RT info:eu-repo/semantics/doctoralThesis
T1 Utilidad de las señales de oximetría y flujo aéreo en el diagnóstico simplificado de la apnea obstructiva del sueño. Diseño de un test automático domiciliario
A1 Cerezo Hernández, Ana
A2 Universidad de Valladolid. Escuela de Doctorado
K1 Sueño, trastornos del - Tesis doctorales
K1 Apnoea
K1 Apnea
K1 Sleep
K1 Sueño
K1 Diagnosis
K1 Diagnóstico
K1 32 Ciencias Médicas
AB Obstructive Sleep Apnea (OSA) is a respiratory disorder characterized by recurrent episodes of total (apnea) or partial (hypopnea) absence of airflow during sleep. Untreated OSA produces a significant decrease in quality of life and is associated with the main causes of mortality in industrialized countries.However, OSA is considered an underdiagnosed chronic disease. Continuous positive airway pressure (CPAP) is the most common therapeutic option. Nocturnal polysomnography (PSG) in a specialized sleep unit is the reference diagnostic method, although it has low availability and accessibility. Consequently, in recent years there has been a significant demand for abbreviated methods, most of them at home, to reduce waiting lists. The fundamental hypothesis that the use of automatic processing techniques based on machine learning tools could allow maximizing the diagnostic accuracy of a reduced set of combined biomedical signals: overnight oximetry and airflow recorded at patient&#8217;s home. The main objective was to evaluate whether the joint analysis by means of machine learning algorithms of unsupervised SpO2 and AF signals acquired at patient's home leads to a significant increase in diagnostic performance compared to single-channel approaches. A prospective observational study was carried out in which a population referred consecutively to the Sleep Unit showing moderate-to-high clinical suspicion of having OSA was analyzed.All patients underwent an unsupervised PSG at home(gold standard) from which the SpO2 and AF signals were extracted, which were subsequently processed offline.The apnea-hypopnea index(AHI) derived from the PSG was used to confirm or rule out the presence of the disease.Three different approaches for screening patients with suspected OSA were assessed in terms of the source of information used: single-channel based on SpO2, single-channel based on AF, and two-channel combining information from both SpO2 and AF.The automatic processing of the SpO2 and AF signals was developed in 4 stages: preprocessing, feature extraction, feature selection, and pattern recognition. Unsupervised SpO2 and AF recordings were parameterized using the fast correlation-based filter(FCBF)algorithm.The following machine learning methods were used: linear regression(MLR), multilayer perceptron neural networks(MLP) and support vector machines(SVM). The population was divided into independent training and test groups. Agreement between the estimated and the actual AHIderived from at-home PSG was assessed, and typical OSA cutoff points(5, 15, and 30 events/h) were applied. A total of 299 unattended PSGs were performed at home, with a validity percentage of 85.6%. The highest agreement between the estimated AHI and the PSG AHI was reached by the SVMSpO2+AF model, with an CCI 0.93 and a 4-class kappa index 0.71, as well as with an overall accuracy for the 4 OSA severity categories equal to 81.25%, significantly higher than the individual analysis of the SpO2 signal and the airflow signal.The SVMSpO2+AF model achieved the highest diagnostic performance of all algorithms for the detection of severe OSA, with an accuracy of 95.83% and AUC ROC 0.98. In addition, the AUC ROC of the dual-channel models was significantly higher (p<0.01) than that achieved by all the single-channel approaches for the cutoff of 15events/h. The proposed methodology based on the joint automatic analysis of the SpO2 and AF signals acquired at home showed a high complementarity that led to a remarkable increase in diagnostic performance compared to single-channel approaches. The automatic models outperformed the conventional indices(desaturation and airflow-derived indexes) both in terms of correlation and concordance with the AHI from PSG, as well as in terms of overall diagnostic accuracy, providing a moderate increase in diagnostic performance, particularly in the detection of moderate-to-severe OSA.Our findings suggest that the joint analysis of oximetry and airflow signals by means of machine learning methods allows a simplified as well as accurate screening of OSA at patient's home.
YR 2021
FD 2021
LK https://uvadoc.uva.es/handle/10324/60998
UL https://uvadoc.uva.es/handle/10324/60998
LA spa
NO Escuela de Doctorado
DS UVaDOC
RD 29-jun-2024