Towards Automating Sleep Stage Scoring to Diagnose Sleep Disorders

Automatic sleep stage scoring;Polysomnography;Likelihood ratio classifier;Decision tree;Biomedical Engineering

Gunnarsdottir, Kristin MariaGamaldo, Charlene E. ;

Johns Hopkins University

关键词: Automatic sleep stage scoring;    Polysomnography;    Likelihood ratio classifier;    Decision tree;    Biomedical Engineering;   

Others  :  https://jscholarship.library.jhu.edu/bitstream/handle/1774.2/39495/GUNNARSDOTTIR-THESIS-2016.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: JOHNS HOPKINS DSpace Repository

【 摘 要 】

Overnight polysomnography (PSG) is an important tool used to characterize sleep and the gold standard procedure for diagnosing many sleep disorders. PSG is a non-invasive procedure that collects various physiological data, such as EEG, EMG, EOG and ECG signals. The data is then scored in a subjective, laborious and time-consuming process by sleep specialists who assign a sleep stage to every 30-second window of the data according to predefined scoring rules by the American Academy of Sleep Medicine (AASM). Finally, clinicians make a diagnosis based on this annotated data. Consequently, the current process is heavily dependent upon human factors, which can result in poor agreement between expert scorers, but inter-scorer reliability has been found to be only around 82%. In this study we developed an automatic sleep stage scoring method, using a likelihood ratio decision tree classifier, with the goal of improving the speed, reliability, accuracy and cost efficiency of the current PSG scoring process. The algorithm was developed using the AASM Manual for Scoring Sleep. We extracted features from various physiological recordings of the PSG, based on the predefined rules of the AASM Manual. The features were computed for each 30-second epoch, in either the time or the frequency domain. The most useful features were selected by looking at probability distributions for each metric conditioned on the sleep stage, and identifying the features giving the greatest separation between stages. Examples of meaningful features include the power in different frequency bands of EEG signals, EMG energy per epoch, and number of spindles per epoch, to mention a few. These features were then used as inputs to the classifier which assigned each epoch one of five possible stages:; N3, N2, N1, REM or Wake.The automatic scoring was trained and tested on PSG data from 39 healthy individuals (age range: 24.2±3.1 years) with no sleep disturbances. The overall scoring accuracy was 76.97% on the test set. Some of the stages, such as stage N2, have more distinctive characteristics and thus yielded a higher per-stage scoring accuracy, whereas the other stages, for example stages N1 and REM, got confused more easily, resulting in lower per-stage accuracies. As expected, most misclassifications occurred between adjacent sleep stages. Although this accuracy may at first seem low, it is likely that the stages that the tool classified inaccurately may be sleep stages that contribute to inter-scorer reliability. Therefore, we see this tool as assisting sleep scorers to enhance efficiency with the further goal of eventually improving inter-scorer reliability.Sleep stage scoring provides an important basis for diagnosis of sleep disorders in general. However, the detection of sleep disturbances is very costly and time-consuming, and relies on subjective measures. Automating the scoring process improves the efficiency and consistency of scoring procedures and offers a way to diagnose sleeping disorders in a more robust, quantitative manner.