Artificial Intelligence Analysis of Mandibular Movements Enables Accurate Detection of Phasic Sleep Bruxism in OSA Patients: A Pilot Study

Abstract

Jean-Benoit Martinot,1,2 Nhat-Nam Le-Dong,3 Valérie Cuthbert,1 Stéphane Denison,3 David Gozal,4 Gilles Lavigne,5 Jean-Louis Pépin6 1Sleep Laboratory, CHU Université Catholique de Louvain (UCL) Namur Site Sainte-Elisabeth, Namur, 5000, Belgium; 2Institute of Experimental and Clinical Research, UCL Bruxelles Woluwe, Brussels, 1200, Belgium; 3Sunrise, Namur, 5101, Belgium; 4Department of Child Health and Child Health Research Institute, University of Missouri, Columbia, MO, 65201, USA; 5Faculté de médecine dentaire, Université de Montréal, Montréal, Québec, H3C 3J7, Canada; 6HP2 Laboratory, Inserm U1042, University Grenoble Alpes, Grenoble, 38000, FranceCorrespondence: Jean-Benoit MartinotCentre du Sommeil et de la Vigilance, CHU UCL Namur Site Ste Elisabeth, 15, Place Louise Godin, Namur, 5000, BelgiumTel [phone]Fax [phone]Email [email protected]: Sleep bruxism (SBx) activity is classically identified by capturing masseter and/or temporalis masticatory muscles electromyographic activity (EMG-MMA) during in-laboratory polysomnography (PSG). We aimed to identify stereotypical mandibular jaw movements (MJM) in patients with SBx and to develop rhythmic masticatory muscles activities (RMMA) automatic detection using an artificial intelligence (AI) based approach.Patients and Methods: This was a prospective, observational study of 67 suspected obstructive sleep apnea (OSA) patients in whom PSG with masseter EMG was performed with simultaneous MJM recordings. The system used to collect MJM consisted of a small hardware device attached on the chin that communicates to a cloud-based infrastructure. An extreme gradient boosting (XGB) multiclass classifier was trained on 79,650 10-second epochs of MJM data from the 39 subjects with a history of SBx targeting 3 labels: RMMA episodes (n=1072), micro-arousals (n=1311), and MJM occurring at the breathing frequency (n=77,267).Results: Validated on unseen data from 28 patients, the model showed a very good epoch-by-epoch agreement (Kappa = 0.799) and balanced accuracy of 86.6% was found for the MJM events when using RMMA standards. The RMMA episodes were detected with a sensitivity of 84.3%. Class-wise receiver operating characteristic (ROC) curve analysis confirmed the well-balanced performance of the classifier for RMMA (ROC area under the curve: 0.98, 95% confidence interval [CI] 0.97– 0.99). There was good agreement between the MJM analytic model and manual EMG signal scoring of RMMA (median bias − 0.80 events/h, 95% CI − 9.77 to 2.85).Conclusion: SBx can be reliably identified, quantified, and characterized with MJM when subjected to automated analysis supported by AI technology.Keywords: masticatory muscular activities, machine learning, jaw movement

Keywords

• masticatory muscular activities
• machine learning
• jaw movement