PROJECT TITLE :

Evaluation of Machine-Learning Approaches to Estimate Sleep Apnea Severity From At-Home Oximetry Recordings

ABSTRACT:

Because of the complexity, costs, and long wait times associated with sleep apnea-hypopnea syndrome (SAHS) diagnosis, a streamlined alternative is required. The blood oxygen saturation signal (SpO2) can be easily obtained through nocturnal oximetry and contains useful information about SAHS. In this work, 320 patients' SpO2 single-channel recordings were gathered at their residences and used to obtain statistical, spectral, nonlinear, and clinical SAHS-related information automatically. Relevant, nonredundant data from these investigations were then utilized to train and test four machine-learning techniques that can classify SpO2 signals into one of four degrees of SAHS severity (no-SAHS, mild, moderate, and severe). The diagnostic ability of the conventionally used 3 percent oxygen desaturation index was exceeded by all of the trained models (linear discriminant analysis, 1-vs-all logistic regression, Bayesian multilayer perceptron, and AdaBoost). The best results came from an AdaBoost model that used linear discriminants as basis classifiers. In the SAHS severity classification, it achieved 0.479 Cohen's, as well as 92.9 percent, 87.4 percent, and 78.7 percent accuracies in binary classification tasks using increasing severity thresholds (apnea-hypopnea index: 5, 15, and 30 events/hour, respectively). These findings imply that machine learning can be used in conjunction with SpO2 data collected at a patient's home to simplify SAHS diagnosis.