[Discrimination of Chin Electromyography in REM Sleep Behavior Disorder Using Deep Learning]
Fumiya Kinoshita1, Meiho Nakayama2, Hiroki Takada3
1Graduate School of Engineering, Toyama Prefectural University.
Objective:
The confirmation of abnormal behavior during video monitoring in polysomnography (PSG) and the frequency of rapid eye movement (REM) sleep without atonia (RWA) during REM sleep based on physiological indicators are essential diagnostic criteria for the diagnosis of REM sleep behavior disorder (RBD). However, no clear criteria have been established for the determination of the tonic and phasic activities of RWA. In this study, we investigated an RWA decision program that simulates visual inspection by clinical laboratory technicians.
Methods:
We used the measurement data of 25 men and women (average age±standard deviation: 72.7±1.7 years) who visited the Sleep Treatment Center for PSG inspection due to suspected RBD. The chin electromyography (EMG) during REM sleep was divided into 30 s intervals, and RWA decisions were made on the basis of visual inspection by a clinical laboratory technician. We compared and investigated two machine-learning methods namely support vector machine (SVM) and convolutional neural network (CNN) for RWA decisions.
Results:
When comparing SVM and CNN, the highest discrimination accuracy for RWA decisions was obtained when using the average rectified value (ARV) processed chin EMG images using CNN as a feature. We also estimated the prevalence of RBD on the basis of the Mahalanobis distance measure using the frequency of occurrence of both tonic and phasic activities calculated from a total of 25 subjects in the patient and healthy groups. Consequently, estimation of RBD prevalence using CNN resulted in misclassification of none of the subjects in the patient group and two subjects in the healthy group.
Conclusions:
In this study, we investigated the automatic analysis of PSG results focusing on RBD, which is a parasomnia. As a result, there were no misclassifications of patients in the 25 subjects in the patient or healthy groups based on the estimates of RBD prevalence using CNN. The prevalence estimation based on our proposed automated algorithm is considered effective for the primary screening for RBD.
