Actigraphy, the simple wrist-worn monitoring of rest-activity cycles, combined with a specialized AI transformer model can distinguish Parkinson’s disease patients from healthy controls with over 93% accuracy and identify those with concurrent REM sleep behavior disorder with over 95% accuracy, according to a study published July 7 in the Journal of Sleep Research.
The findings suggest that the same non-invasive approach may detect early neurodegenerative changes in people with isolated REM sleep behavior disorder (iRBD), a known precursor to Parkinson’s disease and related synucleinopathies, years before motor symptoms become clinically apparent.
What they found
Researchers at Inselspital, Bern University Hospital in Switzerland trained and compared three AI architectures on actigraphy data from patients with Parkinson’s disease (PD), PD with and without RBD, isolated RBD, and non-neurodegenerative controls.
The Pre-trained Actigraphy Transformer (PAT), a transformer model fine-tuned on actigraphy time-series data, outperformed both a convolutional neural network and traditional machine learning approaches:
| Classification Task | AUC | Sensitivity | Specificity |
|—|—|—|—|
| PD vs. controls | 0.937 | 80.5% | 92.9% |
| PD-RBD vs. PD-noRBD | 0.956 | 84.4% | 92.9% |
The CNN achieved an AUC of 0.863 for PD versus controls, while conventional machine learning reached 0.840.
Critically, when the PAT model was tested on patients with isolated RBD, individuals who have the sleep disorder but no clinical signs of Parkinson’s. Their model scores fell squarely between those of the control and PD groups. This intermediate positioning suggests the AI is capturing subtle neurodegenerative changes that precede motor diagnosis.
PD-RBD patients also showed significantly altered actigraphy features compared to both PD patients without RBD and controls, indicating that rest-activity fragmentation carries a distinct signature beyond what standard clinical assessment detects.
Why it matters
Current diagnosis of Parkinson’s disease relies on clinical motor symptoms, which appear only after substantial dopaminergic neuron loss has already occurred. Objective, scalable biomarkers for prodromal detection are among the highest priorities in movement-disorder research.
Actigraphy offers a practical advantage: wrist-worn devices are already widely used in sleep medicine and consumer wearables. Adding an AI analysis layer could extend existing infrastructure into a screening tool for neurodegenerative disease, no specialized equipment, no radioactive tracers, no hospital visits required for the initial triage.
For the estimated 1% of the general population over age 60 with iRBD, the vast majority of whom will go on to develop a synucleinopathy, a low-cost digital biomarker could help identify who is at imminent risk and stratify candidates for neuroprotective trials.
Limits
The study was cross-sectional. The authors explicitly note that validation in longitudinal cohorts is necessary before the model can be used for individual risk prediction. It remains unknown how well the PAT model’s intermediate scores in iRBD patients correlate with actual progression time to phenoconversion. The study also did not address potential confounders such as medication effects on rest-activity patterns or comorbid sleep disorders beyond RBD.
Bottom line
AI-enhanced actigraphy achieves diagnostic performance comparable to more expensive and invasive biomarker modalities for Parkinson’s disease and its prodromal stage. If validated prospectively, it could become the first truly scalable, non-invasive screening tool for population-level neurodegenerative risk assessment, starting from a device many people already wear to bed.
Source
Lopes L, Warncke JD, Filchenko I, Shi K, Bassetti CLA, Schäfer C. “Actigraphy meets AI: A digital biomarker for Parkinson’s disease and isolated REM sleep behaviour disorder.” Journal of Sleep Research. 2026 Jul 7:e70396. DOI: 10.1111/jsr.70396