Validity of Neural Network in Sleep Apnea

Clinical assessment of obstructive sleep apnea (OSA) is poor. Overnight polysomnography (OPG) is the standard reference test, but it is expensive and time-consuming. We developed an artificial neural network (ANN) using anthropomorphic measurements and clinical information to predict the apnea-hypopnea index (AHI). All patients completed a questionnaire about sleep symptoms, sleep behavior, and demographic information prior to undergoing OPG. Neck circumference, height, and weight were obtained on presentation to the sleep center. Twelve variables were used as inputs. The output was an estimate of the AHI. The network was trained with a back-propagation algorithm on 189 patients and validated prospectively on 80 additional patients. Data from the derivation group was used to calculate the 95% confidence interval of the estimated AHI. Predictive accuracy at different AHI thresholds was assessed by the c-index, which is equivalent to the area under the receiver operator characteristic curve. The c-index for predicting OSA in the validation set was 0.96 +/- 0.0191 SE, 0.951 +/- 0.0203 SE, and 0.935 +/- 0.0274 SE, using thresholds of > 10, > 15, and > 20/hour respectively. The actual AHI of the 80 patients in the validation data set fell within the 95% confidence limits of the values predicted by the ANN. This study suggests that ANN may be useful as a predictive tool for OSA.