Prospective evaluation of logistic regression models from overnight oximetry to assist in sleep apnea diagnosis

This study focused on prospectively testing diagnostic performance of different logistic regression (LR) models in the context of sleep apnea hypopnea syndrome (SAHS) detection from blood oxygen saturation (SaO2) recordings. Feature extraction, selection and classification procedures were applied. Time, frequency, linear and nonlinear analyses were carried out to compose the initial feature set. Forward stepwise logistic regression (FSLR) was applied for feature selection. LR was used to measure performance classification of single features and an optimum feature subset from FSLR. A training set composed of 148 recordings from patients suspected of suffering from SAHS was used to obtain LR models, which were further validated on a dataset composed of 50 recordings from normal healthy subjects and 21 recordings from SAHS patients, all derived from an independent sleep unit. Diagnostic performance of one-feature LR models from oximetry in the training set significantly changed on further assessments in the test set. On the other hand, FSLR provided a more general LR model in the context of SAHS, which reached an accuracy of 89.7% on the training set and 87.3% on the test set.