0311 Automated Apnea and Hypopnea Event Detection Using Deep Learning

Abstract

Introduction: Accurate apnea and hypopnea event detection in polysomnography (PSG) is important to the diagnosis of sleep apnea. Automated PSG event detection has the potential to reduce the labor costs and variability experienced with human scorers. Deep learning is a form of machine learning that uses layers of perceptrons (probabilistic classifiers) to learn hierarchical structures and dependencies within the data. Deep learning systems can build highly accurate classifiers by learning intrinsic rules over many examples rather than by programming individual event features. Method(s): This study performed a retrospective evaluation of scored PSG data from the Sleep Heart Health Study. Two deep learning sleep staging classifiers were developed using a convolutional neural network model. The training set was composed of approximately 10,000 30-second PSG epochs labeled with the occurrence of apnea or hypopnea events from 100 randomly selected patients. Data from the abdominal and thoracic respiratory inductance plethysmography channels and the oronasal thermistor served as input to the apnea classifier. The hypopnea classifier used the same inputs as well as the electroencephalography and blood oxygen saturation channels. A hold-out set of PSG epochs from 10 randomly selected non-training patients was used to assess model performance with area under the curve of the receiver operating characteristic curve (ROC AUC) and accuracy. The two classifiers were combined to calculate the apnea-hypopnea index (AHI) for each patient. Result(s): The convolutional neural network classifier for apnea events demonstrated an AUC of 0.93 and accuracy of 97%. The convolutional neural network classifier for hypopnea events demonstrated an AUC of 0.80 and accuracy of 80%. The mean squared error for the model's predicted AHI for the patients in the hold-out set was 3.541. Conclusion(s): The deep learning sleep classifiers for apneas and hypopneas demonstrated excellent accuracy for apnea detection and fair accuracy for hypopnea detection. They achieve comparable results to studies performed in literature for automated sleep event detection with models that do not require hand-engineered features.