Multi-class adaboost to detect Sleep Apnea-Hypopnea Syndrome severity from oximetry recordings obtained at home

Abstract

--This paper aims at evaluating a novel multiclass methodology to establish Sleep Apnea-Hypopnea Syndrome (SAHS) severity by the use of single-channel athome oximetry recordings. The study involved 320 participants derived to a specialized sleep unit due to SAHS suspicion. These were assigned to one out of the four SAHS severity degrees according to the apnea-hypopnea index (AHI): no-SAHS (AHI<5 events/hour), mild-SAHS (5≤AHI<15 e/h), moderate-SAHS (15≤AHI<30 e/h), and severe-SAHS (AHI≥30 e/h). A set of statistical, spectral, and non-linear features were extracted from blood oxygen saturation (SpO2) signals to characterize SAHS. Then, an optimum set among these features were automatically selected based on relevancy and redundancy analyses. Finally, a multi-class AdaBoost model, built with the optimum set of features, was obtained from a training set (60%) and evaluated in an independent test set (40%). Our AdaBoost model reached 0.386 Cohen's kappa in the four-class classification task. Additionally, it reached accuracies of 89.8%, 85.8%, and 74.8% when evaluating the AHI thresholds 5 e/h, 15 e/h, and 30 e/h, respectively, outperforming the classic oxygen desaturation index. Our results suggest that SpO2 obtained at home, along with multiclass AdaBoost, are useful to detect SAHS severity.

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

“…Thus, we finally propose a comprehensive assessment of the athome SpO 2 usefulness by training and validating up to four new machine-learning derived models, ranging from simple to complex ones: linear discriminant analysis (LDA), logistic regression (LR), multi-layer perceptron Bayesian neural network (BY-MLP), and the ensemble learning method adaptive boosting (AdaBoost), arranged along with LDA as base classifiers (AB-LDA). Preliminary studies of our own group have been already conducted regarding BY-MLP and AdaBoost, providing signs of the usefulness of the machinelearning approach for at-home SpO 2 recordings [26], [27]. Nonetheless, the comprehensive approach conducted in the current study has led to the use of different SpO 2 information to train the models, as well as LDA instead of classification and regression trees as base classifiers for AdaBoost.…”