Validation of a new data-driven automated algorithm for muscular activity detection in REM sleep behavior disorder

Background: Documentation of REM sleep without atonia is fundamental for REM sleep behavior disorder (RBD) diagnosis. The automated REM atonia index (RAI), Frandsen index (FRI) and Kempfner index (KEI) were proposed for this, but achieved moderate performances. New method: Using sleep data from 27 healthy controls (C), 29 RBD patients and 36 patients with periodic limb movement disorder (PLMD), we developed and validated a new automated data-driven method for identifying movements in chin and tibialis electromyographic (EMG) signals. A probabilistic model of atonia from REM sleep of controls was defined and movements identified as EMG areas having low likelihood of being atonia. The percentages of movements and the median inter-movement distance during REM and non-REM (NREM) sleep were used for distinguishing C, RBD and PLMD by combining three optimized classifiers in a 5-fold cross-validation scheme. Results: The proposed method achieved average overall validation accuracies of 70.8% and 61.9% when REM and NREM, and only REM features were used, respectively. After removing apnea and arousal-related movements, they were 64.2% and 59.8%, respectively. Comparison with existing method(s): The proposed method outperformed RAI, FRI and KEI in identifying RBD patients and in particular achieved higher accuracy and specificity for classifying RBD. Conclusions: The results show that i) the proposed method has higher performances than the previous ones in distinguishing C, RBD and PLMD patients, ii) removal of apnea and arousal-related movements is not required, and iii) RBD patients can be better identified when both REM and NREM muscular activities are considered.