TLDR: This research introduces a novel method to screen for Obstructive Sleep Apnoea (OSA) by estimating respiratory effort directly from nocturnal breathing sounds recorded with a smartphone. Current diagnostic methods like polysomnography are costly and inconvenient. The proposed latent-space fusion framework combines estimated respiratory effort embeddings with acoustic features, improving OSA detection sensitivity and AUC compared to audio-only approaches. This sensor-free technique offers a scalable and non-invasive solution for at-home and longitudinal OSA monitoring.
Obstructive Sleep Apnoea (OSA) is a widespread condition characterized by repeated interruptions in breathing during sleep. It affects a significant portion of the adult population and is linked to serious health issues like cardiovascular disease and diabetes. Despite its prevalence and health risks, many individuals with OSA remain undiagnosed due to the high cost and complexity of traditional overnight polysomnography (PSG).
PSG, the gold standard for OSA diagnosis, involves numerous sensors and requires patients to spend a night in a specialized laboratory. This process is not only expensive and labor-intensive but can also disrupt sleep and lead to long waiting lists, hindering early detection and continuous monitoring. Furthermore, the severity of OSA can vary significantly from night to night, making a single PSG assessment potentially insufficient for a complete picture of the disease burden.
Acoustic-based monitoring has emerged as a promising, non-invasive alternative. Snoring and breathing sounds contain valuable physiological information about airway obstruction. While previous studies have shown the feasibility of acoustic screening, their performance can be limited by environmental noise and a lack of physiological context. Respiratory effort, a crucial signal in clinical OSA scoring, typically requires additional contact sensors, which reduces the scalability and comfort benefits of acoustic methods.
A Novel Approach: Estimating Respiratory Effort from Sound
A recent study introduces a groundbreaking approach to overcome these limitations by estimating respiratory effort directly from nocturnal audio recordings. This innovative method allows for the recovery of physiological context using sound alone, eliminating the need for additional sensors. The researchers propose a latent-space fusion framework that integrates these estimated effort embeddings with acoustic features to enhance OSA detection.
The study utilized a dataset of 157 nights from 103 participants, with recordings captured in real home environments using smartphones placed near the bed. The respiratory effort estimator achieved a concordance correlation coefficient of 0.48, demonstrating its ability to capture meaningful respiratory dynamics from audio. This indicates that even subtle imprints of respiratory activity in snoring and breathing sounds can be effectively modeled.
Improved Screening Performance
The fusion of estimated respiratory effort with acoustic features significantly improved OSA screening performance compared to audio-only baselines. This was particularly evident at lower apnoea–hypopnoea index (AHI) thresholds, which are critical for detecting mild OSA cases that might otherwise be missed. For instance, at an AHI cut-off of 5 events/h, the proposed latent-space fusion (LSF) model achieved the highest sensitivity (0.88), outperforming both audio-only and even oracle systems that use measured effort.
At moderate AHI cut-offs (15 events/h), the LSF model with estimated effort also showed improved sensitivity and achieved the highest Area Under the Curve (AUC) of 0.88, highlighting its strength in distinguishing between moderate and severe OSA. While performance converged at the highest AHI threshold (≥30), the overall results underscore that audio alone contains latent physiological cues capable of supporting effective at-home OSA screening.
Also Read:
- Engineered Faults Enhance Anomaly Detection in Electric Vehicle Cabin Sounds
- Enhancing ECG Foundation Models with a Targeted Post-Training Strategy
Scalability and Future Implications
The key advantage of this proposed method is its scalability and non-invasiveness. At the time of testing, it requires only smartphone audio, eliminating the need for additional sensors that can reduce patient comfort or limit long-term monitoring. This sensor-free approach paves the way for more accessible, scalable, and longitudinal OSA monitoring in home environments.
The researchers acknowledge that while the gains are modest compared to systems using direct physiological measurements, they are significant given the challenges of noisy real-world smartphone recordings. Future work aims to explore simpler physiological targets like breathing-rate variability, improve effort prediction through advanced models, and develop robust training objectives to account for noise and temporal misalignments.
This research represents a significant step towards making OSA screening more accessible and convenient, potentially leading to earlier diagnosis and better management of this prevalent sleep disorder. You can read the full research paper here.
