Abstract
Obstructive sleep apnea (OSA) is a common, but severely under-diagnosed sleep disorder characterized by repeated periods of reduced or paused breathing during sleep. Polysomnography (PSG) is the gold standard for the diagnosis of OSA, but requires overnight monitoring in a sleep laboratory, which is both resource-demanding and uncomfortable for patients. The main objective of the CESAR project is to increase the percentage of detected/diagnosed OSA patients with the use of cheaper sleep monitor solutions that can be applied at home. This should be achieved by utilizing low-cost consumer electronics instead. Which in turn should reduce the time and resources needed, as no PSG in a sleep laboratory is necessary for the initial diagnostic step. The idea is to use machine learning for automatic classification of OSA to eliminate the need for sleep experts in the first step. In this thesis, we evaluate the usefulness of a low-cost strain-gauge respiratory effort sensor (FLOW) from SweetZpot for overnight monitoring of sleep apnea in a home respiratory polygraphy study. During the A3 study from Oslo University Hospital, we collect 57 FLOW and NOX recordings from 34 patients diagnosed with atrial fibrillation. We measure the signal quality produced by FLOW against the respiratory inductance plethysmography (RIP) NOX T3 sensor from NOX Medical by evaluating the classification performance of several machine learning models. This process requires the signal data from both sensors to be synchronized to utilize annotated scoring from a sleep expert. Our analysis reveals several data quality issues with FLOW related to connection loss, unreliable timestamping- and sampling rate, which proves to be a non-trivial problem to correct for synchronization. We discuss several approaches for timestamp adjustment and finally design a flexible window model that can both identify connection loss and adjust timestamps. We evaluate the FLOW recordings on a window-based approach to validate the timestamp adjustment (synchronization) and to analyze how the signal quality changes overnight. The signal quality evaluation is based on the breath detection accuracy metrics sensitivity, positive predictive value (PPV), and clean minute proportion (CMP), along with the breath amplitude accuracy metric weighted absolute percentage error (WAPE). These metrics are similar to how apnea and hypopnea events are scored by medical personal. We achieve a sensitivity, PPV, CMP, and WAPE score of 97.2%, 94.2%, 59.4%, and 18.4%, respectively, indicating that our preprocessing is sufficient for mitigating the original data quality issues. Our results show that common behaviors during sleep, such as movement and changes in sleeping position, significantly affect the signal quality produced by FLOW, which we attribute to belt entrapment or misplacement. We are able to significantly increase the machine learning classification performance on FLOW data by applying a simple standardization. Using ten-fold-cross-validation, we achieve a classification accuracy of 76.1% using convolutional neural network. Our improvements suggest that preprocessing of the data results in better classification accuracy. For comparison, we achieve a classification accuracy of 79.6% on NOX.