Automatic Respiratory Event Scoring in Obstructive Sleep Apnea Using a Long Short-Term Memory Neural Network

Area of Science:

• Sleep Medicine
• Artificial Intelligence
• Biomedical Engineering

Background:

• Obstructive sleep apnea (OSA) diagnosis relies on manual scoring of respiratory events from polysomnography, a process that is time-consuming and costly.
• Automated scoring methods offer a potential solution to enhance diagnostic efficiency and reallocate resources within sleep medicine.

Purpose of the Study:

• To develop and evaluate a long short-term memory (LSTM) neural network for the automatic scoring of respiratory events in sleep apnea diagnosis.
• To assess the accuracy and agreement of the automated scoring system compared to manual scoring methods.

Main Methods:

• An LSTM neural network was trained using polysomnography data, including peripheral blood oxygen saturation, airflow, and respiratory effort signals.
• The network was trained on data from 787 patients and validated on an independent set of 100 patients.

Main Results:

• The automated scoring achieved high epoch-wise agreement with manual scoring (88.9%, κ = 0.728).
• The apnea-hypopnea index (AHI) calculated by the neural network closely matched the manually determined AHI (mean absolute error of 3.0 events/hour, ICC = 0.985).

Conclusions:

• The developed LSTM neural network demonstrates high accuracy and strong agreement with manual scoring for respiratory events.
• This automated approach has potential for analyzing large datasets and future clinical applications in sleep apnea diagnostics, with the option for manual review.