Obstructive sleep apnea (OSA) is a respiratory disorder highly correlated with severe cardiovascular diseases that has unleashed the interest of hundreds of experts aiming to overcome the elevated requirements of polysomnography, the gold standard for its detection. In this regard, a variety of algorithms based on heart rate variability (HRV) features and machine learning (ML) classifiers have been recently proposed for epoch-wise OSA detection from the surface electrocardiogram signal. Many researchers have employed freely available databases to assess their methods in a reproducible way, but most were purely tested with cross-validation approaches and even some using solely a single database for training and testing procedures. Hence, although promising values of diagnostic accuracy have been reported by some of these methods, they are suspected to be overestimated and the present work aims to analyze the actual generalization ability of several epoch-wise OSA detectors obtained through a common ML pipeline and typical HRV features. Precisely, the performance of the generated OSA detectors has been compared on two validation approaches, i.e., the widely used epoch-wise, $k$ -fold cross-validation and the highly recommended external validation, both considering different combinations of well-known public databases. Regardless of the used ML classifiers and the selected HRV-based features, the external validation results have been 20 to 40% lower than those obtained with cross-validation in terms of accuracy, sensitivity, and specificity. Consequently, these results suggest that ML-based OSA detectors trained with public databases are still not sufficiently general to be employed in clinical practice, as well as that larger, more representative public datasets and the use of external validation are mandatory to improve the generalization ability and to obtain reliable assessment of the true predictive power of these algorithms, respectively.

Obstructive sleep apnea (OSA) represents a significant health concern. While polysomnography (PSG) remains the gold standard, its resource-intensive nature has encouraged the exploration of further alternative approaches. Most of these were based on the heart rate variability (HRV) analysis, but only a few of them have presented a recurrence-based approach. The present paper addresses this gap by integrating convolutional neural networks (CNNs) with HRV recurrence analysis. Employing three different and publicly available databases from PhysioNet’s official repository (Apnea-ECG, MIT-BIH, and UCD-DB), the presented method was able to expose concealed patterns within the distance matrix of HRV’s phase space, which is discernible at an appropriate level of abstraction through CNNs. Under the challenging context of external validation (MIT-BIH and UCD for training, and Apnea-ECG for testing), the results obtained were comparable to those presented in the state of the art, achieving a peak accuracy of 75%, while maintaining balanced sensitivity and specificity at 74% and 75%, respectively. Moreover, these results obtained by the proposed CNN-based recurrence analysis of HRV also outperformed traditional time–frequency models, which have yielded values of accuracy lower than 65%. Hence, this paper highlights the importance of the proposed method in gaining new insights into OSA’s HRV dynamics, offering a contribution that adds to the existing analytical approaches in the state of the art.

This study introduces and evaluates a novel nonlinear technique for synchronizing electrocardiogram (ECG) and photoplethysmogram (PPG) signals by embracing the intrinsic relationship between heart rate variability (HRV) and pulse rate variability (PRV). The proposed method utilizes normalized mutual information (NMI) derived from cross-distance matrices (CDM) of HRV and PRV, alongside the distance matrix of a reference HRV extract. We tested the methodology on two databases containing simultaneous ECG and PPG signals, including one publicly available on Physionet, and benchmarked it against various synchronization techniques, both linear (Pearson coefficient, dynamic time warping) and nonlinear (conventional mutual information, NMI with recurrence plots, cross-recurrence plot, and determinism). Results showed that the proposed method (NMI of CDM) outperformed all others, achieving synchronization rates near 50% within a 1.2-second lag threshold. This study also comprehensively examines how signal quality and recording methodology variations affect synchronization outcomes, confirming the importance of lag threshold adjustment for real accuracy assessment. However, some limitations must be kept in mind: the proposed approach is not suited for blood pressure estimation based on HRV-PRV differences, nonlinear methods generally require higher computational resources than linear ones, and further validation is needed with additional databases from real-world scenarios similar to those used in this study.