0626 A METHOD FOR IDENTIFICATION OF INSPIRATORY FLOW LIMITATION USING RESPIRATORY AIRFLOW

Inspiratory flow limitation (IFL) is a common component of sleep-disordered breathing, but current methods of identification require catheterization of the pharynx. The objective of this study was to develop and validate an accurate, non-invasive method of identifying IFL. Participants with high upper airway resistance underwent a full-night polysomnogram wearing temporary dental trays connected to a remotely controlled mandibular positioner (MATRx). A pressure transducer connected to a saline-filled nasopharyngeal catheter measured supra-glottic pressure, and respiratory airflow was calculated from naris pressure. These two signals were processed by an auto-labeler (AL) which identified IFL breaths that served as a “gold standard”. A neural network (NN) was trained on six patients using airflow as input and the gold standard as ideal output. The trained NN was then validated on three new patients using airflow as the sole input with no ideal output. In training, the AL identified 17652 breaths (61%) as IFL and 11220 (39%) as non-IFL; in validation, the AL identified 8123 breaths (59%) as IFL and 5650 (41%) as non-IFL. A 5-fold cross validation on the training data yielded an area under ROC curve of 0.86. The area under the ROC using the trained NN on all validation breaths was 0.89, and this increased to 0.90 when equivocal breaths, comprising 10% of the total, were excluded. A non-invasive method of identifying IFL was developed by training a NN using a “gold standard” and respiratory airflow. When applied to a new population of patients, the trained NN, using airflow alone, showed reasonable accuracy in identifying IFL breaths. This was improved by excluding equivocal breaths, yielding an area under ROC curve of 0.90 on 90% of all breaths. The authors acknowledge Alberta Innovates - Technology Futures, NRC-IRAP, and Zephyr Sleep Technologies for supporting this research.