A Cascaded Two-Stage CNN Pipeline for Audio-Based Snore and Sleep Apnea Detection on Smartphones: Reference Baselines with Multi-Seed Validation

Smartphone-based audio analysis offers a non-contact, low-cost path to sleep health screening at population scale. We present a cascaded two-stage CNN pipeline for audio-based sleep monitoring on consumer mobile devices: a short-window 2D CNN processes 1-second audio segments represented as 15 × 13 Mel-frequency cepstral coefficient (MFCC) matrices and produces a per-segment binary snore-presence indicator; that per-second indicator, together with two scalar acoustic features (sound pressure level and SPL change rate), is assembled into a 200 × 3 time-series feature matrix at 1 Hz sampling that forms the input to a long-window 1D CNN classifying the 200-second observation window as containing a sleep-disordered breathing event or not. The cascade structure — Stage-1 detector output forming a feature channel of the Stage-2 classifier's input — provides an ultra-compact intermediate representation (200 × 3 ≈ 600 floats per analysis window) suitable for on-device inference on consumer mobile hardware. We evaluate both stages under a multi-seed protocol with 95% bootstrap confidence intervals (10,000 iterations). On a dataset of 13,538 labeled 1-second snore segments and 2,953 labeled 200-second apnea-detection windows — derived from 80 person-nights across 40 participants, with PSG event annotations from a mix of in-laboratory PSG (10 nights) and portable / ambulatory PSG with nasal-airflow cannula (70 nights) — the two stages achieve, respectively, 94.29% (93.60, 95.02) accuracy with F1 90.28% (89.29, 91.38) for Stage-1 (snore detection), and 83.82% (82.61, 85.14) accuracy with F1 83.99% (82.87, 85.05) for Stage-2 (sleep apnea detection). The two-stage cascade and its compact intermediate representation are presented as reference baselines, with multi-seed bootstrap metrics that can serve as comparison points for downstream methodological work. Data and code availability: The model architectures, training pipeline, and multi-seed bootstrap analysis code are publicly available at https://github.com/somnisense/audio-sleep-cnn-baselines under the MIT License. By design, no audio recordings, labels, or derived feature matrices are distributed with the repository: the original recordings were collected under participant consent that does not cover public release of either the waveforms or the derived features. Competing interests: The author is the founder of SomniAI LLC, the entity that may commercialize the technology described in this work. The methods described are also the subject of three co-filed U.S. provisional patent applications by SomniAI LLC. No external funding was received.