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

Abstract

Abstract-Complexity, costs, and waiting lists issues demand a simplified alternative for sleep apnea-hypopnea syndrome (SAHS) diagnosis. The blood oxygen saturation signal (SpO2) carries useful information about SAHS and can be easily acquired from overnight oximetry. In this study, SpO2 single-channel recordings from 320 subjects were obtained at patients' home. They were used to automatically obtain statistical, spectral, non-linear, and clinical SAHS-related information. Relevant and non-redundant data from these analyses were subsequently used to train and validate four machine-learning methods with ability to classify SpO2 signals into one out of the four SAHS-severity degrees (no-SAHS, mild, moderate, and severe). All the models trained (linear discriminant analysis, 1-vs-all logistic regression, Bayesian multilayer perceptron, and AdaBoost), outperformed the diagnostic ability of the conventionally-used 3% oxygen desaturation index. An AdaBoost model built with linear discriminants as base classifiers reached the highest figures. It achieved 0.479 Cohen's  in the SAHS severity classification, as well as 92.9%, 87.4%, and 78.7% accuracies in binary classification tasks using increasing severity thresholds (apnea-hypopnea index: 5, 15, and 30 events/hour, respectively). These results suggest that machine learning can be used along with SpO2 information acquired at patients' home to help in SAHS diagnosis simplification.

A review of approaches for analysing obstructive sleep apnoea‐related patterns in pulse oximetry data

“…This work has focused on SpO 2 data only, but incorporating information from the photoplethysmogram (i.e. pulse rate variability) and other physiological channels or utilizing home‐based studies may improve the performance and/or utility of predictive tools respectively.…”

“…SpO 2 values are capped at 100%, and consequently, most individuals show an asymmetric distribution with a tail of values below the mean. This negative tail may be extended with increasing OSA severity …”

“…The first to fourth statistical moments of the frequency spectrum are analogous to the mean, variance, skewness and kurtosis defined above for the frequency histogram . The peak in frequency content in the ‘OSA band’ leads to a broader overall spectrum with two peaks: one at frequencies approaching 0 Hz (DC) and one within the OSA band.…”

“…OSA patients are expected to have a greater variance in SpO 2 values (whereas healthy individuals are expected to have relatively consistent values) …”

“…Kurtosis quantifies the ‘sharpness’ of the distribution, and OSA patients are expected to have less sharp distribution leading to a lower value …”