Classification of Central and Obstructive Sleep Apnea Using Respiratory Inductance Plethysmography and Convolutional Neural Networks

In this study, we employ a pre-trained Convolutional Neural Network (CNN), specifically ResNet18, for automated sleep apnea classification using spectral images derived from Respiratory Inductance Plethysmography (RIP) signals. Using data from 130 patients who underwent diagnostic polysomnogram (PSG) at the Sleep Disorders Clinic of the Royal Ottawa Mental Health Centre, we compare the performance of three models fine-tuned using spectrograms of 60 sec segments of the RIPsum, RIPflow, and the concurrent thoracic (THO) and abdominal (ABD) RIP signals. In the case of the THO/ABD model, concurrent spectrograms of the THO and ABD signals are combined into a single image. The resulting models classified any apnea events (central and obstructive apnea/hypopnea) from normal respiration accuracy ranging from 87-90% and F1 scores ranging 90-91%. We also examine the models' performance on classification of central sleep apnea, obstructive sleep apnea, and normal respiration. The THO/ABD model classified central apnea events with 89.4% precision, 90.2% recall, 89.8% F1 score and 94.4% specificity. Our results show that RIP belts can be used as an effective screening tool for sleep apnea, and central sleep apnea more specifically, utilizing smaller datasets and minimal training. This approach may be useful for non-invasive, cost-effective diagnostic methodologies at home and in clinical environments.