Artificial Neural Network Modeling to Predict the Severity of Oxygen Desaturation and Improve the Nocturnal Hypoxemia after General Anesthesia in Patients with Morbid Obesity

Obese patients have more probability than the general population to develop nocturnal oxygen desaturation. Bariatric surgery is an efficient way to reduce obesity. About 250 obese patients received bariatric surgeries every year at Taipei Medical University Hospital. Because the surgery is performed under general anesthesia, the patient is prone to develop nocturnal oxygen desaturation postoperatively. We would explore the risk factors of nocturnal hypoxemia in obese patients, and to develop computer-based predictive models which can be used preoperatively to evaluate the severity of postoperative nocturnal hypoxemia for obese patients and help clinicians to alleviate the occurrence of postoperative nocturnal hypoxemia. In the first part of the study, total 150 obese patients who receive surgeries under general anesthesia will be enrolled. We will record age, gender, body mass index, waist, and other associated parameters of patients preoperatively. Nocturnal oximeters will be used to monitor and record the oxygen saturation of patients at night before operation, at night of the operation day, and in the first postoperative night. We will analyze the risk factors of nocturnal hypoxemia for these patients. Patients will be randomly assigned into three groups in which different positive end-expiratory pressure(PEEP) is set during general anesthesia. We will explore the improvement of nocturnal oxygen desaturation by different setting of PEEP. In the second part of the study, we will use the related risk factors to develop three different computer-based predictive models: artificial neural networks, support vector machine, and logistic regression models. After the models are constructed, the data of 60 new obese patients will be collected as the test set for external validation. The predictive performance of the models are validated and compared according to the discrimination and calibration of the models. Discrimination is evaluated by the area under the receiver operating characteristic (ROC) curve. Calibration is evaluated by the Hosmer-Lemeshow goodness of fit statistic. Finally we will apply the best predictive model in clinical practice for a period of six months. Obese patients who receive surgeries under general anesthesia will be enrolled and randomly assigned into two groups. In study group, the model will be used preoperatively to predict the severity of postoperative nocturnal oxygen desaturation and inform the clinician. In the control group, the clinician well not be informed of the risk of nocturnal oxygen desaturation. We will analyze if application of the predictive model in clinical practice will alleviate the severity of postoperative nocturnal hypoxemia.