Sam Rusk, BS1 • Joseph Cataline, MD2 • Yoav N. Nygate, MS1 • Jiansan Gu, MD2 • Natalie Witek, MD2 • Chris R. Fernandez, MS1 • Nathaniel F. Watson, MD, MS3 • James Herdegen, MD2
Introduction
The prevalence of polysomnography (PSG) confirmed REM sleep behavior disorder (RBD) has been estimated at 0.68% of the general population, and that of probable RBD at 5.65%. Isolated RBD is considered a pre-clinical marker of neurodegeneration with strong predictive value. Large longitudinal cohort studies demonstrate 81–91% of idiopathic RBD patients, followed for ≥14-years, will develop a definite neurodegenerative disease or mild cognitive impairment. Furthermore, high prevalence of sleep apnea (SA) exists in patients with RBD/REM without atonia (RWA), which can confound analysis of EMG tone and arousal activity. We demonstrate predictive AI models for RBD/RWA in subjects with and without SA.
Methods
Patients with known RBD/RWA who underwent PSG were divided into sub-groups with (n=86, RBDSA) and without SA (n=19, RBDNSA). Patient demographics, PSG characteristics, and co-morbid conditions were reviewed. Random forest (RF) models of PSG indices and Deep Learning (DL) methods identified predictive characteristics of RBD/RWA, sensitivity and specificity for RBD/RWA detection were estimated, and feature importance analysis was performed.
Results
Subjects with RBDSA compared to RBDNSA demonstrated greater BMI, AHI, Arousals (ArI), and Periodic Limb Movements (PLMs). In a Parkinson’s disease subgroup (n=12, RBDPD) compared to non-Parkinson’s sub-group, sleep was more fragmented, showing higher WASO, ArI, and PLMs. RF modeling demonstrated sensitivity/specificity of 67%/96% for RBDNSA subjects, and 79%/52% for RBDSA. Feature importance analysis resulted in several top-features for RBDNSA: N3 time, TST, AHI, REM duration, and sleep latency, while different top-features were observed for RBDSA: ratio of leg events in REM vs. non-REM, TST, ratio of arousals in REM vs. non-REM, REM duration, and REM time.
Conclusions
AI approaches including RF models produced high specificity and moderate sensitivity for RBDNSA subjects. Observed specificity was lower for RBDSA patients. We hypothesize the performance difference attributable to similarity and additional complexity in sleep disturbance characteristics between RBD and SA. Broad implementation of AI methods show potential to expand early detection and diagnosis of RBD and associated neurodegenerative conditions through expanded analysis, and follow-up, including RWA spectrum patients that don’t meet clinical thresholds for diagnosable RBD.
