ApneaSense · Research
Acoustic sleep apnea research: tasks, metrics and limits
Read the work as acoustic classification research, not an OSA diagnostic or screening claim. PSG annotations label 200-second windows as normal or containing apnea or hypopnea. That research task is distinct from the candidate acoustic moments shown in SomniSense and from a future evaluation on a buyer device.
Review the SDK evaluation path →Dataset and evaluation unit
Paper A describes 80 person-nights from 40 participants: 10 in-lab PSG nights and 70 ambulatory PSG nights. Audio used varied consumer phone and tablet microphones. The data comprise 13,538 labeled one-second snore segments and 2,953 labeled 200-second breathing windows.
The breathing labels contain 1,498 normal windows and 1,455 apnea-or-hypopnea windows. The latter combine 1,089 apnea-labeled and 366 hypopnea-labeled windows into one class. These are window counts, not a count of individual clinical events or diagnosed patients.
Keep each number with its configuration
Paper A: 94.29% mean accuracy and 90.28% F1 for one-second snore-segment classification, five seeds. Its baseline 200-second breathing-window accuracy is 83.82%; do not substitute this baseline for the compressed result.
Paper E: Pruned 50% + fine-tuning + QAT reports 88.49% accuracy, 88.06% F1, 9,416 parameters and a 56.4 KB stored model. Compression uses seed 42 on the 2,953-window corpus with a stratified 60/20/20 split; the reported test set has 591 windows.
Separately, Paper E measures CoreML Pruned 50% + fine-tuning forward-pass latency on an Apple M2 MacBook Pro: 0.064 ms mean, 0.076 ms P95, over 200 calls after 20 warmups. Audio preprocessing and post-processing are excluded. This is not the same measured configuration as the QAT row or a buyer-device end-to-end guarantee.
What remains unproven
The split is described as stratified per seed, not as a participant-independent prospective clinical trial. The cohort is small; cross-device, cross-site and population generalization need further evaluation. Window metrics do not establish candidate start/end accuracy, duration accuracy, clinical AHI or person-level screening sensitivity.
Paper E describes retrospective exploratory work with data-provenance limitations, including no available formal ethics-review record and consent that did not explicitly address downstream research publication. It should not be presented as a formally ethics-cleared clinical study.
Methods are public; health data are not
The method and training code are public. Participant audio, labels and derived feature matrices are not distributed; the compression repository does not supply trained checkpoints. You can inspect and run methods on a suitably authorized corpus, but cannot independently reproduce the original dataset result from the public code alone.
Research
Research sources are English originals. These are preprints, not evidence of peer-reviewed acceptance. Code access does not include the original health dataset or a buyer SDK.
Coordinate Attention · Stage 2
Paper E · Compression
Questions before you proceed
Does 88.49% mean it finds 88.49% of patients with OSA?
No. It is accuracy across labeled 200-second windows. It is not patient-level sensitivity, clinical event-boundary accuracy, or a guarantee for a personal night.
Are apnea and hypopnea separate model outputs here?
The reported breathing benchmark merges apnea-labeled and hypopnea-labeled windows into one class. It does not establish separate clinical diagnoses from sound.