Noncontact Sleep Monitoring With Infrared Video Data to Estimate Sleep Apnea Severity and Distinguish Between Positional and Nonpositional Sleep Apnea: Model Development and Experimental Validation

Sina Akbarian, Nasim Montazeri Ghahjaverestan, Azadeh Yadollahi, Babak TaatiJournal of Medical Internet Research 2021 · DOI

Abstract

This study presents a noncontact system for screening sleep apnea at home using infrared (IR) video and 3D convolutional neural networks (3D-CNNs). By analyzing subtle chest and abdominal movements, the model estimates the Apnea-Hypopnea Index (AHI) and distinguishes between positional and nonpositional sleep apnea, offering a scalable alternative to traditional clinical diagnostics.

Introduction

Sleep apnea is a prevalent respiratory disorder characterized by repeated interruptions in breathing during sleep. It is currently diagnosed via polysomnography (PSG), a costly and intrusive procedure involving over 20 sensors. A critical subset of patients suffers from positional sleep apnea, where events primarily occur while sleeping in the supine position. Identifying these patients is vital, as they can often be treated with simple behavioral therapies rather than continuous positive airway pressure (CPAP) machines. This research aims to bridge the gap between clinical accuracy and home accessibility using computer vision.

Methodology

The researchers developed a multi-stage pipeline to process video data captured by a single near-infrared camera:

  1. Motion Extraction: To focus on respiratory mechanics rather than visual appearance, the system uses FlowNet 2.0 to generate dense optical flow maps. These maps track the precise x and y movement of pixels, highlighting the physical rise and fall of the chest.
  2. 3D-CNN Architecture: The core of the system is a 3D-CNN trained on 18-frame windows of optical flow. Unlike 2D models, 3D-CNNs capture temporal dynamics, allowing the model to recognize the specific rhythms and pauses indicative of respiratory events.
  3. Position and Classification: A secondary CNN identifies body position (supine vs. lateral). A Random Forest classifier then integrates AHI estimates and position data to categorize the apnea type.
Comparison of movement signals (v) over time (s) for Apnea (A), Hypopnea (B), and Normal Breathing (C).

Experimental Validation

The model was validated using a dataset of 41 participants in a sleep laboratory, with simultaneous PSG providing the ground truth. The cohort represented a typical clinical population with a mean age of 53 and a mean BMI of 30.

Key Results

  • AHI Estimation: The 3D-CNN achieved a Spearman correlation of 0.79 with PSG, significantly outperforming baseline vision models (0.55).
  • Screening Accuracy: At an AHI threshold of 15 (moderate to severe apnea), the system reached 83% accuracy and an F1-score of 86%.
  • Positional Classification: The system correctly identified positional status with 65.7% accuracy, nearly doubling the performance of previous noncontact methods.

Discussion and Future Work

By prioritizing motion patterns over raw pixel data, the model remains robust against variations in patient appearance or bedding. The ability to distinguish positional apnea is a significant clinical advancement for noncontact monitoring, as it directly informs treatment strategies.

Future development will focus on expanding the dataset to include more diverse home environments and bedding types (such as heavy duvets) to ensure the system's reliability outside of controlled laboratory settings. The ultimate goal is to integrate this technology into consumer devices, providing a sophisticated medical screening tool via a smartphone or tablet app.

Summary updated

Abstract

This study demonstrates the possibility of using a camera-based method for developing an accessible and easy-to-use device for screening sleep apnea at home, which can be provided in the form of a tablet or smartphone app.
BibTeX
@article{akbarian2021noncontact,
  author = {Sina Akbarian and Nasim Montazeri Ghahjaverestan and Azadeh Yadollahi and Babak Taati},
  title = {{Noncontact Sleep Monitoring With Infrared Video Data to Estimate Sleep Apnea Severity and Distinguish Between Positional and Nonpositional Sleep Apnea: Model Development and Experimental Validation}},
  journal = {Journal of Medical Internet Research},
  year = {2021},
  doi = {10.2196/26524},
}
Figure