Objective: To evaluate the performance of 4 published prognostic models for postoperative onset of nausea and vomiting (PONV) by means of discrimination and calibration and the possible impact of customization on these models.
Design: Prospective, observational study.
Setting: Tertiary care university hospital.
Patients: 748 adult patients (>18 years old) enrolled in this study. Severe obesity (weight > 150 kg or body mass index > 40 kg/m) was an exclusion criterion.
Interventions: All perioperative data were recorded with an anesthesia information management system. A standardized patient interview was performed on the postoperative morning and afternoon.
Measurements: Individual PONV risk was calculated using 4 original regression equations by Koivuranta et al, Apfel et al, Sinclair et al, and Junger et al Discrimination was assessed using receiver operating characteristic (ROC) curves. Calibration was tested using Hosmer-Lemeshow goodness-of-fit statistics. New predictive equations for the 4 models were derived by means of logistic regression (customization). The prognostic performance of the customized models was validated using the "leaving-one-out" technique.
Main results: Postoperative onset of nausea and vomiting was observed in 11.2% of the specialized patient population. Discrimination could be demonstrated as shown by areas under the receiver operating characteristic curve of 0.62 for the Koivuranta et al model, 0.63 for the Apfel et al model, 0.70 for the Sinclair et al model, and 0.70 for the Junger et al model. Calibration was poor for all 4 original models, indicated by a P value lower than 0.01 in the C and H statistics. Customization improved the accuracy of the prediction for all 4 models. However, the simplified risk scores of the Koivuranta et al model and the Apfel et al model did not show the same efficiency as those of the Sinclair et al model and the Junger et al model. This is possibly a result of having relatively few patients at high risk for PONV in combination with an information loss caused by too few dichotomous variables in the simplified scores.
Conclusions: The original models were not well validated in our study. An antiemetic therapy based on the results of these scores seems therefore unsatisfactory. Customization improved the accuracy of the prediction in our specialized patient population, more so for the Sinclair et al model and the Junger et al model than for the Koivuranta et al model and the Apfel et al model.