At night, the bedroom hosts an orchestra, with breathing as the soloist. Under the sheet lies a mat with 512 sensors, each capturing the slightest body movements. When breathing stops for ten seconds, the orchestra falls silent. The challenge is to hear that silence.

Previously, it was unclear which notes in this music matter most. Researchers dissected the signal into parts: statistics, zero crossings, energy, wavelets, fractal dimension. Ten groups in total, 191 features. But which ones truly indicate a pause in breathing?

It turned out that the conductor of this orchestra is breathing frequency. Power in the narrow band of 0.1–0.4 Hz, where respiratory movements occur, is most important. Wavelets and nonlinear measures are almost unnecessary: their contribution is only 10.4% across 59 features.

The Random Forest model achieved an AUC-ROC of 0.967, meaning it distinguishes apnea events from clean windows almost without error. For 79% of patients, accuracy exceeded 0.95. Strict leave-one-patient-out validation ensures the result is not due to chance.

Now that we know which notes to listen for, we can assemble a compact orchestra. Instead of 191 features, 48 frequency and temporal features suffice. Such a mat could work at home, without attached sensors or complex equipment.

Let's return to the orchestra. Now we know that a pause in breathing is heard not in complex patterns but in a simple rhythm. The mat under the sheet could become a home guardian of sleep, one that doesn't interfere but simply listens.