Ensemble-learning regression to estimate sleep apnea severity using at-home oximetry in adults.

Overnight pulse oximetry has shown usefulness to simplify obstructive sleep apnea (OSA) diagnosis when combined with machine-learning approaches. However, the development and evaluation of a single model with ability to reach high diagnostic performance in both community-based non-referral and clinical referral cohorts are still pending. Since ensemble-learning algorithms are known for their generalization ability, we propose a least-squares boosting (LSBoost) model aimed at estimating the apnea-hypopnea index (AHI), as the correlate clinical measure of disease severity. A thorough characterization of 8,762 nocturnal blood-oxygen saturation signals (SpO) obtained at home was conducted to extract the oximetric information subsequently used in the training, validation, and test stages. The estimated AHI derived from our model achieved high diagnostic ability in both referral and non-referral cohorts reaching intra-class correlation coefficients within 0.889-0.924, and Cohen's within 0.478-0.663 when considering the four OSA severity categories. These resulted in accuracies ranging 87.2%-96.6%, 81.1%-87.6%, and 91.6%-94.6% when assessing the three typical AHI severity thresholds, 5 events/hour (e/h), 15 e/h, and 30 e/h, respectively. Our model also revealed the importance of the SpO predictors, thereby minimizing the 'black box' perception traditionally attributed to the machine-learning approaches. Furthermore, a decision curve analysis emphasized the clinical usefulness of our proposal. Therefore, we conclude that the LSBoost-based model can foster development of clinically applicable and cost saving protocols for detection of patients attending primary care services, or to avoid full polysomnography in specialized sleep facilities, thus demonstrating the diagnostic usefulness of SpO signals obtained at home.