Towards a Hybrid Expert System Based on Sleep Event’s Threshold Dependencies for Automated Personalized Sleep Staging by Combining Symbolic Fusion and Differential Evolution Algorithm

摘要

Identification of sleep stages is a fundamental step in clinical sleep analysis. Existing automatic sleep staging systems ignore two major issues: 1) Most of existing automatic sleep staging systems are using numerical classification methods without involving medical knowledge. These kinds of systems are not yet understood and accepted by physicians. 2) Individual variability sources are ignored. However, individual variability is observed in many aspects of sleep research (such as polysomnography recordings, sleep patterns, and sleep architecture). In this paper, a hybrid expert system is proposed to mimic the decision-making process of clinical sleep staging in accordance with the medical knowledge by using symbolic fusion. To formalize the medical guideline and knowledge, thresholds are used for translating the sleep events into symbols and the sleep event's threshold dependencies are analyzed for fully understanding the thresholds dependencies among different sleep stages and subjects. Meanwhile, the differential evolution algorithm is adopted to automate the setting-up of thresholds that are used in the symbolic fusion model and to provide personalized thresholds, which allows taking the individual variability into consideration. The robustness and clinical applicability of the proposed system are evaluated and demonstrated on a clinical dataset. The dataset is composed of 16 patients (nine males and seven females) and scored by physicians. Only 5% of the dataset is used for the training process to obtain the personalized thresholds. Then, these personalized thresholds are passed to the classification process, and the overall accuracy on the identification of five sleep stages reaches 80.09%. Using a small dataset for the training process, the proposed system not only drastically reduces the training set but also achieves favorable results compared with most of the existing works.