Research On Sleep Staging Method Based On Contrastive Learning

Sleep stage classification is an essential task in the diagnosis of obstructive sleep apnea hypopnea syndrome.Nowadays,the application of deep learning technology to the field of sleep stage classification has been increasingly embedded in researches.In sleep staging tasks,sleep staging models are generally trained through supervised learning methods,which require a great deal of labeled data.However,lack of the high-quality labeled poly-somnography data presents a challenge for analysis.The cumbersome and time-consuming process of labeling poly-somnography data requires doctors who feature relevant knowledge in sleep medicine.In this regard,this paper researchs the sleep staging method based on contrastive learning.Two methods of sleep stage classification are proposed,which are conductive to the performance of sleep stage classification with a small amount of labeled poly-somnography data.The specific contributions of this paper are listed as follows:1.This paper proposes a novel sleep staging model based on contrastive learning combined with attention mechanism.The unlabeled poly-somnography data are used for model’s unsupervised pre-training,which extracts the consistent features of unlabeled poly-somnography data through the encoder in the model.The parameters of the encoder are fixed after pre-training.Then the model fine-tunes the parameters of bidirectional long short-term memory network and the binary view attention module with a small amount of labeled poly-somnography data.Validated by experiments,the method based on contrastive learning combined with attention mechanism can effectively improve the performance in sleep staging classification.2.This paper proposes a novel sleep staging model based on contrastive learning combined with pseudo-label method.A small amount of labeled data are used for pretraining to obtain a classification network with appropriate accuracy.Meanwhile,the encoder parameters of the pre-trained network are used as the initializing parameters of the model encoder.With the pseudo-label method,the model train the unlabeled polysomnography data and labeled poly-somnography data end-to-end.The model separates the unlabeled poly-somnography data into the data with high low confidence and the data with low confidence,and uses the unlabeled data with high confidence to generate pseudo-label for training.For the unlabeled data with low confidence,the model performs contrastive learning method on them to improve the encoder’s ability to extract consistent features.Validated by experiments,the method based on contrastive learning combined with pseudo-label can effectively improve the performance in sleep staging classification.Lastly,the model is improved based on the concept of exponential moving average,experiments imply that the improved model has superior performance of sleep stage classification.