Sleep stages automatically scored from mandibular movement signal recorded by a unique chin sensor

Context: Automated sleep stages scoring from surrogate signals is challenging. Mandibular movements (MM) are reliable markers of the neural respiratory drive, with activity patterns reflecting sleep stages. We introduce a new automated sleep stage scoring method using MM signal. Method: Study was conducted on 96 adults (18 to 58 yrs, AHI MM signals were split into 30 seconds epochs, from which 217 features were extracted to classify 3 labels: wake, nonREM and REM sleep. The model was trained on PSG data from 68 subjects consisting of 169 hrs of nonREM, 26 hrs of REM and 21 hrs of wake. The model was validated on unseen data from 28 subjects composed of 147 hrs of nonREM, 38 hrs of REM sleep and 33 hrs of wake. Results: Subjects had a mean total sleep time of 7.14 hrs (CI: 5.53 to 8.49), a mean AHI of 3.57 events/h (CI: 1.00 to 6.92) and a mean arousal index of 9.22 events/h (CI: 4.36 to 16.69). The model showed good global performances, with a balanced accuracy of 0.79 (CI: 0.78 to 0.80) and an area under the receiver operating curve of 0.94, 0.91 and 0.93 for wake, nonREM and REM. Conclusion: This study demonstrated for the first time that MM signals are suitable for automated sleep stage scoring, providing a promising solution for home monitoring of sleep architecture.