Pattern recognition applied to airflow recordings to help in sleep apnea-hypopnea syndrome diagnosis

Resumen

The sleep apnea-hyponea syndrome (SAHS) is a disease characterized by episodes of complete absence (apneas) or significant reduction (hypopneas) of breathing during sleep. The apneic events recurrence leads to inadequate gas exchange which causes hypoxia and hypercapnia, resulting in oxygen saturation drops, periodic arousals, as well as sleep fragmentation. As a consequence, SAHS patients are not able to get restful sleep, which affects their quality of life. Hypersomnolence, decrease in the short-memory function, and depression are some of the daytime symptoms reported by affected people. Additionally, SAHS has been associated with major cardiovascular and metabolic illnesses such as heart failure, stroke, sudden death, and diabetes. Recently, it has been also associated with an increase in cancer incidence. These SAHS consequences make a fast diagnosis the key action to improve health and quality of life of patients. SAHS is a very prevalent illness, affecting from 2% to 7% of adult population and up to 6% of children. It is also considered as an underdiagnosed disease, with a growing incidence due to the obesity epidemic present in developed countries. Overnight polysomnography (PSG), conducted in a specialized sleep unit, is the gold standard to diagnose SAHS. However, it is technically complex due to the high number of physiological signals to be recorded, costly due to the need for patient's hospitalization, as well as time-consuming due to the online inspection of the recordings, which is required to reach diagnosis. This is obtained by computing the apnea-hypopnea index (AHI) after carefully reviewing of the recorded signals. Moreover, PSG test deprives patients of their natural sleep environment. These drawbacks, the high prevalence of SAHS, as well as the limited availability of specialized facilities, have led to the search for new ways to simplify the diagnostic process. One common approach is the analysis of a reduced set of signals among those involved in full PSG. In this Doctoral Thesis it is posed the automatic analysis of single-channel airfow (AF) as a simple and reliable alternative to PSG. In addition, pattern recognition is proposed as the main approach to conduct an automatic SAHS diagnosis, including binary classifcation (presence or absence of SAHS) as well as determination of SAHS severity degree (multiclass classifcation and AHI estimation by means of regression). We hypothesize that it is possible to reduce the complexity of the SAHS diagnostic process by means of an automatic pattern recognition analysis of AF. Consequently, the general goal of this work is the comprehensive study and assessment of the diagnostic potential of the AF signal as a surrogate for full PSG in SAHS detection. Our methodology is based on three main steps. First, a feature extraction stage is implemented to obtain information of SAHS from single-channel AF. Physiological signals are known to behave both in deterministically and chaotically ways. For this reason, different methodologies were used to extract SAHS-related information, such as spectral and non-linear analyses. The purpose of this approaches was the optimum characterization of SAHS by means of obtaining complementary information. The second step is an automatic feature selection stage. The comprehensive analysis conducted in the previous stage may lead to the extraction of useless features for SAHS diagnosis or features sharing similar information than others. Thus, it has been implemented a feature selection stage to eliminate those non-relevant or redundant. Two different approaches have been used for this purpose: the well-known forward-selection backward-elimination method (SLR-FSBE) and the fast correlation-based filter (FCBF) algorithm. The former is a wrapper method since it is closely related to a specific classifer (logistic regression), whereas the latter is a filter since it is independent from subsequent analyses. Finally, the third stage is pattern recognition. In this Doctoral Thesis, it has been used to obtain an automatic SAHS diagnosis by the application of different classification and regression methods to data obtained and selected in previous stages. The main purposes of this step have been determining the presence or absence of SAHS (binary classification), classifying subjects into one out of the four SAHS severity degrees (multiclass classification), and the estimation of AHI (regression). This approach differs from the common approach followed in the state of the art, where the main studies focus on detecting each of the apneic events present in the recordings. After applying our methodology to single-channel AF, the results showed that our proposal outperformed a classic event-detection algorithm applied to our databases. Thus, in the case of binary classification, an ensemble learning model based on AdaBoost, built with decision trees, reached 89.0% sensitivity (Se), 80.0% specificity (Sp), 86.5% accuracy (Acc), 0.950 area under the receiver-operating characteristics curve (AROC), and 0.672 Cohen's k, in contrast to the classic event-detection algorithm which obtained 75.8% Se, 54.3% Sp, 64.0% Acc, 0.635 AROC, and 0.286 Cohen's k. Regarding multiclass classi fication, another AdaBoost model, built with linear discriminant classifiers, obtained 86.5%, 81.0%, and 82.5% accuracies when evaluated in the AHI cutoffs which establish each of the four SAHS severity degrees (AHI = 5 events/hour, 15 e/h, and 30 e/h). The event-detection algorithm obtained lower statistics for each threshold, reaching 81.0%, 68.3% y 63.5%, respectively. Finally, when applying regression to estimate AHI, an artificial neural network model based on multi-layer perceptron (MLP) obtained an intra-class correlation coeficient (ICC) of 0.849, and 79.7%, 91.5%, 79.7%, and 88.1% diagnostic accuracies for AHI cutoffs = 5 e/h, 10 e/h, 15 e/h y 30 e/h, respectively, each of them associated with corresponding 0.903, 0.956, 0.904, and 0.973 AROC values. By contrast, the event detection algorithm reached 0.840 ICC, and accuracies of 79.7% (0.823 AROC), 78.0% (0.833 AROC), 66.1% (0.867 AROC) y 74.6% (0.982 AROC). On the other hand, our methodology applied to at-home AF recordings from children showed higher performance than the oxygen desaturation index (ODI), which is commonly used in clinical practice. Additionally, the combination of spectral information from these recordings with ODI achieved 85.9% Se, 87.4% Sp, 86.3% Acc, 0.947 AROC and 0.720 k. Our proposal achieved high diagnostic performance comparing with PSG diagnosis, state-of-the-art studies focused on detecting apneic events in other AF databases, as well as studies reporting similar approaches to ours applied in different PSG signals. Consequently, the main conclusion obtained from this Doctoral Thesis is that pattern recognition methods applied to single-channel AF are useful to improve the automation of the SAHS diagnosis process. Hence, it is also concluded that this process can be reliably simplified by means of the automatic analysis of AF.