Expert-level probabilistic breathing event detector informs phenotyping of sleep apnea

Area of Science:

• Sleep Medicine
• Artificial Intelligence
• Respiratory Physiology

Background:

• Manual sleep study annotation is time-consuming and variable.
• Existing automatic methods lack generalizability across different centers.
• Accurate diagnosis of sleep disordered breathing (SDB) is crucial for patient management.

Purpose of the Study:

• To develop an automatic deep learning model for localizing and classifying various apneic breathing events.
• To assess the model's performance against expert annotations and its generalizability across diverse datasets.
• To introduce a probabilistic output, "apnotyping," for deeper insights into SDB etiology.

Main Methods:

• An end-to-end deep learning architecture was trained on 5456 polysomnographies and tested on 1099 from six cohorts.
• The model detects and classifies obstructive apneas, central apneas, hypopneas, and isolated respiratory events.
• Performance was evaluated using correlation coefficients and F1 scores, compared against expert raters on independent datasets.

Main Results:

• The model achieved a strong correlation (r² = 0.84) with expert apnea-hypopnea index annotations.
• Overall F1 score was 0.78, with specific F1 scores of 0.71 (obstructive apnea), 0.51 (central apnea), and 0.65 (hypopnea).
• The model performed comparably or better than individual expert raters on independent datasets.

Conclusions:

• The developed deep learning model accurately detects and classifies SDB events, outperforming or matching expert raters.
• "Apnotyping" provides novel insights into SDB etiology, correlating better with physiological traits than traditional indexes.
• This probabilistic approach holds potential for enhancing diagnostic accuracy and guiding personalized SDB treatment strategies.