J.-B. Martinot, N. Le-Dong, C. Letesson, V. Cuthbert, D. Gozal, J.-L. Pepin.
Introduction: To facilitate a more streamlined diagnostic approach of obstructive sleep apnea syndrome (OSAS), we
validated an artificial intelligence (AI) based diagnostic platform (Sunrise, Namur, Belgium), by combining single channel recordings of sleep mandibular movements (MM) with self-reported symptoms, co-morbidities and anthropometry. Materials and method: The Sunrise system consists of a wearable sensor attached on the chin which records MM and communicates with a smartphone. The collected data were transferred to a cloud-based infrastructure and processed with a machine laming algorithm. First, the algorithm estimates total sleep time (TST), arousal index (Arl), duration (ORDd!) and hourly index (ORDI) of obstructive respiratory disturbances, then combined these scores with other 25 clinical features including self-reported symptoms, smoking status, Epworth score, co-morbidity and anthropometric data, to establish a diagnosis at the cut-off of 5 n/h (OSAS class 1) or 15 n/h even in the absence of associated symptoms or morbidities (OSAS class 2) according to ICSD-3. The experiment was conducted on 2 independent subsets (model development (n-462) and validation (n=192)) of patients who underwent in-lab polysomnography (PSG). PSG scoring and OSAS diagnosis were based on AASM 2012 rules. Results: The integrated diagnostic rule has a very good diagnostic performance when validated on unseen data of 192 persons (balanced accuracy of 0.84 and kappa coefficient of 0.81). The ROC-AUC values for 3 classes: non-OSAS, OSAS class 1 and OSAS class 2 were 0.96, 0.92 and 0.96, respectively. caption: The model interpretation using the Shapley additive explanation method also revealed that ORDI and body mass index contributed the most to the detection of OSAS class1 whilst both ORDI and ORDdt were the most important contributors to OSAS class 2. Conclusion: Sleep MM
recordings are the most important contributors when interfaced with self-reported symptoms and co-morbidities to provide optimized OSAS screening approach.