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 <15) with 90% sensitivity and 85% specificity. Integrating ultrasonic mandibular markers with audio-based AI and DNDF enable accurate binary discrimination of moderate-to-severe OSA, supporting scalable, posture-aware HSAT screening and longitudinal management.
Tani et al. (2026) studied this question.