Actigraphy-based detection of isolated REM sleep behavior disorder: multicenter validation across devices and populations

Scalable home-based detection of isolated REM sleep behavior disorder (iRBD) is essential for early care, prevention trials, and identifying candidates for neuroprotective interventions against synucleinopathies. We previously showed that high-resolution wrist actigraphy (Axivity AX6) could identify iRBD based on abnormal sleep (AUC of 0.916) and rest-activity-rhythms (RAR, AUC of 0.856) using machine learning. Here, we aimed to assess generalizability across: (1) other actigraphs using lower resolutions; (2) different populations. We tested the analysis pipeline directly in cohorts from the International RBD Study Group using: Axivity AX6 (50-100 Hz), Philips Actiwatch (60-second epoch), and MicroMini-Motionlogger (30-second epoch). The cohorts included a total of 352 iRBD and 258 non-RBD participants from 4 centers (Mount Sinai, Oxford, Hong Kong, and Innsbruck). Two conversion pipelines were created to map activity counts from Actiwatch and MicroMini-Motionlogger to AX6 from 14 volunteers co-wearing two devices. In addition to the actigraphy analysis, four synucleinopathy prodromes-RBD symptoms, hyposmia, constipation, orthostatic hypotension-were tested in a two-stage screening approach. The sleep model achieved AUCs of 0.838-0.865 across centers, and the RAR model 0.520-0.818. Screening based on prodromes followed by actigraphy achieved sensitivities, specificities, and positive predictive values (PPVs) of 59.4-78.3%, 84.1-98.2%, and 56.0-98.6% (RBD symptoms), 46.5%, 99.0%, and 98.9% (hyposmia), 25.8-43.3%, 95.5-98.8%, 96.3-98.0% (constipation), and 11.6-36.8%, 96.0-100%, and 96.2-100% (orthostatic hypotension), respectively. Resolution (high versus low) did not affect the performance. After adjusting for a real-world iRBD prevalence of 1.5%, the corresponding PPVs would range from 6.3% to 100.0% depending on the prodromes. This multicenter study shows that the original actigraphy-based detection model of iRBD using sleep features but not RAR features generalizes well across independent cohorts and devices. Combined with key prodromes of synucleinopathies, it could enable precise, scalable population-level screening.