How accurate is acoustic breathing-event detection against PSG-labeled data?

Across 80 person-nights from 40 participants, the published production benchmark reaches 88.49% accuracy for 200-second breathing-event-window classification; 1-second snore-event detection reaches 94.29% accuracy. Full results, per-seed metrics, and the code are public. These are research benchmarks, not a diagnosis.

Can a phone microphone detect sleep apnea events?

It can detect the acoustic signature of snoring and breathing events on-device, validated against PSG. It is an acoustic proxy — it does not measure airflow or blood oxygen directly, so PSG and continuous SpO₂ remain more direct for the physiological event.

Is the code available to reproduce the results?

Yes — all three papers ship an MIT-licensed code companion on GitHub. The repositories release the algorithm framework and training code (no data, by participant-consent constraints), so you can run the methodology on your own corpus.

Research & validation

The methodology, the numbers, and the code — checkable.

Name: ApneaSense PSG validation corpus
Creator: SomniAI LLC
License: https://apneasense.com/research/#data-availability

These are SomniAI LLC's internal validation results, validated against polysomnography. Three papers, three open-source repositories. Don't take the numbers on faith — run the code on your own data.

The validation set

Detection was evaluated against in-lab and ambulatory polysomnography (PSG): 80 paired PSG nights across 40 participants — 10 in-lab PSG and 70 ambulatory PSG nights with a nasal-airflow cannula, smartphone audio captured simultaneously. The ambulatory majority means most of the data is real home environments, not a quiet lab.

Two time-aligned tracks — acoustic detection above, PSG reference below — with most events lining up between them. — Illustrative A schematic of how acoustic detection is scored against the PSG reference, event by event — not plotted data. The actual agreement is the accuracy numbers above and the per-seed results in the papers and code below.

The complete system — one two-stage cascade

The production model is a complete two-stage on-device audio cascade: Stage-1 snore detection feeds Stage-2 breathing-event detection, with Coordinate-Attention 1D as the Stage-2 method. The whole system is written up as a single paper — a preprint on Research Square with a citable DOI (not peer-reviewed yet, so I'll call it a preprint). Read it — Research Square → · Google Scholar →

How it's built — baseline → attention → production

The same work, broken into the three steps it was built from — a cascade, its attention-based Stage-2 classifier, then that classifier compressed for the device. Each is its own Zenodo preprint with an MIT-licensed code companion.

Step 1 · baseline & cascade

Two CNN baselines + the cascaded two-stage pipeline

The baseline snore and apnea CNNs, with multi-seed bootstrap validation — and the Stage-1→Stage-2 cascade that the next two build on.

94.29%snore accuracy (93.60–95.02)

83.82%apnea baseline accuracy

0.983snore AUC-ROC

Read the paper (PDF) → Run the code (GitHub) → Cite — Zenodo DOI →

Step 2 · attention (Stage-2)

Coordinate-Attention 1D — the Stage-2 classifier

Replaces the baseline apnea CNN above as the cascade’s Stage-2 classifier: a 14,001-parameter attention model that keeps temporal position — a 93.2% parameter cut, with accuracy preserved or improved.

87.14%accuracy (85.14–89.68)

14,001parameters

93.2%param reduction vs baseline

Read the paper (PDF) → Run the code (GitHub) → Cite — Zenodo DOI →

Step 3 · compression → production

Compression → the on-device production model

Compresses the Stage-2 model above to production size: INT8 quantization-aware training + 50% structured pruning + CoreML. On this dataset, compression raises test accuracy.

88.49%breathing-window accuracy (INT8 + pruned)

56.4 KB9,416 INT8 params

0.064 msinference · Apple M2 Neural Engine

Read the paper (PDF) → Run the code (GitHub) → Cite — Zenodo DOI →

What we publish, and what the patent covers

The papers and code disclose the full method — the cascaded two-stage architecture over a compact 200×3 @ 1 Hz representation, the Coordinate-Attention 1D formulation, the quantization and pruning protocol, and the evaluation methodology — so the results are reproducible. Certain production specifics — multi-stage gating, event-driven triggering, and the privacy-preserving on-device system architecture — are covered by three co-filed U.S. provisional patent applications (SomniAI LLC) and are not described.

Data availability

The validation corpus — the 80 paired PSG nights and the simultaneous smartphone audio — is not publicly distributed. Participants consented to internal validation, not redistribution, so the raw recordings and labels stay private. What's public is the method and the code: each paper above ships an MIT-licensed companion repository — the algorithm framework and training code, no data — so the methodology is fully reproducible on your own corpus.

Limitations we hold to

Acoustic proxy. We detect sound, not airflow or blood oxygen — PSG and continuous SpO₂ measure the physiological event more directly. We don't claim to beat them.
Screening, not diagnosis. ApneaSense is not a medical device and is not FDA-cleared.
Precision-first. Tuned to avoid false positives, so it under-counts — read the rate as a conservative lower bound.
Sample. Recruited with sleep-breathing complaints; generalization across BMI and healthy cohorts is ongoing.

Citing this work

Authored at SomniAI LLC by the inventor on the pending U.S. patent (PAT-001). Each paper is published on Zenodo with a citable DOI — linked under each paper above; the combined cascade paper is on Research Square. Full profile on Google Scholar and ORCID.

Want to evaluate it on your own data? See how the SDK works or email the founder.