Sleep Structure Recognition For Small Amount Of Labeled Data

Sleep is closely related to human physical and psychological health.Sleep structure identification,also known as sleep staging,is an important basis for objectively assessing sleep quality and diagnosing diseases such as sleep disorders.Clinically,physicians use polysomnography to monitor sleep throughout the night.The objective analysis of sleep structure mainly relies on the technician’s interpretation of the collected EEG signals,with the EEG signals as auxiliary reference signals.The annotation of sleep data for a single person throughout the night usually takes several hours,and the annotation of sleep data for patients with sleep breathing disorders can take up even more time.In addition,manual annotation is subjective,and the agreement rate between multiple technicians is low.The privacy of medical data also contributes to the scarcity of annotated data to some extent.Automatic sleep structure recognition can be achieved with the help of new technologies such as artificial intelligence,but nowadays,sleep structure recognition based on fully supervised learning strategy follows the idea of being driven by a large amount of highquality labeled data,and all the training data of the model need to be labeled,and the small amount of labeled data leads to the problem of low performance of intelligent sleep structure recognition and difficulty for engineering scholars to carry out their work.In order to solve the problems brought by the small amount of labeled data to the sleep structure recognition task and to achieve medical-industrial cooperation,this paper applies semi-supervised learning strategy and self-supervised learning strategy to the field of sleep structure recognition,and the specific research work is as follows:(1)A resampling-based semi-supervised sleep structure recognition algorithm is proposed.In addition to a small amount of labeled data,the current sleep data also has the category imbalance problem,and the classification accuracy of the few-sample categories is low.Therefore,this paper introduces a two-branch network,consisting of original sampling branch and resampling balance branch,where the original sampling branch learns sleep data features and the resampling balance branch learns sleep stage features of the few-sample categories,and the model training learns both labeled and unlabeled data features.In the training set,the proportion of labeled data is 40%,and the unlabeled data features are learned by maximizing the consistency of the predicted output of the two enhanced data to filter the pseudo-label and improve the classification effect.Experiments were conducted on 2 public and 1 private datasets using single-channel EEG signals,in which the accuracy on the ST dataset of Sleep-EDFx was 0.807,the kappa coefficient was0.720,and the macro F1 score was 0.750.Better sleep structure recognition performance was achieved compared with existing semi-supervised sleep structure recognition methods,and the fully supervised sleep structure recognition with similar results.Ablation experiments were conducted on whether to resample or not,and the results demonstrated that increasing resampling branches helped the model to learn features of less sample categories.(2)A self-supervised sleep structure recognition algorithm based on cross-modal learning and representation reconstruction is proposed.To address the problems that data augmentation in self-supervised contrast learning tends to lose important information and the auxiliary task tends to make the model learn information irrelevant to sleep,this paper proposes a sleep structure recognition algorithm based on cross-modal learning and representation reconstruction.Specifically,positive sample pairs with large differences are constructed for mutual supervision using symmetric physiological signals? meanwhile,a representation reconstruction task is introduced to increase the mutual information between representations and inputs to approximate more information relevant to the sleep structure recognition task.The algorithm was evaluated on 2 public datasets and 1 private dataset using 5% labeled data for model fine-tuning,where the accuracy on Sleep-EDFx was 0.885 with a kappa coefficient of 0.789 and a macro F1 score of 0.711,achieving a better performance compared to fully supervised sleep structure recognition using the same percentage of labeled data,outperforming most existing self-supervised sleep structure recognition methods.In addition,this paper compares different data transformation methods and directly employs cross-modal data enhancement with left and right eye electrical signals and different EEG signals,which can bring better recognition results.In comparing other advanced methods,it is found that increasing the representation reconstruction loss can improve the recognition effect in non-rapid eye movement periods.This paper also conducts experiments on different scales of labeled data and visualizes the representations.The results show that the proposed algorithm can well extract features from unlabeled data and improve the sleep structure recognition accuracy.Finally,the self-supervised structure recognition algorithm is integrated into the toolbox,and the main sleep-related functions implemented are data reading and waveform display,sleep structure recognition,sleep stage occupancy and whole-night sleep structure map to assist clinicians in sleep quality assessment.In summary,for the limited supervised information brought by a small amount of labeled data,this paper proposes a semi-supervised sleep structure recognition algorithm based on resampling and a self-supervised sleep structure recognition algorithm based on cross-modal learning and representation reconstruction,which fully exploit the potential features of unlabeled data and can achieve similar or better performance than fully supervised sleep structure recognition,reducing label dependency for future objectified sleep structure recognition and providing a solid backbone for medical-industrial integration.