Respiratory Pattern-Based Detection of Abnormal Sleep Epochs Using Logistic Regression and Segment Length Analysis

Abstract

In polysomnographic examinations, sleep technologists visually analyze respiratory waveforms, also known as respiratory curves, using respiratory inductive plethysmography belts and other sensors to identify and record abnormal respiratory events, such as apnea and hypopnea. To automate this analysis process, we have been developing a method for automatically detecting abnormal events from respiratory curves. Our previous work focused on extracting respiratory motion patterns using change-point detection and clustering techniques. In this study, we extend these efforts by constructing a model to detect abnormal sleep epochs. The baseline model estimates the probability of abnormality for a target epoch using logistic regression based on the frequency of occurrence of each respiratory motion pattern. To improve the prediction performance of this model, we investigate the effectiveness of incorporating information on the segment lengths obtained by dividing the respiratory curves at detected change points.