Automatic analysis of overnight airflow to help in the diagnosis of pediatric obstructive sleep apnea

Abstract

Obstructive sleep apnea (OSA) is a sleep-related breathing disorder characterized by presenting recurrent oronasal airflow cessations. These breathing cessations can be complete (apnea events) or partial (hypopnea events). The presence of these respiratory events can adversely affect the physiological and cognitive functions of children. In this regard, OSA can cause serious long-term neurocognitive deficiencies, behavioral disorders, as well as cardiovascular, metabolic and endocrine dysfunctions, drastically reducing their health and quality of life. Consequently, it is of the utmost importance that children are timely diagnosed and treated to prevent the negative consequences associated to OSA. Pediatric OSA has a high prevalence, since it affects 5.7% of children between 2 and 18 years of age. According to the primary care clinical data requested from the Subdirección General de Información Sanitaria Española in March 2021, it is estimated that this disease affects 9.56‰ of children under 15 years in Spain. Despite its high prevalence, OSA is an underdiagnosed disease, estimating that 90% of affected children still do not have a medical diagnosis. In order to diagnose it, the subjects are referred to a specialized pediatric sleep unit where they undergo nocturnal polysomnography (PSG). This sleep study is based on simultaneously recording several neurophysiological and cardiorespiratory signals while the child sleeps. After, these recordings are visually inspected by sleep medical specialists for manual scoring of respiratory events and computing the apnea-hypopnea index (AHI). In children, this index is used to determine the presence and severity of OSA according to the thresholds 1, 5, and 10 events/h (e/h). PSG is effective, but also uncomfortable to children, complex, time-consuming, and relatively unavailable, which lead to long waiting lists and diagnostic delays. Then, great efforts have been made to search and develop simpler alternative methods that help diagnose pediatric OSA. In this regard, several studies have focused their research on automatically analyzing a minimum set of cardiorespiratory signals involved in PSG. In this Doctoral Thesis, we propose to exhaustively characterize the behavior of nocturnal airflow (AF) in children to obtain relevant and useful information that helps to simplify the pediatric OSA diagnosis. This signal reflects the respiratory activity during sleep time, including the breathing pauses associated with OSA. In addition, AF can be easily acquired at the patient’s home using a portable monitoring device with built-in thermistor. Thereby, AF analysis is a promising way to simplify the diagnosis of childhood OSA. Thus, we hypothesize that the characterization of overnight AF by means of novel approaches can help and simplify pediatric OSA diagnosis. Accordingly, the main objective of this Doctoral Thesis is to design, implement, and assess novel automatic signal processing methods that allow exhaustively characterizing the overnight AF from children and helping in the pediatric OSA diagnosis. In order to achieve this goal, a four-stage methodology is proposed. Firstly, the recordings were subjected to a pre-processing stage to resample them and automatically remove noise and artifacts. Moreover, AF signals were standardized to minimize the effects caused by particular features unrelated to OSA. It would improve the quality of the AF recordings and would increase the effectiveness of subsequent analysis. Afterwards, a feature extraction stage was performed to comprehensively characterize the behavior of pediatric overnight AF by means of different techniques. In this regard, cardiorespiratory signals, and therefore AF, are dynamic, non-linear, and non-stationary. Consequently, non-linear, spectral, bispectral, recurrence plot (RP), and wavelet analyses have been conducted for adapting to the intrinsic properties of overnight AF and getting useful OSA-related information from it. The features derived from each of these methodological approaches could provide redundant information about the AF behavior. Thus, a feature selection stage has been applied to identify those features that provide relevant and complementary information, maximizing the diagnostic ability of AF. In this regard, forward stepwise logistic regression (FSLR) wrapper method and fast correlation-based filter (FCBF) method were used for this purpose. Finally, supervised machine-learning techniques have been applied to recognize patterns in AF features, infer behaviors from them, and use this information to automatically detect the presence and severity of OSA in children. This stage was conducted from three different approaches: discrimination between OSA-negative and OSA-positive pediatric subjects (binary classification task), classification of children according to their OSA severity degree (multiclass classification task), and AHI estimation of each child (regression task). The binary and multiclass classification tasks were performed by means of logistic regression (LR) and adaptive boosting (AdaBoost.M2) algorithms, respectively. Regarding the regression task, it was performed through a multi-layer perceptron neural network (MLP) and a MLP with Bayesian approach (BY-MLP). In addition, the 3% blood oxygen desaturation index (ODI3), a clinical parameter used as a suboptimal alternative to PSG, was incorporated to the study. This allowed us to evaluate its complementarity with the information obtained from AF through the different methodological approaches. Each of proposed characterization approaches enabled us to uncover behaviors of pediatric nocturnal AF that were previously unknown in OSA context. In this regard, the central tendency measure and spectral entropies showed that this disease reduces the variability and increases the irregularity of pediatric AF. The characterization conducted by means of RP-derived features revealed that OSA modifies the underlying dynamics and the phase-space of AF. Concretely, the occurrence of apneic events decreases the variability, the stationarity, and the complexity of AF signal, as well as the exponential divergence of its phase-space. Moreover, it also increases the dwell time at a certain phase state of AF (i.e., it does not change, or changes very slowly), its average prediction time, and its irregularity. In the case of the bispectral features, they showed that OSA reduces the non-gaussianity of AF, as well as the non-linear interaction of its harmonic components. Childhood OSA also decreases the phase coupling in the normal breathing band, shifting the coupling focus towards low frequency components related to the occurrence apneic events. In addition, the irregularity of AF signal increases in terms of amplitude and phase when the OSA severity is higher. Regarding the wavelet features, they revealed that OSA disturbs the energy distribution and the frequency components of AF signal. Concretely, apneic events reduce the AF detail signal amplitude and the energy produced in the normal breathing band. The frequency components of AF decrease, while its irregularity increases in terms of energy as the AHI is higher. In addition, it was observed that the information provided by AF through the different methodological approaches is complementary to the information from the classic ODI3. This complementarity was not only manifested in the selection stage, but also in the pattern recognition stage. In this regard, moderate-to-high accuracies (Acc) were achieved by the predictive models fed only with AF features: 60.0% – 81.1% for 1 e/h, 57.1% – 76.0% for 5 e/h, and 70.5% – 80.6% for10 e/h. However, significantly higher diagnostic accuracies were obtained when AF features and ODI3 were combined: 78.0% – 83.2% for 1 e/h, 78.5% – 82.5% for 5 e/h, and 90.2% – 91.0% for 10 e/h. Thereby, RP from AF signal and ODI3 achieved the highest Acc for 1 e/h (83.2%). Moreover, this approach obtained the lower negative likelihood ratio (LR– = 0.1), which is considered as a reliable clue to confirm the disease absence when it is ≤ 0.1. In the case of 5 e/h, the highest Acc was achieved by the bispectral analysis of AF and the ODI3 (82.5%). Regarding the AHI threshold 10 e/h, both RP and wavelet features from AF obtained 91.0% Acc in combination with the ODI3. However, AdaBoost model reached a remarkably higher positive likelihood ratio using wavelet features (LR+= 19.0), which is considered as a strong inkling to confirm the disease presence when it is ≥ 10. Based on the aforementioned considerations, the different methodological approaches proposed in this Doctoral Thesis allow adapting to the intrinsic properties of pediatric overnight AF, characterizing its behavior, and providing useful OSA-related information. These approaches enhance the ability of automatic AF analysis to determine the presence and severity of OSA in children. In this regard, the predictive models based on RP, bispectrum, and wavelet features obtained a high overall diagnostic performance along with the ODI3, outperforming other state-of-the-art studies and conventional approaches previously applied in adults. Thus, we can conclude that the characterization of overnight AF by means of these novel methods can help to simplify the OSA diagnosis in children. In addition, the high performance of the proposed models suggests that they could be incorporated into clinical practice as reliable automatic screening methods for pediatric OSA.