💻 computer science

A Two-Stage Audio Cascade with Coordinate Attention for On-Device Snore and Sleep Apnea Detection

This paper presents an ultra-compact, two-stage on-device audio cascade that utilizes a lightweight 2D CNN for snore detection and an adapted Coordinate Attention mechanism for classifying sleep apnea events, achieving high accuracy with a 93.2% parameter reduction compared to baselines.

Original authors: Yang L.

Published 2026-07-03✓ Author reviewed ⓘ

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yang L.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: A Sleep Detective on Your Phone

Imagine you want to know if someone is having trouble breathing while they sleep (a condition called sleep apnea), but you don't want to strap them to a machine in a hospital. This paper proposes a way to do this using just the microphone in a regular smartphone.

The author, Yang L., built a "two-stage detective team" that lives entirely on the phone. It listens to the room, figures out if the person is snoring, and then decides if that snoring pattern suggests a dangerous breathing pause.

The Two-Stage Detective Team

Think of the system as a relay race with two runners:

Runner 1: The "Snore Spotter" (Stage 1)

The Job: This runner listens to the audio in tiny 1-second chunks.
The Tool: It's a small, fast computer brain (a neural network) that looks at a visual map of the sound (called a Mel-frequency map).
The Output: Every second, it shouts a simple "Yes" or "No": Is there snoring right now?
Why it matters: It doesn't decide if the person is sick yet; it just keeps a running log of when snoring happens.

Runner 2: The "Pattern Analyst" (Stage 2)

The Job: This runner looks at a longer window of time (200 seconds, or about 3 minutes). It takes the "Yes/No" log from Runner 1, plus two other clues: how loud the room is and how quickly the volume is changing.
The Goal: It decides if that 3-minute window contains a "breathing event" (an apnea or hypopnea).
The Innovation: This is where the paper gets fancy. The author adapted a new way for this runner to pay attention.

The Secret Sauce: "Coordinate Attention"

Usually, computer brains look at data like a blurry photo where they just guess what's important.

The Old Way (Standard CNN): Imagine looking at a long strip of paper with notes on it. The old method might squint at the whole strip and say, "The middle part seems important," but it forgets exactly where on the strip the important note was.
The SE-Attention Way: This is like having a highlighter that marks the whole strip with the same color. It knows what to look for, but it treats every second of time the same.
The New Way (Coordinate Attention): The author adapted a technique originally used for spotting objects in 2D images (like finding a cat's ear in a photo) and turned it into a 1D time-traveler.
- The Analogy: Imagine a librarian who doesn't just know which books are important, but also knows exactly which shelf and which row they are on. This new method creates a map that says, "At second 45, the 'loudness' clue is super important," but at second 46, the 'snoring' clue is the star. It preserves the position of the sound in time.

The Results: Small Brain, Big Smarts

The author tested three different "brains" for the second runner:

The Old Brain: A standard, heavy computer model.
The SE-Brain: A model with a standard attention mechanism.
The New Brain: The model with the new "Coordinate Attention."

The Findings:

Size Matters: The new brain is tiny. It has 93% fewer parts (parameters) than the old heavy brain. It's like shrinking a full-size truck into a compact car without losing the engine's power.
Accuracy: The new brain was more accurate than the heavy truck. It correctly identified sleep issues about 87% of the time, compared to 84% for the heavy model.
The "False Alarm" Win: When comparing the new brain to the SE-Brain (which is also small), the new brain was better at not crying wolf. It made fewer false alarms (high precision and specificity). This is crucial for a phone app; you don't want to wake someone up or scare them with a false alert.

Important Limitations (What the Paper Doesn't Say)

The paper is very honest about what it hasn't proven yet:

The "Stranger Danger" Test: The study tested the model on data from 40 people, but the training and testing sets mixed up the data from those people. The paper admits it hasn't proven the model works perfectly on a brand new person it has never met before. That is the next step.
The "One Night" Snapshot: The model looks at 3-minute windows. It doesn't yet claim to give a full diagnosis for a whole night's sleep or calculate a specific medical score (AHI) for a doctor. It just detects if an event happened in that window.
Privacy: The system works entirely on the phone. No audio is sent to the cloud, which keeps the user's data private.

The Bottom Line

The author built a super-efficient, two-step system that can run on a smartphone to detect sleep breathing problems. By using a clever new "attention" trick that remembers when things happen in time, they made a model that is tiny, fast, and better at avoiding false alarms than previous methods. It's a promising step toward making sleep health screening accessible to everyone, right from their bedside.

Problem Statement

Obstructive Sleep Apnea (OSA) is a prevalent chronic disorder with significant health risks, yet an estimated 80–90% of cases remain undiagnosed globally. The clinical gold standard, in-laboratory polysomnography (PSG), is expensive, inaccessible to many, and provides only a single-night snapshot. While smartphone audio offers a potential low-cost, non-contact screening alternative, existing deep learning approaches face two primary engineering challenges:

Temporal Structure: Standard 1D Convolutional Neural Networks (CNNs) often apply uniform weighting across time steps, failing to capture the clinically meaningful temporal phases of apnea events (onset, silence plateau, and recovery).
Edge Deployment: Many high-accuracy models require cloud inference due to their size, compromising privacy and real-time capability on consumer devices.

The core engineering challenge addressed is whether a lightweight neural network, operating entirely on a smartphone using only the built-in microphone, can detect sleep-disordered breathing events with clinically meaningful sensitivity while maintaining low inference latency.

Methodology

The authors propose a two-stage on-device audio cascade designed for edge inference.

Stage 1: Short-Window Snore Detection

Input: 1-second audio segments converted into 15 × 13 Mel-frequency cepstral coefficient (MFCC) matrices.
Architecture: A compact 2D CNN (301,473 parameters) classifies each segment as "snore" or "non-snore."
Output: A binary flag (0 or 1) sampled at 1 Hz.

Stage 2: Long-Window Apnea/Hypopnea Detection

Input: A 200-second window represented as a 200 × 3 feature matrix sampled at 1 Hz. The three channels are:
1. Sound Pressure Level (SPL) in dB.
2. Rate of change of SPL (dB/sec).
3. The binary snore-presence flag generated by Stage 1.
Task: Binary classification of whether the 200-second window contains an apnea or hypopnea event (defined as ≥10 seconds of reduced or absent airflow).
Proposed Architecture (CA-1D): The authors adapt Coordinate Attention (CA), originally designed for 2D vision, to 1D time-series data.
- Mechanism: Instead of the global average pooling used in Squeeze-and-Excitation (SE) attention (which collapses time and produces a time-invariant $1 \times C$ vector), the CA-1D module employs a global–local fusion scheme. It computes a global context vector, tiles it to the temporal dimension, and concatenates it with local features. This produces an attention map of shape $T \times C$ (time $\times$ channels), preserving temporal position information.
- Comparison Baselines: The proposed model is compared against a vanilla 1D CNN (heavy classifier head) and an SE-Attention CNN (time-invariant weights).

Experimental Protocol

Dataset: 2,953 PSG-labeled 200-second windows from 40 participants (80 person-nights), comprising both hospital and home recordings.
Evaluation: A rigorous five-seed bootstrap protocol. Models were trained five times with different random seeds, and results were reported as means with 95% bootstrap confidence intervals. Pairwise comparisons used paired bootstrap resampling to control for split-induced variability.

Key Contributions

Two-Stage Cascade: A novel pipeline where a short-window snore detector feeds a binary feature into a long-window event classifier, creating an ultra-compact intermediate representation (~600 values per window) suitable for mobile devices.
CA-1D Adaptation: A methodological adaptation of Coordinate Attention for 1D biomedical time-series, utilizing global-local fusion to preserve temporal position in the attention map.
Rigorous Multi-Seed Evaluation: An empirical study comparing three architectures across five random seeds, addressing the common weakness of single-seed reporting in small-dataset biomedical AI papers.
Parameter Efficiency: A reduction of model parameters by 93.2% (from 204,801 to 14,001) while improving performance metrics.
Honest Characterization of Attention: A precise delineation of where Coordinate Attention adds value (Precision and Specificity) versus where it does not (F1 or AUC) compared to SE-Attention.

Results

Evaluated on the 2,953-sample dataset:

Stage 1 (Snore Detection): Achieved 94.29% accuracy and 90.28% F1-score.
Stage 2 (Apnea Detection):
- Coord-Attn vs. Original CNN: The proposed CA-1D model achieved 87.14% accuracy and 86.94% F1, significantly outperforming the 204k-parameter baseline (83.82% accuracy) across six of seven metrics (Accuracy, Precision, Specificity, F1, AUC-ROC, AUC-PR).
- Coord-Attn vs. SE-Attention: When compared to the SE-Attention baseline (13,629 parameters), the CA-1D model showed statistically significant improvements in Precision (+2.62 pp) and Specificity (+3.40 pp). However, differences in F1, AUC-ROC, and AUC-PR were not statistically significant under the multi-seed bootstrap analysis.
- Parameter Efficiency: The CA-1D model achieved >86% accuracy with only ~14k parameters, a >14x reduction compared to the baseline, while maintaining high accuracy.

Significance and Claims

The paper claims that the proposed architecture successfully demonstrates that position-aware attention mechanisms can be adapted for 1D audio time-series to improve on-device sleep apnea detection.

Clinical/Engineering Impact: The model offers a path to scalable, privacy-preserving sleep screening on consumer smartphones by reducing model size by over 90% while improving detection accuracy.
Methodological Insight: The study highlights that while Coordinate Attention offers a genuine advantage in reducing false positives (higher Precision and Specificity) compared to SE-Attention, this does not necessarily translate to a statistically significant gain in overall F1 or AUC scores. This nuance is critical for practitioners designing screening tools where false alarms drive user disengagement.
Reproducibility: The authors emphasize the importance of multi-seed bootstrap evaluation to avoid misleading conclusions when comparing similar-sized architectures on small datasets.

Limitations Noted by the Authors:

The evaluation is window-level, not participant-disjoint; the model has not been validated on entirely unseen individuals (leave-subjects-out).
The dataset, while diverse in recording environments, is limited to 40 participants.
The study focuses on the detection of the event window, not the aggregation of scores into a night-level AHI estimate or full clinical sub-typing.