Multi-modal sleep staging in the clinic for REM sleep behaviour disorder
- 1Nuffield Department of Clinical Neurosciences and Institute of Biomedical Engineering, University of Oxford, Oxford, UK
Abstract
Accurate REM identification is critical for diagnosing REM sleep behaviour disorder (RBD), yet many automated sleep-staging systems, especially single-channel EEG models trained on healthy cohorts, do not generalise well to real-life polysomnography (PSG) performed in patients. We compared a feature-based Random Forest (RF) model tuned for RBD with a state-of-the-art single- EEG deep architecture (AttnSleep), and assessed the impact of cohort adaptation and multimodal inputs (EEG, EOG, EMG, ECG). Experiments used 89 multi-site in-clinic PSGs (SleepWearables Phase-1) plus 53 MASS healthy controls (mean age 63 {+/-} 5 years), with 10-fold cross-validation and out-of-fold evaluation. When applied out-of-the-box after training on open-source healthy datasets, the RF outperformed AttnSleep on the clinical cohort (Cohens{kappa} = 0.56 versus 0.49), with the largest difference in REM sleep in RBD subjects (RF Cohens{kappa} = 0.51 versus AttnSleep Cohens{kappa} = 0.18), highlighting limited cross-cohort generalisation. Cohort-specific retraining substantially improved AttnSleep, and the multi-modal model further improved overall agreement (Cohens{kappa} 0.61 versus 0.63), with the largest stage-level gains in N1 and N3; in RBD subjects, however, overall agreement was unchanged (Cohens{kappa} 0.46 for both configurations). Attention-based modality analysis identified EEG as the dominant signal, increased EOG contribution during REM and N1, and elevated ECG importance during N3. In RBD subjects, EOG weighting increased relative to subjects without RBD ({Delta} = +0.096) while EEG weighting fell ({Delta} = -0.056), suggesting physiological relevance. Guided by these weights, a reduced four-channel EEG model matched full multimodal performance in non-RBD subjects, and adding EOG matched it overall (Cohens{kappa} = 0.63) while achieving the best performance in RBD (Cohens{kappa} = 0.49) and improved REM classification in RBD (60.6% compared to 53.0% recall). Inclusion of EOG also halved inter-dataset variability in REM staging between the clinical and MASS cohorts. Nonetheless, staging performance in RBD remained lower than in controls, particularly for REM, and within the RBD subgroup only the REM improvement from the four-channel EEG + EOG configuration reached statistical significance at this sample size (n = 22). These results highlight (1) the limited generalisability of minimal-sensor models trained on healthy cohorts, (2) the value of mixed cohort-specific training, and (3) that the composition of the sensor set matters more than its size: targeted addition of EOG improved REM staging in RBD specifically, whereas full multimodal integration improved performance mainly outside the RBD subgroup.
Related
- The protocol for a phase 2A, multi centre, double-blind, randomized, placebo-controlled, parallel-group study to assess the effect of 12 weeks treatment with oral PXS-4728A on microglia activation, as measured by positron emission tomography, in participants with isolated rapid eye movement sleep behaviour disorder
- Athlete-AI: a multi-modal artificial intelligence framework for personalized athlete monitoring and injury risk forecasting
- Use of multi-modal artificial intelligence model for prediction of atrial fibrillation in hospitalized patients
- Multi-modal data Integration for Metastatic Risk Prediction of AJCC Stage I/II Melanomas: Results from the Population-based Dutch Early-stage Melanoma Study (D-ESMEL)
- Multi-modal AI: Integrating electrocardiography with molecular biomarkers for noninvasive detection of cardiac allograft rejection
BibTeX
@misc{gunter2026multimodal,
title = {{Multi-modal sleep staging in the clinic for REM sleep behaviour disorder}},
author = {K. M. Gunter and N. Bijlani and G. Dennis and C. Lo and T. Quinnell and M. Symmonds and J. Welch and P.-L. Ratti and M. T. Hu and M. Villarroel},
howpublished = {medRxiv},
year = {2026},
month = jun,
doi = {10.64898/2026.06.30.26356905},
url = {https://www.medrxiv.org/content/10.64898/2026.06.30.26356905v2},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Multi-modal sleep staging in the clinic for REM sleep behaviour disorder
AU - K. M. Gunter
AU - N. Bijlani
AU - G. Dennis
AU - C. Lo
AU - T. Quinnell
AU - M. Symmonds
AU - J. Welch
AU - P.-L. Ratti
AU - M. T. Hu
AU - M. Villarroel
PY - 2026
DA - 2026/06/30
PB - medRxiv
DO - 10.64898/2026.06.30.26356905
UR - https://www.medrxiv.org/content/10.64898/2026.06.30.26356905v2
AB - Accurate REM identification is critical for diagnosing REM sleep behaviour disorder (RBD), yet many automated sleep-staging systems, especially single-channel EEG models trained on healthy cohorts, do not generalise well to real-life polysomnography (PSG) performed in patients. We compared a feature-based Random Forest (RF) model tuned for RBD with a state-of-the-art single- EEG deep architecture (AttnSleep), and assessed the impact of cohort adaptation and multimodal inputs (EEG, EOG, EMG, ECG). Experiments used 89 multi-site in-clinic PSGs (SleepWearables Phase-1) plus 53 MASS healthy controls (mean age 63 {+/-} 5 years), with 10-fold cross-validation and out-of-fold evaluation. When applied out-of-the-box after training on open-source healthy datasets, the RF outperformed AttnSleep on the clinical cohort (Cohens{kappa} = 0.56 versus 0.49), with the largest difference in REM sleep in RBD subjects (RF Cohens{kappa} = 0.51 versus AttnSleep Cohens{kappa} = 0.18), highlighting limited cross-cohort generalisation. Cohort-specific retraining substantially improved AttnSleep, and the multi-modal model further improved overall agreement (Cohens{kappa} 0.61 versus 0.63), with the largest stage-level gains in N1 and N3; in RBD subjects, however, overall agreement was unchanged (Cohens{kappa} 0.46 for both configurations). Attention-based modality analysis identified EEG as the dominant signal, increased EOG contribution during REM and N1, and elevated ECG importance during N3. In RBD subjects, EOG weighting increased relative to subjects without RBD ({Delta} = +0.096) while EEG weighting fell ({Delta} = -0.056), suggesting physiological relevance. Guided by these weights, a reduced four-channel EEG model matched full multimodal performance in non-RBD subjects, and adding EOG matched it overall (Cohens{kappa} = 0.63) while achieving the best performance in RBD (Cohens{kappa} = 0.49) and improved REM classification in RBD (60.6% compared to 53.0% recall). Inclusion of EOG also halved inter-dataset variability in REM staging between the clinical and MASS cohorts. Nonetheless, staging performance in RBD remained lower than in controls, particularly for REM, and within the RBD subgroup only the REM improvement from the four-channel EEG + EOG configuration reached statistical significance at this sample size (n = 22). These results highlight (1) the limited generalisability of minimal-sensor models trained on healthy cohorts, (2) the value of mixed cohort-specific training, and (3) that the composition of the sensor set matters more than its size: targeted addition of EOG improved REM staging in RBD specifically, whereas full multimodal integration improved performance mainly outside the RBD subgroup.
ER -