0786 Association Between Machine Learning Actigraphy Model of RBD Probabilities with Incident Parkinson’s Disease and Dementia in Older UK Biobank Patients

Abstract Introduction Isolated rapid-eye-movement sleep behavior disorder (iRBD) is a strong predictor of Parkinson’s disease (PD) and related α-synucleinopathies. Our group and others have shown that wrist accelerometer data and machine learning methods can be used to detect RBD. We aimed to validate this approach in a large community cohort by assessing whether iRBD predictions were associated with an increased risk of incident PD or dementia. Methods We analyzed UK Biobank baseline actigraphy data from 26,912 participants older than 60 (mean age 63 ± 2.4 years; mean recording duration 6.7 ± 0.6 days; mean follow-up 8 ± 0.6 years) and applied an iRBD classifier using previously published features to output iRBD probability scores from 4 main, explainable, actigraphy-derived features previously published. A subset of 4,863 participants was used to determine RBD-probability thresholds, defining a “low-risk” (≤90th percentile) versus a “high-risk” (>90th percentile) category. Incident diagnoses were identified from ICD-10 codes, and risk ratios (RRs) with 95% CIs were estimated using log-transformed Wald methods. Results In the remaining, unseen cohort (n=22,049, 63.8 ± 2.8 years), the high-risk group had an RR of 2.28 (95% CI 1.50–3.45) for PD (27 cases/1,440 controls vs. 119 cases/14,603 controls for low-risk) and 2.48 (95% CI 1.55–3.99) for non-Alzheimer's disease (AD) dementia. Although not reaching statistical significance, high risk also predicted dual diagnoses of PD and AD (RR 1.84; 95% CI: 0.41–8.31), PD and non-AD dementia (2.30; 95% CI: 0.87–6.07), and dementia with Lewy body (RR 1.56; 95% CI: 0.35–6.90) which could be due to an insufficient number of incident cases. Conclusion Wearable-based 10% “highest risk” prediction of iRBD was associated with more than a two-fold increase in incident risk of PD and non-AD dementia, supporting its utility as a prodromal marker. These results lend support to scalable prodromal risk stratification for synucleinopathies using wearable devices on average 6-7 nights prior to diagnosis. Support (if any)