Sleep detection in heart rate signals from photoplethysmography

The obstructive sleep apnea/hypopnea syndrome (OSAHS) is one of the most prevalent sleep pathologies. The OSAHS is characterized by repetitive interruptions of respiratory flow caused by upper airway obstructions. The apnea/hypopnea index (AHI) allows to quantify OSAHS severity. It is important to develop methods that allow automatic screening of OSAHS, improving accessibility and costs associated with polysomnography studies. In this paper, machine learning techniques are applied to determine whether a patient is awake or asleep from heart rate signals provided by pulse oximetry. The estimation of total sleep time will allow to improve the existing algorithms that estimate IAH and implement new algorithms using only pulse oximetry. In the feature extraction stage, information theory, continuous wavelet transform and classical statistics measures were used. As an objetive measure to evaluate performance, optimal parameters were selected using the area under the ROC curve. Finally, a support vector machine (SVM) classifier was trained with the optimal features and the accuracy (0.79), sensitivity (0.70) and specificity (0.82) were calculated. Sleep times for each patient were evaluated, obtaining a low average relative error (0.15). In addition, a Bland-Altman analysis was performed to compare the estimated and real sleep times.