Evaluation of a Sound-Based Deep Learning Model with Polysomnography in a patient with Obstructive Sleep Apnea using Positive Airway Pressure Therapy

Oct 10, 2023
Chest 2023

Abstract

Introduction:

Since noise occurs during the use of positive airway pressure (PAP) therapy, it may be difficult to analyze sleep using a sound-based deep learning algorithm that does not require any wearable or nearable device. The purpose of this study is to determine whether the AI model shows comparable validity during the use of PAP therapy by comparing the accuracy of sleep stage and apnea/hypopnea prediction between a sound-based deep learning model and polysomnography (PSG) in one patient with obstructive sleep apnea (OSA) using PAP therapy.


Case Presentation:

We retrospectively analyzed data from a patient diagnosed with OSA (AHI 70.9) who underwent baseline PSG and split-night PSG with PAP titration at a sleep clinic. We compared the accuracy of sleep stage and apnea/hypopnea prediction between the AI algorithm and PSG. During the sleep test, recorded sound was analyzed using an AI algorithm. The performance of the model was assessed by epoch-by-epoch prediction accuracy. During baseline PSG and split-night PSG, the macro F1 scores of the 4-class sleep stage (Wake, Light, Deep, and REM sleep) were 0.597 and 0.646, respectively. Epoch-by-epoch agreement with PSG results of sleep stage was 68.6% and 69.3%. At baseline PSG, the model had an accuracy of 68.8% for wake, 67.4% for light sleep (N1 and N2), 79.0% for deep sleep (N3), and 65.6% for REM sleep. On the other hand, these numbers were 79.1%, 64.9%, 81.5% and 69.0%, respectively, from the split-night PSG with PAP titration. For the 2-class respiratory event detection task (no event, apnea/hypopnea), our model achieved 64.3% and 73.4% epoch-by-epoch agreement with PSG results, respectively, at baseline PSG and split-night PSG, with macro F1 scores of 0.64 and 0.72 achieved.

Discussion:

This AI model does not require any devices other than a mobile device to measure sound. The use of this AI model during PAP therapy could reduce the size, power consumption, and complexity of PAP machines, improving patient comfort and adherence. AI models can provide a more comprehensive user profile for data analysis and personalized therapy. Lastly, this sound-based deep learning algorithm works well even in situations where noise may occur during the use of PAP therapy.

Conclusions:

The AI model showed comparable accuracy in sleep stage and apnea/hypopnea prediction to PSG in patients with OSA using PAP therapy. The AI algorithm could be a useful tool during PAP model, which may lead to improved PAP adherence and potential health outcomes.

Authors

Taeyoung Lee
Kwang Su Cha
Seunghun Kim
Moonsik Keum
AHRUM SIM
Daewoo Kim
Dongheon Lee
Clete Kushida, MD, PhD
Jeong-Whun Kim
Younghoon Cho

Acknowledgments