Validation of an artificial intelligence-driven novel microphone-based wireless jaw movement and respiratory event detection system for home sleep apnea testing

Home sleep apnea testing (HSAT) is increasingly used for obstructive sleep apnea (OSA) it captures sleep in natural, at-home settings. However, audio-only approaches are challenged by hardware heterogeneity and the modest association between snoring loudness and the apnea–hypopnea index (AHI). This study utilized a smartphone-recorded HSAT combining broadband respiratory audio (∼200–15,000 Hz) with a chip-less, angle-triggered ultrasonic emitter affixed to the mandible. The emitter writes narrowband tones into the same 48-kHz track—18.0 kHz as a mandibular-related anchor and 19.5 kHz as a cue—providing posture/movement anchors without inertial sensors. A multimodal AI pipeline fused handcrafted ultrasonic/audio features with learned representations from SoundSleepNet and a Vision Transformer (ViT), integrated by a Deep Neural Decision Forest (DNDF) to label 30-s epochs (Awake/Normal/Event). HSAT-derived AHI was compared with simultaneous polysomnography (PSG) scored per AASM-2023 rules (hypopnea defined according to the AASM recommended criteria, requiring a ≥3% oxygen desaturation or an associated EEG arousal) in a prospective single-center cohort ( ClinicalTrials.gov NCT06862297; n = 96). SoundSleepNet many-to-many outperformed single-to-single modeling while ViT achieved optimal accuracy using the embedded 19,475–19,525 Hz band. In the prediction set, DNDF-AHI correlated with PSG-AHI (R 2 = 0.6986) with a mean Bland–Altman bias of +5.85 (95% limits −24.24 to +35.94) events·h −1 . OSA-severity accuracy was 92% for binary scheme (AHI ≥ 15 vs