AI Models Push Wrist Actigraphy Beyond RBD Screening to Flag Parkinson's Itself

The Swiss team behind an actigraphy tool for spotting REM sleep behavior disorder has trained deep learning models on the same wrist recordings to separate Parkinson's disease patients from controls, and to test whether the same digital signature shows up early in isolated RBD.

Researchers at Bern University Hospital in Switzerland report that a pretrained transformer model applied to wrist actigraphy recordings distinguished patients with Parkinson's disease from non-neurodegenerative controls with an area under the curve of 0.937 (sensitivity 80.5 percent, specificity 92.9 percent), outperforming a convolutional neural network (AUC 0.863) and a standard feature-based machine learning model (AUC 0.840). The findings were published online July 7, 2026 in the Journal of Sleep Research.

The work extends an actigraphy screening approach the same Bern sleep-medicine group described in Annals of Neurology in 2022, which used movement patterns captured by a wrist-worn sensor to flag REM sleep behavior disorder (RBD) in Parkinson's patients, reaching accuracy above 90 percent in a small in-clinic sample. In the new study, the same transformer model separated Parkinson's patients with RBD from those without RBD (AUC 0.956, sensitivity 84.4 percent, specificity 92.9 percent), and the researchers then ran it on six patients with isolated RBD, a condition considered an early stage of alpha-synucleinopathies including Parkinson's disease and dementia with Lewy bodies. On both classification tasks, the isolated RBD patients' scores fell between those of the comparison groups, according to the paper.

The cohort was small and drawn from a single center: 45 Parkinson's patients with RBD, 42 without RBD, 47 controls with restless legs syndrome, and only six people with isolated RBD were included. The study compared existing groups at a single point in time; it did not follow the isolated RBD patients forward to see whether their actigraphy score predicted an eventual Parkinson's or Lewy body diagnosis. The authors describe the isolated RBD result as an early signal pointing toward possible future use in tracking prodromal disease, not a validated predictive test.

The authors say the models would need testing in larger, longitudinal cohorts of isolated RBD patients, followed over time, before an actigraphy-based AI score could be used to gauge an individual's risk of progressing to Parkinson's disease or a related condition.