Automated analysis of the oximetry signal to simplify the diagnosis of pediatric sleep apneafrom feature-engineering to deep-learning approaches

Resumen

Obstructive sleep apnea (OSA) is a high prevalent respiratory disorder in the pediatric population (1%-5%). Untreated pediatric OSA is associated with significant adverse consequences affecting metabolic, cardiovascular, neurocognitive, and behavioral systems, thus resulting in a decline of overall health and quality of life. Consequently, it is of paramount importance to accelerate the diagnosis and treatment in these children. Overnight polysomnography (PSG) is the gold standard to diagnose OSA in children. This test requires an overnight stay of pediatric subjects in a specialized sleep laboratory, as well as the recording of up to 32 biomedical signals. These recordings are used to quantify respiratory events in order to obtain the apnea-hyponea index (AHI), which is used to establish pediatric OSA severity. Nonetheless, PSG is technically complex, time-consuming, costly, highly intrusive for the children, and relatively unavailable, thus delaying the access for both the diagnosis and treatment. Consequently, simplified diagnostic techniques become necessary. In an effort to overcome these drawbacks and increase the accessibility of pediatric OSA diagnosis, many simplified alternative procedures have been developed. Among these, a common approach is the analysis of the blood oxygen saturation (SpO2) signal from overnight oximetry due to its easy acquisition and interpretation, as well as its suitability for children. Many studies have demonstrated the utility of the automated analysis of SpO2 recordings to help in adult OSA diagnosis. Conversely, the preceding studies focused on pediatric patients reported lower accuracies than those reached in the case of adults, suggesting the need to seek novel signal processing algorithms that provide additional information from the SpO2 signal for the particularities of childhood OSA. In the present Doctoral Thesis, we hypothesize that the application of novel feature extraction and deep-learning algorithms could increase the diagnostic ability of the oximetry signal in the context of pediatric OSA. Consequently, the general objective of this Doctoral Thesis is to design, develop, and assess novel clinical decision-support models in the context of childhood OSA based on the automated analysis of the SpO2 signal. To achieve this goal, 3196 SpO2 recordings from three different databases of children were involved: (i) the Childhood Adenotonsillectomy Trial (CHAT) database, (ii) the University of Chicago (UofC) database, and (iii) the Burgos University Hospital (BUH) database. These recordings were automatically analyzed using feature-engineering and deep-learning methodologies. On one hand, feature-engineering methodologies were conducted in three phases. First, a set of OSA-related features were extracted from the SpO2 signal using different analytical approaches: statistical parameters, conventional oximetric indices, frequency domain methods, and nonlinear analysis. Particularly, we have evaluated the usefulness of bispectrum, wavelet, and detrended fluctuation analysis (DFA) to provide additional and complementary information to conventional approaches linked to pediatric OSA and its severity. As a second step, the fast correlation-based filter algorithm was applied to select optimum subsets of features that provide relevant and non-redundant information related to pediatric OSA and its severity. Finally, pattern recognition algorithms were applied to these optimum subsets of features in order to estimate pediatric OSA and its severity. To this effect, different approaches were explored: binary (OSA negative vs. OSA positive) and multi-class (OSA severity degrees) classification and regression (estimation of the AHI). On the other hand, a deep-learning methodology based on convolutional neural networks (CNN) was employed to automatically estimate pediatric OSA severity from raw oximetry data. A high performance was obtained with both the proposed feature-engineering and deep-learning approaches. Thus, in the case of feature-engineering, our results showed that the application of bispectrum, wavelet, and DFA allowed to obtain features that provide relevant and complimentary information to conventional methods regarding OSA-related changes in the oximetry signal. Specifically, a multiclass multi-layer perceptron (MLP) neural network was fed with an optimum subset composed of the mean amplitude of the bispectrum, the mean of the bispectrum invariant, variables from the power spectral density (PSD), the 3% oxygen desaturation index (ODI3), and anthropometric variables. This MLP model reached 81.3% and 85.3% accuracy (Acc) in the AHI cutoffs of 5 (moderate OSA) and 10 (severe OSA) events per hour (e/h), respectively, outperforming a MLP model trained without bispectral features. In addition, a binary support vector machines (SVM) model was trained with an optimum subset composed of the skewness and energy of the wavelet coefficients in the 9th detail level and the wavelet entropy, together with ODI3, statistical moments in the time domain and PSD-derived parameters. This optimum SVM model showed a high capability as a screening tool to detect moderate-to-severe pediatric OSA (AHI >= 5 e/h), with 84.0% Acc and a positive likelihood ratio of 14.6, which are higher than the obtained with every single feature. Finally, a regression MLP model trained with a subset of features composed of the ODI3 and the slope in the first scaling region of the DFA obtained 82.7%, 81.9%,and 91.1% Acc for the AHI cutoffs of 1 e/h, 5 e/h, and 10 e/h, respectively. This regression MLP model outperformed the conventional ODI3, commonly used in clinical settings. On the other hand, it was found that deep-learning approaches can automatically learn additional information from the oximetry signal linked to apneic events. A CNN-based deep-learning architecture trained to estimate the AHI from raw SpO2 segments reached 0.515, 0.422, and 0.423 Cohen’s kappa in three independent datasets (CHAT, UofC, and BUH). In addition, the proposed CNN-based model reached high accuracies for the AHI severity cutoffs of 1 e/h (77.6%, 80.1%, and 79.2%), 5 e/h (97.4%, 83.9%, and 83.5%), and 10 e/h (97.8%, 92.3%, and 91.3%) in the CHAT, UofC, and BUH datasets. This CNN-based model achieved a higher overall performance than feature-engineering approaches. The application of this deep-learning model as a screening protocol would avoid the need for 73.7% (CHAT), 50.0% (UofC), and 45.9% (BUH) of full PSGs in pediatric subjects. Our proposed methodologies also achieved a higher overall performance than state-of-the-art studies, especially for moderate-to-severely affected pediatric subjects. Therefore, the results obtained in this Doctoral Thesis suggest that bispectrum, wavelet, and DFA are able to further characterize changes in the SpO2 signal caused by apneic events in pediatric subjects. Furthermore, it is also concluded that deep-learning algorithms can learn complex features from oximetry dynamics that allow to enhance the diagnostic capability of nocturnal oximetry in the context of childhood OSA. We feel that these studies could contribute to the use of clinical screening tools to diagnose pediatric OSA based on the automated analysis of the oximetry signal, aiming at providing an early and timely diagnosis and treatment of the affected children.