What can be expected from risk scores for predicting postoperative nausea and vomiting?

Several risk scores have been developed to calculate the Probability of postoperative nausea and vomiting (PONV). However, the power to discriminate which individual will suffer from PONV is still limited. Thus, we wondered how the number of predictors in a score affects the discriminating power and how the characteristics of a population - which is needed to measure the power of a score - may affect the results. For ethical reasons and to be independent from centre specific populations, we developed a computer model to simulate virtual populations. Four populations were created according to number, frequency, and odds ratio of predictors. Population I: parameters were derived from a previously published paper to verify whether calculated and reported values are in accordance. Population II: a gynaecological population was created to investigate the impact of the study setting. Populations III and IV: to meet ideal assumptions a model with up to seven predictors with an odds ratio of 2 and 3 was tested, respectively. The discriminating power of a risk score was measured by the area under a receiver operating characteristic curve (AUC) and an increase of more than 0.025 per predictor was considered to be clinically relevant. The AUC of population I was similar to those reported in clinical investigations (0.72). The study setting had a considerable impact on the discriminating power since the AUC decreased to 0.65 in a gynaecological setting. The AUC with the 'idealized' populations III and IV was at best in the range of 0.7-0.8. The inclusion of more than five predictors did not lead to a clinically relevant improvement. The currently available simplified risk scores (with four or five predictors) are useful both as a method to estimate individual risk of PONV and as a method for comparing groups of patients for antiemetic trials. They are also superior to single predictor models which are just using the patients' history of PONV or female gender alone. However, our analysis suggests that the power to discriminate which individual will suffer from PONV will remain imperfect, even when more predictors are considered.