Abstract
Introduction
Evaluation of sleep apnea involves manual annotation of Polysomnography (PSG) file, a time-consuming process subject to interscorer variations. The DOSED algorithm has been shown to be helpful in detecting Central Sleep Apnea (CSA), Obstructive Sleep Apnea (OSA), and Hypopnea when merged into a single event type. This work uses a modified version of DOSED capable of detecting each event type separately.
Methods
The network consists of 3 blocks of 1D convolutional layers followed by 6 blocks of 2D convolutional layers. The network has 2 classification layers, one determines the probability of each class, and the other determines the start and duration time of the event with the highest probability. Four channels from nasal and mouth airflow and position of abdomen and thorax are used as input to the model. The model was trained using 2800 PSG from 4 different cohorts (MESA, MROS, SSC, WSC) and tested on 70 PSG, which have been scored by six technicians (Stanford, U Penn, St Louis).
Results
On an event by event basis, model F1-scores versus a weighted consensus score based on 6 technicians were 0.60 for OSA, 0.43 for CSA, and 0.34 for Hypopnea. Average F1-scores for the 6 technicians were 0.48 (std 0.04) for OSA, 0.29 (std 0.145) for CSA, and 0.54 (std 0.183) for Hypopnea, indicating that the model functions better on an event-by-event basis than an average technician. Correlations between indices/hr for central apnea, obstructive apnea, and hypopnea indicate excellent correlations for apneas, but poor correlation for hypopnea. We are now adding the snoring channel to explore if predictions can be improved.
Conclusion
The result shows that deep learning-based models can detect respiratory events with an accuracy similar to technicians. The poor agreement between technicians from different universities indicates that we need better definitions of hypopnea.
Support
A pacemaker transthoracic impedance sensor with an advanced algorithm to identify severe sleep apnea: The DREAM European study