Machine Learning Methods to Detect Obstructive Sleep Apnea under a Robust Testing Framework

The obstructive sleep apnea (OSA) detection problem has been addressed through different approaches to overcome the limitations of polysomnography, the official clinical gold standard. On this matter, the combination of heart rate variability with modern machine learning (ML) techniques has become quite popular in recent years. Indeed, the evolution in performance of the models reported in the literature has been promising; however, clinical conventions have been urging the scientific community to employ more robust testing frameworks oriented to clinical practice, since most of such models had not been tested with newly unseen data. Hence, the present work follows the guidelines found in the state of the art to build a set of ML-based models and evaluate their performance in more real-life scenarios. More specifically, three publicly available databases were employed to assess the generalization capability of five different classifiers, i.e., support vector machines, k-nearest neighbors, decision trees, and two random forest-based algorithms. Eventually, although the obtained results were different from those reported in the corresponding literature, the experiments conducted in the present work have revealed interesting findings on further insights that could improve the clinical value of the modern ML-based OSA classifiers.