What PSG-paired research means for your phone.

PSG (polysomnography) is a clinical sleep study. Pairing audio with its reference labels lets us test whether the methods behind SomniSense, a sleep apnea detector app, recognize the breathing patterns being studied.

Here is what our studies tested, what they found and how that evidence connects to the breathing record you can inspect on your own phone.

Why paired recordings matter

Paired recordings link the audio and the sleep-study reference to the same person, on the same night. That gives the model a relevant reference for the audio patterns being studied, rather than comparing unrelated group averages.

The dataset covers 40 participants and 80 person-nights: 10 in-laboratory nights and 70 nights in participants’ home environments with portable sleep-study equipment. Consumer smartphone and tablet microphones captured audio across different rooms and conditions.

What our studies measured

Our five-seed component evaluation reports 94.29% accuracy for 13,538 one-second snore segments, and 87.14% mean accuracy for 2,953 labeled 200-second breathing windows. Each result names the task and unit, so you can see what the evidence supports. The split is at window level, not participant-disjoint; it does not establish performance on a wholly unseen group of people.

The separate compression study reports 88.49% accuracy and 88.06% F1 on the breathing-window task in its single-seed evaluation. It demonstrates that a compact model can retain useful classification performance; it is not a second estimate of the same five-seed result.

What the results let you judge

These are concrete results for recognizing the labeled audio patterns in the research data. Paired reference data, explicit tasks and reported results let you examine the method behind SomniSense rather than taking “validated” on trust.

The app puts that work to use in a record you can inspect: possible breathing marks, estimated times and intervals, waveforms and available surrounding sound. The window-classification studies do not measure the clinical correctness of every app event, nightly BRI or diagnosis.

Where the reference came from

Reference: audio paired with PSG-based event annotations. The exact scorer certification and blinding procedures are not documented in these sources. Research · Accuracy