Computer aided diagnosis of pediatric sleep apnea through the analysis of airflow and oximetry signalsfrom ensemble learning to explainable deep learning algorithms

Abstract

Obstructive sleep apnea (OSA) is a sleep disorder in which intermittent obstruction or narrowing of the upper airway causes recurrent pauses and cessations of normal respiration (apneas and hypopneas, respectively) during sleep. The most common symptoms of OSA in chidren are snore, labored respiration or breathing pauses, and daytime hypersomnolence. Pediatric OSA affects 1% - 5% of children, with several negative consequences that range from cardiometabolical comorbidities to neurobehavioral disorders. The gold standard in the diagnosis of childhood OSA is the polysomnography (PSG), a sleep study that involves the recording of several biomedical signals in a sleep laboratory. Pediatric OSA is diagnosed and quantified by computing the apnea-hypopnea index (AHI), the rate of apneas/hypopneas per hour (e/h) of sleep: no OSA (AHI < 1 e/h), mild OSA (1 < AHI < 5 e/h), moderate OSA (5 < AHI < 10 e/h), and severe OSA (AHI > 10 e/h). However, PSG has low availability due to scarce sleep units, high complexity, and associated costs. Alternatives to PSG usually comprise the analysis of less signals, and simplified tests involving the analysis of airflow (AF) and oximetry (SpO2) are a suitable alternative. AF and SpO2 analyses can be automated by means of signal processing and pattern recognition algorithms that can simplify the diagnosis of pediatric OSA. In this doctoral thesis, different algorithms such as ensemble learning (AdaBoost) and deep learning (DL) combined with explainable artificial intelligence (XAI) methods are proposed to facilitate the diagnosis of pediatric OSA through the automatic analysis of AF and SpO2 signals. The main objective of this doctoral thesis was to study, develop and validate ensemble learning and DL methods together with new XAI techniques in the context of automatic analysis of AF and SpO2 signals, so that these methods can be used to help diagnose pediatric OSA. The characterization of AF and SpO2 revealed that both signals exhibited relevant and complementary information, which enabled AdaBoost to reach the highest diagnostic performance in comparison with single signal approaches. These results suggest that AF and SpO2 can be complementary and useful together to detect pediatric OSA. On the other hand, the DL architectures outperformed the previous approaches in terms of diagnostic performance. A combination of convolutional and recurrent neural networks surpassed all the previous approaches focused on detecting pediatric OSA. Regarding XAI analysis, heatmaps derived from the Gradient-weighted Class Activation Mapping (Grad-CAM) algorithm revealed that this DL model focused in AF amplitude changes and desaturations to detect and quantify OSA. These heatmaps also revealed that the model usually misses some hypopneas associated to arousals. The use of Grad-CAM enabled the discovery of relevant OSA-related patterns that are automatically detected by the DL algorithms and can aid users to reinforce their trust in DL models. In view of these results, the methods proposed in this doctoral thesis could be used to develop a screening test of pediatric OSA that would alleviate the waiting lists of pediatric sleep laboratories. The results achieved by the methods proposed in this research allow us to conclude that the automatic analysis of AF and SpO2 based on ensemble and DL methods combined with XAI have demonstrated a remarkable diagnostic usefulness, and can be used to deploy alternative, simple, reliable and trustworthy screening methods to serve as an aid in the diagnosis of OSA in children.