be stored remote database.By analysising the result of expert group and non-expert group through the online spindles marking system, we can get the following conclusions: For the expert group, when the group threshold T-group is 0.3 and the overlap threshold T-overlap is 0.45, the gold standard of expert group is the best. The average performance of all the gold standard of expert group is 0.84007±0.023(mean±variance), which indicates that each expert has good agreement with the gold standard of expert group. For the non-expert group, when the group T-group is 0.35 and the overlap threshold T-overlap is 0.3, the standard of the non-expert group is the best. The average performance of all non-expert and non-expert group is 0.7246±0.1008(mean±variance), which is obviously smaller than the average performance of the expert group. This shows that the consistency between the non-experts is not as high as the experts. The F1 value of the non-expert group and the gold standard of the expert group is 0.7557, which means that though the agreement of non-expert group below the agreement of expert group, the difference between the standard of the non-expert group and the gold standard of the expert group is acceptable. At the same time, the non-expert group and RMS automatic detection algorithm are compared, and the non-expert group is better than the RMS algorithm.