Validation of a new data-driven method for identification of muscular activity in REM sleep behaviour disorder

Objectives/Introduction:REM sleep behaviour disorder (RBD) isa parasomnia characterized by lack of atonia during REM sleep. Gold standard methods for RBD diagnosis require manual REM sleep without atonia (RSWA) scoring, which is time‐consuming and subjective. We propose and validate a new data‐driven algorithm and com-pare it to other automatic methods based on RSWA detection for identifying RBD patients.Methods:We included 27 control subjects (C), 29 idiopathic RBDpatients and 36 patients with periodic limb movement disorder(PLMD).

After artefact removal, mean absolute amplitude values of1‐s windows of chin, tibialis left and right EMG signals during REMfrom 9 randomly selected controls were used to define a probabilistic model delineating atonia. For the remaining subjects, each 1‐s window was labelled as movementif its probability of being atonia was lower than an optimized threshold.

For each EMG signal, we calculated the percentages of 1‐s windows with movements and themedian intra‐movement distance during REM and NREM. Usingthese indices, a classification algorithm was trained and tested (5‐fold cross‐validation) to distinguish the three subject groups. For comparison, the REM atonia index (RAI), Frandsen index (FRI) and Kempfner index (KEI) were calculated for the same cohort and ananalogous classification algorithm was applied to each of them.

The overall test accuracies, sensitivities and specificities for C, RBD and PLMD were calculated for each method.Results:The following test performances were achieved (mean and standard deviation across the five folds in %): Overall accuracy:79.58±9.16 (this work), 44.56±6.27 (RAI) 46.73±5.40 (FRI),49.08±11.82 (KEI); RBD sensitivity: 81.67±17.08 (this work),53.67±13.03 (RAI), 58.71±10.94 (FRI), 58.81±28.37 (KEI); RBD speci-ficity: 83.98±5.09 (this method), 83.59±4.12 (RAI), 85.48±4.18 (FRI),77.80±5.54 (KEI).

Further, the proposed method achieved higher sen-sitivity and specificity for identifying C and PLMD than the other ones.Conclusions:The proposed data‐driven method outperforms other automatic methods in distinguishing C, RBD and PLMD subjects andis more sensitive for RBD detection. Compared to the other meth-ods based only on RSWA detection, this method uses also NREM muscular activity to characterize patients groups.

The obtained high performances thus confirm previous findings of increased NREM muscular activity in RBD patients.