Sleep Spindle Detection Using Deep Learning

Sleep spindles are significant transient oscillations observed on the electroencephalogram from brain scalp in stage 2 of non-rapid eye movement sleep, which are caused by the neural activity with the oscillation from 12 to 14 Hz, according to the standard principle released by the American Academy of Sleep Medicine. In the past, neurophysiologists have an affinity to visual identification by hand which is a laborious and time consuming task. Moreover,there are different results within neurophysiologists on same data. On the other hand, deep learning bear on some really challenging problems and have gained a great successes on pattern classification, especially in images and speech.In this paper, we make a research on detecting sleep spindle using deep learning. The training datasets are collected from different subjects and scored by the method of crowsourcing.As a new technology to score sleep spindle, beside for making an comparison of deep learning to traditional method, we also get results on the performance of different deep learning architectures.The first experiment, crowdsourcing replacing gold standard was applied, which is feasible and reliable, to obtain training datasets. Meanwhile, the sleep alpha waves are easily regarded as sleep spindle for the similarity of their attribute of frequency. In the second experiment,The datasets are transformed as input samples by power spectrum density. Based on four different power spectrum density features, deep belief network get the greater performance as a classifier than k nearest neighbour, decision trees, and get the better result than support vector machine. Then, two different features between raw and extracted feature are used to study the the effect of feature on deep belief network, with the former one gain 10 percent larger than latter one. Finally, we investigate the possibility of applying the deep belief network trained in the dataset to raw electroencephalogram activity, which is comparable to neurophysiologist's method.There are three deep learning architectures discussed in this paper, multi layer perception,stacked denoising auto-encoder, and deep belief network. For the third experiment, the performance of three different classifiers are estimated by scoring sleep spindles which are collected by same method consisting of 1 second length data. The deep belief network perform worse than multi layer perception which is not consistent to other papers, but the stacked denoising auto-encoder shows a good result, with 1 percent better than multi layer perception. Therefore, the stacked denoising auto-encoder are more competent than others on the study of sleep spindle.As crowdsourcing is a new method to get datasets that have an great impact on performance of classifier, further research is still needed to validate the reliable on these matters. In addition, a larger number of subjects and recordings should be gained to estimate the deep learning when it may be used on clinical monitoring and data mining task. There is no double that deep learning gain a unbelievable result of F1-score on classifying spindle as a new architecture, with the ability of self learning. However, except the three deep learning architectures, is there an other deep learning architecture which is more appropriate to sleep spindle? On the other hand, the effect of different feature on the performance of classifier is also an interesting and valuable area. Finally, the paper use the python library of theano and Tesla cuda of Nvidia to code and speed up the performance of algorithm, but seeking new parallel method, to improve the speed and efficiency of parallel, is still the challenge task of the study.