Resumen de Analysis of nocturnal oximetry recordings using pattern recognition techniques to assist in the diagnosis of the sleep apnoea-hypopnoea syndrome
Sleep apnoea-hypopnoea syndrome (SAHS) is the most common form of sleep disordered-breathing. SAHS is characterised by repetitive occlusion of the upper airway during sleep, causing intermittent cessations of breathing (apnoeas) or reduction in airflow (hypopnoeas). It is associated to hypoxaemia, bradycardia and fragmented sleep. SAHS has been pointed out as a major cause of traffic and industrial accidents due to excessive daytime sleepiness. Long-term effects can lead to severe cardiovascular and cerebrovascular diseases. Its prevalence is estimated at 5% of the adult population in western countries. Moreover, it is suspected that a high percentage of patients suffering from SAHS remain undiagnosed. Therefore, SAHS can be considered as a risk factor for public health.
Nowadays, nocturnal polysomnography (PSG) is the gold-standard for SAHS diagnosis. PSG must be performed in a special sleep unit and under supervision of a trained technician. Different physiological recordings and data are monitored during a complete night. They must be manually analysed by an expert to obtain a definitive diagnosis. It is based on the value of the apnoea-hypopnoea index (AHI), which measures the ratio of apnoeas and hypopnoeas per hour of sleep. Despite its high diagnostic performance, PSG presents some drawbacks since it is complex, expensive and time-consuming. Additionally, the demand for PSG studies is progressively growing as people and clinicians are becoming aware of SAHS while the available infrastructure is insufficient to support it. Thus, simplified diagnostic techniques are desirable.
This Thesis proposes to analyse oxygen saturation (SaO2) signals recorded through nocturnal pulse oximetry to assist in SAHS diagnosis. Pulse oximetry is a non-invasive technique to monitor arterial blood oxygenation. It can be performed at patient's home, resulting in reduced complexity and cost in comparison with PSG. SaO2 recordings reflect hypoxaemia due to airflow reduction during apnoea/hypopnoea events. As a result, SaO2 signals from SAHS patients tend to be more unstable than those from control subjects due to the recurrence of apnoeas during sleep. This different behaviour can be exploited to detect SAHS. In addition to visual inspection, several conventional oximetry indices have been suggested for automated interpretation of SaO2 signals. The oxygen desaturation index over 3% (ODI3) and 4% (ODI4) as well as the cumulative time spent below 90% of saturation (CT90) are the most popular oximetry indices. In addition, other measurements such as the minimum SaO2 value (minSaO2), the saturation impairment time at 90% (SIT90) or the delta index have been proposed. However, none of them has been accepted as a reliable predictor of SAHS.
A novel method for automated SaO2 analysis is proposed in the Thesis. Pattern recognition techniques were used to model SAHS diagnosis from oximetry data. The applied methodology comprised three different stages: 1) feature extraction, 2) normalisation and dimensionality reduction, and 3) pattern analysis. In the first one, a total of 14 time-domain and frequency-domain features were extracted from SaO2 recordings in order to characterise their dynamical behaviour. In the second stage, the distribution of each feature was normalised to have zero mean and unit variance. The utility of dimensionality reduction by means of principal component analysis (PCA) after normalisation was evaluated. In the third stage, pattern recognition techniques were used to process the features (or components) from the previous phase. Two different approaches were proposed to model SAHS diagnosis: classification and regression. The output of classification methods represents a categorical variable indicating one of two possible groups for the input pattern: SAHS-negative and SAHS-positive. The regression approach aims to provide accurate estimations of the AHI from information contained in the input pattern.
A database composed of 240 SaO2 signals was available for this study. It was randomly divided into a training set (40%) with 96 SaO2 signals (32 SAHS-negative and 64 SAHS-positive) and a test set (60%) with 144 SaO2 signals (48 SAHS-negative and 96 SAHS-positive). The former was used for model selection and optimisation while the latter was allocated for assessing the trained algorithms. Several pattern recognition methods were evaluated for both tasks. To analyse the effect of dimensionality reduction, each of them was assessed with and without PCA in the second stage. A total of 16 classification algorithms and 14 regression algorithms were built. The classifier with the highest performance was based on the analysis of the complete set of normalised features by means of a multilayer perceptron (MLP) network. A classification accuracy of 92.36% on the test set was reached (94.79% sensitivity and 87.50% specificity). This algorithm outperformed the delta index, which provided the best classification results (88.19% accuracy) among the conventional oximetry indices. The highest accuracy in the regression task was achieved by two algorithms: one based on multivariate adaptive regression splines (MARS) for processing the complete set of normalised features and another based on Bayesian MLP networks using the components retained from PCA. Both reached an intraclass correlation coefficient (ICC) higher than 0.9. Moreover, the diagnostic accuracy provided by the estimated AHI was 87.50% (88.54% sensitivity and 85.42% specificity) for both of them. Despite ODI3 also achieved high ICC, it showed to be significantly inaccurate for predicting small AHI values. Most of the subjects misdiagnosed by the selected classification and regression algorithms had mild SAHS (5 h-1 ¿ AHI ¿ 15 h-1). Thus, these algorithms represent valuable screening tools that could contribute to reduce the number of required PSG tests.
