Wearable sleep staging using photoplethysmography and accelerometry across sleep apnea severity: a focus on very severe sleep apnea

This study demonstrates that wearable sleep staging performance using photoplethysmography and accelerometry declines with increasing sleep apnea severity, particularly in very severe cases, highlighting the critical need for training datasets that adequately represent severe phenotypes and staging schemes tailored to specific clinical use cases.

The Gist

Imagine your body is a busy city, and sleep is the time when the city goes into "night mode" to clean up, repair roads, and recharge batteries. To know if the city is truly resting, we usually need a team of experts (doctors) to watch every single streetlight and power line in a special hospital room. This is called a sleep study, but it's expensive, uncomfortable, and hard to do every night.

This paper is about testing a smartwatch to see if it can do the job of those experts, but with a specific focus on people whose "city" is in total chaos due to severe sleep apnea.

Here is the story of their experiment, broken down simply:

1. The Goal: Can a Watch Tell the Difference?

The researchers wanted to see if a wrist-worn device could tell the difference between light sleep, deep sleep, and dreaming (REM), just like a human expert. They used two sensors on the watch:

• The Heartbeat Sensor (PPG): Like a tiny detective listening to the rhythm of the city's traffic (your heart rate).
• The Motion Sensor (Accelerometer): Like a security camera checking if the city is still or moving (your body movement).

They built a "brain" (a computer program) to learn how to read these signals and guess what stage of sleep you are in.

2. The Test Groups: The Quiet Library vs. The Stormy City

They tested their watch on two groups of people:

• Group A (The Sleep Lab): These people were in a quiet, controlled room. Their sleep patterns were relatively normal.
• Group B (The Hospital): These people were in a real-world hospital setting. Crucially, this group included many people with very severe sleep apnea.
  • The Analogy: Imagine Group A is a calm library where everyone is whispering. Group B is a city during a massive hurricane, with traffic jams, power outages, and sirens blaring. The "sleep apnea" is the hurricane disrupting the city's ability to rest.

3. The Big Discovery: The Watch Got Confused in the Storm

The results showed that the watch worked okay in the quiet library (Group A). But in the stormy city (Group B), especially for the people with the worst "hurricanes" (very severe apnea), the watch got confused.

• The Problem: When the breathing stops and starts violently (which happens in severe apnea), the heart rate goes crazy. The watch's "detective" couldn't tell if the heart was racing because you were dreaming, or because you were gasping for air.
• The Result: The watch's accuracy dropped significantly for the sickest patients.

4. The Lesson: You Need to Train the Watch on the Right Stuff

The researchers realized something important about how they taught the computer.

• The Mistake: They initially trained the computer mostly on data from the "quiet library" (Group A) and only a few "stormy city" nights.
• The Fix: They tried training the computer again, but this time they made sure it saw plenty of "stormy city" data.
• The Analogy: It's like training a pilot. If you only train a pilot on calm, sunny days, they will crash when they hit a thunderstorm. But if you train them specifically on how to handle turbulence, they can fly safely even in the worst weather.

They found that when the computer was trained on more severe cases, it got better at guessing sleep stages for those severe patients.

5. The "Coarser" Map

They also tried simplifying the map. Instead of trying to tell the difference between 5 specific types of sleep (like distinguishing between "light sleep" and "very light sleep"), they just asked the watch to tell the difference between 4 broad categories.

• The Result: This simplified map actually worked better! It was like using a map that just shows "City" vs. "Country" instead of every single street. It was less precise, but much more reliable when the weather was bad.

The Bottom Line

This study tells us that while smartwatches are great for tracking sleep, they struggle when the user has very severe sleep apnea.

However, the solution isn't to give up on the watch. The solution is to teach the watch differently. We need to make sure the computer learns from people who have severe breathing problems, not just healthy people. If we do that, and maybe simplify how we describe the sleep stages, these watches could become powerful tools to help doctors manage severe sleep apnea from the comfort of a patient's home, without needing a hospital visit every night.

Technical Summary

Problem Statement

The management of sleep apnea, particularly in its most severe forms, currently relies on repeated laboratory polysomnography (PSG), which is impractical for long-term monitoring. While wearable devices offer a scalable alternative, their reliability in staging sleep for patients with very severe sleep apnea (defined as an Apnea-Hypopnea Index [AHI] $\ge$ 50 events/h) remains unclear. A critical gap exists in understanding how the composition of training datasets affects model performance when applied to this specific, high-risk clinical subgroup.

Methodology

The study employed a deep learning approach to evaluate wearable sleep staging across a spectrum of sleep apnea severity.

• Data Sources: The analysis utilized 552 overnight recordings from two distinct cohorts:
  • Sleep Lab Dataset: 318 recordings.
  • Hospital Dataset: 234 recordings, where 26.5% of subjects suffered from very severe sleep apnea.
• Input Features: The model utilized physiological signals captured by wrist-worn devices:
  • Photoplethysmography (PPG): Specifically, RR intervals derived from the PPG signal.
  • Accelerometry: Three-axis motion data.
• Model Architecture & Training:
  • A deep learning model was developed for sleep staging.
  • Staging Granularity: Performance was assessed under two schemes: 5-stage (standard) and 4-stage (coarser) classification.
  • Validation Strategy: Baseline performance was evaluated using cross-validation.
  • Ablation Study: To isolate the impact of training data composition, the authors compared the baseline model against an "ablation model." The ablation model was trained on the same number of recordings but with a higher proportion of Sleep Lab data and fewer low-AHI Hospital recordings, effectively reducing the representation of very severe cases in the training set.
• Analysis: The study examined night-level associations between model performance and AHI severity and specifically evaluated the models on the subgroup of 62 recordings with very severe sleep apnea.

Key Results

• Performance Disparity by Cohort:
  • In 5-stage classification, Cohen's kappa was 0.586 for the Sleep Lab Dataset but dropped significantly to 0.446 for the Hospital Dataset.
  • In 4-stage classification, the performance gap narrowed but persisted, with kappa values of 0.632 (Sleep Lab) vs. 0.525 (Hospital).
• Impact of Severity: Within the Hospital Dataset, sleep staging performance declined as AHI severity increased.
• Training Data Sensitivity: The ablation study revealed that reducing the representation of high-AHI cases in the training data negatively impacted the model's ability to handle severe cases. Specifically, in the very severe subgroup, lowering the high-AHI representation in training caused Cohen's kappa to drop from 0.365 to 0.303.

Key Contributions

1. Severity-Specific Evaluation: The study provides empirical evidence that wearable sleep staging performance is not uniform across the sleep apnea spectrum; it degrades significantly in very severe cases ($AHI \ge 50$).
2. Granularity Optimization: It demonstrates that adopting a coarser staging scheme (4-stage) can mitigate performance gaps between controlled lab environments and real-world clinical settings.
3. Data Composition Insight: The research highlights that training models without sufficient representation of the target severity spectrum (very severe apnea) leads to suboptimal performance in that specific population, emphasizing the need for severity-aware model development.

Significance

This work underscores the limitations of current wearable sleep staging technologies when applied to the most clinically severe patients without tailored training data. The findings suggest that for wearables to be viable for long-term management of severe sleep apnea:

• Training datasets must explicitly include and represent very severe cases.
• The granularity of sleep staging should be selected based on the intended clinical use case (e.g., 4-stage may be more robust than 5-stage for severe populations).
• Future validation efforts should focus on multi-night home monitoring with reliability cues to ensure clinical utility in real-world, high-severity scenarios.