Clinical Validation of Artificial Intelligence Algorithms for the Diagnosis of Adult Obstructive Sleep Apnea and Sleep Staging From Oximetry and Photoplethysmography-SleepAI

Publication: https://onlinelibrary.wiley.com/doi/full/10.1111/jsr.70093

Diagnosing sleep disorders like obstructive sleep apnea (OSA) has traditionally relied on resource-intensive, in-lab polysomnography (PSG). While home sleep apnea tests (HSATs) offer convenience, their accuracy often falls short. In our latest study, we introduce SleepAI, a remote diagnostic system that leverages cutting-edge artificial intelligence (AI) to transform how we monitor sleep and diagnose OSA using just oximetry data.

SleepAI builds on our previous work developing state-of-the-art deep learning models for sleep staging using photoplethysmography (PPG) (Kotzen et al., IEEE JBHI, 2023) and for estimating the apnea-hypopnea index (AHI) from oximetry (Levy et al., Nature Communications, 2023). These models, trained on large-scale retrospective datasets, achieved best-in-class performance for their respective tasks.

In this new study, we sought to validate the clinical utility of these models prospectively. We evaluated SleepAI on a cohort of 53 patients with suspected OSA, comparing its outputs to those of full PSG, the gold standard in sleep medicine. Despite operating with only oximetry input, SleepAI accurately classified OSA severity with 89% accuracy, far surpassing the average 61% accuracy reported for traditional HSATs (Massie et al., 2022). For sleep staging, it achieved a Cohen’s kappa of 0.75 for three-class classification, reflecting substantial agreement with manual PSG scoring.

Figure 1: Performance of SleepAI in AHI estimation and OSA severity estimation.

Figure 2: Performance of SleepAI for sleep measures estimations.

What makes SleepAI particularly compelling is its scalability. Using only a single-channel oximeter, it can offer high-fidelity sleep analytics remotely, enabling wider access to sleep diagnostics while reducing cost and burden on sleep labs.

As AI continues to gain traction in medicine, SleepAI illustrates its transformative potential in sleep health. Our findings suggest that remote, AI-driven tools can match or even outperform conventional diagnostic pathways, marking an important step toward accessible and accurate digital sleep medicine.

We believe this work will resonate with researchers and clinicians alike, especially those exploring the intersections of AI, sleep staging, and digital health innovation.