During the last decades, life-threatening complications associated with anesthesia have become very rare. This safety record has encouraged anesthesiologists to focus attention on minor morbidity. Of these, postoperative nausea and vomiting (PONV) is one of the “big little problems” after general anesthesia (1). PONV may decrease parental satisfaction, increase the use of resources, including medical and nursing care, IV fluids, drugs, and other supplies (2–5). Furthermore, in the ambulatory setting, PONV is a major cause of unanticipated admission (6).
The incidence of this distressing problem can be reduced by using a total IV anesthetic (TIVA) technique instead of inhaled anesthetics and by administering antiemetics prophylactically. However, routine efforts to prevent PONV are not indicated because of the potential for adverse effects, the perception that there are increased costs, and the lack of evidence that patient satisfaction is affected (7).
For these reasons, tools to predict an increased risk for developing nausea and vomiting are certainly useful in clinical practice. Several scores have been developed for adults. However, their use in pediatric patients is limited, because several of the proven risk scores for adults are difficult to assess or not applicable to children. For example, nonsmoking and a history of PONV after previous anesthesia are known risk factors for PONV in adults. Obviously, there are only a few children younger than 14 years old who smoke and fewer children presenting for elective surgery have had a previous anesthesia compared with an adult population and are thus more often erroneously classified as having a “negative” history of postoperative vomiting (PV).
Thus, the aim of this analysis was also to create a simplified model1 to facilitate the prediction of POstoperative VOmiting in Children (POVOC-score).
The data of 1401 children (0–14 yr old) included in this prospective survey were collected during a period of 22 mo at 2 university hospitals, a community children’s hospital, and an outpatient surgical center. The local ethics committee approved the study and the parents gave informed written consent.
Preoperatively, the following data were obtained from the parents: history of PV or motion sickness in the child, history of PONV in the mother, the father, or in any siblings, preoperative anxiety of the child. The children were fasted 2–4 h preoperatively from clear fluids and at least 6 h from milk and solids. All received oral premedication with midazolam. If no contraindications were present, the children received a non-opioid analgesic (acetaminophen, metamizole, and/or diclofenac) intraoperatively or immediately after operation. The first oral intake was allowed depending on the length and site of operation. In most cases (90%), this was within the first 4 h postoperatively. Because of the observational character of the study, the anesthesia technique was not standardized but was performed according to the local standards (Table 1).
Using the anesthesia recordings, the following data were extracted: age, sex, weight, height, type of induction (volatile versus IV), duration of anesthesia and surgery, type of surgery, type and dosage of anesthetics drugs, and additional measures (regional blocks, nasogastric tubes, etc.).
Postoperatively, vomiting or retching (PV) was assessed in the postanesthetic care unit by specially instructed nurses or anesthesiologists. PV was chosen as the main end-point of the survey, because nausea is a subjective phenomenon, and the smaller child often may not be able to describe it (9). Twenty-four hours postoperatively, the children and/or their parents were interviewed. Additionally, all medical records were screened and the nursing staff was asked in order not to miss an emetic episode. The parents of patients having surgery on an outpatient basis were interviewed by telephone using a structured interview on the first postoperative day.
Preprocessing of the Data
To reduce the amount of variables for the final analyses, several clinical data (e.g., administration of any muscle relaxants, opioids or non-opioid analgesics separated according to intra- and postoperative administration, or the use of any local or regional anesthesia performed intraoperatively) were simplified and collapsed into separate dichotomous variables. Weight and height were used to calculate the body mass index that was used for further analysis. Type of surgery was classified mainly according to the frequency, the anatomical location, and the complexity of the surgery (Table 2). The main outcome data (incidence of PV during the first 24 h) were dichotomized into nonvomiters and vomiters.
Of the1401 children that were observed, 88 receiving a prophylactic antiemetic, including corticosteroids, were withdrawn from the analysis. A further 56 patients were lost for follow-up or were excluded from analysis because of incomplete recordings. The remaining 1257 patients were randomly split into an evaluation dataset (n = 657) and a validation set (n = 600).
Creation of the Statistical Models
The evaluation set was subjected to stepwise forward logistic regression analysis using the maximum likelihood function. To allow the model to compensate for nonlinearity with regard to the influence of continuous variables (age, duration of anesthesia, duration of surgery), these data were also dichotomized (14 groups with intervals of 1 yr for age 0–1, 0–2, 0–3 yr etc., and each 8 groups for duration of surgery and anesthesia with 15-min intervals, respectively). Both, the continuous variables and the dichotomized values were offered to the logistic regression analysis. The goodness of fit of a model was judged using Nagelkerke’s R2. All analyses were performed using SPSS 11.0 for Windows. The factors included in the initial model were used to calculate the probability of vomiting for each child of the validation dataset. The discriminating properties of a predictive model were investigated by calculating the area under a receiver operating characteristics (ROC) curve. This graph can be constructed by correlating true- and false-positive rates (sensitivity and 1 minus specifity, respectively) for a series of cut-off points for a test in which the cut-off point is the predicted risk. The area under the curve (AUC) represents the probability that a random pair of test results will be ranked correctly as to their disease state (10). Theoretically, a 45° bisector would be a score predicting not better than a random guess. Thus, the area under this “random score” would be 0.5. A score performing significantly better than chance has an AUC >0.5 with the lower limit of the 95% confidence interval exceeding the value of 0.5 (8).
Adjustments of the Initial Risk Model
Because our aim was to create a simplified score that can be easily calculated by counting the number of unweighed risk factors present in an individual patient, several modifications and adjustments were made in the initial (basic) model. These are explained step-by-step in detail in the Results section. Each adjustment of the risk model was accompanied by a subsequent calculation of Nagelkerke’s R2 as a measure of the goodness of fit of the regression model and by calculating the area under the ROC curve for the validation set.
Initial (Basic) Risk Model
When all available data (including the dichotomized continuous variables of “age,” “duration of surgery,” and “duration of anesthesia” were entered into the first logistic regression analysis, seven risk factors were included in the model (Table 3). Nagelkerke’s R2 (=0.271) showed acceptable goodness of fit. The AUC of the ROC curve was 0.76 (95% confidence interval: 0.72–0.80) when the score was applied to the evaluation dataset. However, from the practical point of view, this first model was found to be too complex and was not compatible with the predefined aim of the study to create a simplified model.
First Adjustment of the Risk Model
In this step, the information about the previous experiences of PV by the patient and of PV/PONV by the relatives [both were significant risk factors with an odds ratio (OR) of 2.6 and 1.4, respectively] were grouped by using a Boolean “OR” connection. Furthermore, age was removed because it was the only continuous variable in the basic risk model. Multiplication of the actual age of a patient with the coefficient complicates a risk model and counteracts the intention to create a model that is easy to use. While running a subsequent logistic regression analysis, “age” (OR in the basis model: 1.057 per year) was substituted by the variable “8 yr or older” (OR of the dichotomous variable: 1.5).
Combining the two variables on anamnestic information whether PV/PONV had been present in the child or its relatives had an OR of 3.7 that is approximately the product of the two previous ORs (Table 3). To further reduce the number of variables, it was decided to replace the variable “duration of anesthesia longer than 90 min” by the duration of surgery, because duration of anesthesia and surgery are obviously highly correlated (quantified by the equation: duration of surgery = 0.941 * duration of anesthesia − 14.6 min; R2 = 0.982; P < 0.0001). Thus, the variable “anesthesia longer than 90 min” was replaced by “duration of surgery longer than 75 min.” A subsequent reanalysis of the data revealed that the goodness of fit of the regression model was only slightly affected (Nagelkerke’s R2 = 0.252) as were the predictive properties when the model was applied to the validation dataset (AUC value of 0.73 (0.68–0.75).
Second Adjustment of the Risk Model
After the first adjustment of the model, there were two factors that were present in duplicate. Increasing age of the child undergoing surgery shows a two-stage increase of the risk to vomit postoperatively. In the present model, these two levels were age of 3 (OR: 2.707) and 8 yr or older (OR: 1.463), respectively. The same applies for the duration of surgery, where 30 min (OR: 2.535) and 75 min of duration (OR: 1.895) were determined as a critical period in this dataset. It was decided to manually remove the less predictive variable of these two factors because of three major reasons. First, from a practical point of view, it seems reasonable to further reduce the amount of risk factors that need to be remembered by the anesthesiologist when applying such a score. Second, “age of 8 yr and older” and “duration of surgery longer than 75 min” both were the weakest predictors for PV in the model. Third, removing one variable of a two-step risk factor strengthens the remaining variable in the model. For example, children 8 yr or older will also increase the odds of PV in children in the group of 3 yr and older. The same applies for the duration of surgery. The results of this step are given in Table 4.
Third (Final) Adjustment of the Risk Model
To meet the predefined aim to develop a simplified model, where the risk of PV can be calculated by simply counting the number of risk factors, the coefficients were finally removed from each risk factor. Because the risk factors all had ORs approximately between 3 and 4, no relevant alteration in predicting properties was expected. Using the evaluation dataset, the observed incidences for the presence of 1, 2, 3, and 4 risk factors were determined. These were 9%, 10%, 30%, 55%, and 70% for 0, 1, 2, 3, and 4 risk factors observed. Using these incidences as cut-off values in the validation dataset, the area under this ROC curve was 0.72 (95% confidence interval: 0.68–0.77). All ROC curves for this and for the previous initial and intermediate risk models are plotted in Figure 1.
A recent evaluation of risk scores that was created for adult patients has shown that none of the investigated risk models was suitable for children (11). As pointed out, this is because several risk factors shown to be relevant in adults are not applicable in most children (e.g., smoking status).
Comparing risk factors for PV/PONV in adults and those identified as independent and relevant in the present study can give interesting insights in the etiology of PONV.
Younger age is a risk factor that was identified in adults (12). However, there is good evidence from clinical trials that toddlers are less susceptible to emetic stimuli than school children and adolescents (13). Our findings support the latter observation. Around the age of three years, the risk to develop PV increases dramatically. This is represented by an OR of 2.4 in the initial model. Furthermore, there was a continuous increase of PV in the children with increasing age. The OR per year was 1.057 that is (depending on the presence of other risk factors) an increase of 0.2%–0.8% per year. This aspect of the influence of age on the occurrence of PV was not included in the final model in order to keep it simple and easy to use in clinical practice. Manually removing the continuous variable “age” subsequently strengthened the factor “age 3 years and older” that had an OR of 3.3 in the final model and thus reaches the impact of other important predictors of PV in children.
There is a continuing debate about whether the type or the site of surgery influences the occurrence of PONV in adults. There are risk scores in which several types of surgery were identified as risk factors [e.g., Sinclair et al. (14)]. However, there is also good evidence that in a large majority of patients the observed frequent (e.g., females undergoing gynecological laparoscopies) or infrequent incidence (e.g., elderly men undergoing transurethral resections of the prostate) of PONV in adults can be explained with biometric and anamnestic risk factors. Thus, most of the established and validated risk scores do not include any type of surgery in their risk model. In children, we came to a very similar conclusion. Most surgery does not have an influence on PV, even though this might have been expected by theoretical pathophysiological considerations (e.g., middle ear surgery is often considered to be a risk surgery). However, our analysis revealed that strabismus surgery was an independent risk factor for PV. It was the most significant predictor with an OR of 5.2 in the basic and 4.3 in the final adjusted model. This result is in agreement with findings in a systematic review on antiemetic prophylaxis in children undergoing strabismus repair where an extraordinary frequent incidence (up to 87%) could be observed in several trials (15).
Duration of Surgery
The duration of surgery and anesthesia respectively have an impact on PV symptoms (12,14,16). Although the exact pathophysiological background is still unknown, there is a rational biological basis. The longer an emetic stimulus (e.g., administration of volatile anesthetics and opioids) is present, the more likely it is that this trigger leads to nausea and vomiting.
The positive history of PONV is an unequivocally accepted risk factor for further PONV symptoms at future anesthesia (14,16–18). Thus, it was not surprising to notice that this was also the case in children. More interesting is that children with parents or siblings who have experienced PV or PONV after a previous anesthesia are at increased risk. The question is whether this family association with PV/PONV is genetically or behaviorally determined. There is some evidence in the literature that genetic aspects might be involved. For example, monozygotic twins have more frequent congruent behavior with regard to developing PONV than heterozygote twins.2
Factors Not in the Initial Model
There were several potential risk factors for PV in our patients that were not statistically significant in the present trial and thus were not included in the basic risk model. However, reviewing them might give interesting insights into the etiology of PV in children. For this purpose, an explorative backward logistic regression analysis was performed. The factors that were removed during the last steps of this analysis and thus are most likely to affect PV in children are listed in Table 5.
Administration of Local or Regional Anesthesia
Administration of local or regional anesthesia is an important measure to reduce the intensity of postoperative pain and to improve postoperative recovery and well-being after surgery. Moreover, these techniques also seem to positively influence the occurrence of PV. Although not significant in the statistical analysis (P = 0.065), they should be taken into account as a part of a multimodal approach to reduce the incidence of PV.
In our analysis, all types of locoregional techniques were grouped into one dichotomous variable to obtain an adequate statistical power for the variable. This means that wound infiltration of the surgical site as well as caudal blocks and peripheral nerve blocks, which were performed in some complex orthopedic surgery along with general anesthesia, were represented within one variable.
The mechanism for the potential antiemetic effect of performing locoregional anesthesia in children remains speculative. When performed intraoperatively, a locoregional block reduced the need for opioids and also for large doses of volatile anesthetics that were shown to be a main cause for PV during the early stage of recovery (19). During the postoperative period, these blocks can reduce the need for postoperative opioids.
Intraoperative Administration of Opioids
The use of intraoperative opioids showed a tendency to reduce the incidence of PV, which seems at first surprising. Most procedures that were performed without the use of opioids were short-lasting procedures, e.g., adenectomies or myringotomies. However, the multivariate analysis revealed that omission of opioids is not the reason for the infrequent incidence of PV in these types of surgery but it is the surgery’s short duration. One possible explanation for the (nonsignificant) PV-reducing effects of using intraoperative opioids might be their dose-sparing effect on volatile anesthetics (19).
Postoperative Administration of Opioids
It is not surprising that the administration of postoperative opioids had a tendency to increase PV, because opioids are known to cause PONV. In two of four risk scores for PONV in adults, the administration of opioids was an independent and statistically significant predictor for the occurrence of these symptoms (16,20).
Previous studies have shown that even in older children, gender does not have a major role in the occurrence of PV (21). In our trial, gender was removed but was the eleventh strongest predictor in the backward regression analysis. Similar to adults, female patients are (not significantly) more prone to PV. It can be speculated that this result is because some girls were included in the analysis that were already at the age to menstruate. This view is supported by the fact that an interaction term between age (11 years and older) and female gender was a better predictor than female gender alone.
Limitations of the Risk Model
One major limitation of our study and our results is that the underlying population does not represent all potential heterogeneity of clinical practice seen in anesthesia as well as in all surgical specialties. As a consequence, techniques that were under-represented in our dataset had no chance to be a significant risk factor or protective factor. As an example, there were only 54 children who had received TIVA. Thus, TIVA had no chance to show its antiemetic properties in this analysis, although it is known from several previous clinical studies and quantitative systematic reviews that propofol used for induction and maintenance of anesthesia reduces the incidence of PV/PONV (22,23). Another example is the types of surgery that were not surveyed in our trial (e.g., neurosurgical, cardiac surgical). Thus, on one hand, the results of our analysis only apply for anesthesia techniques, types of surgeries, and perioperative management of the patients who were described in the Methods section and in Tables 1 and 2. On the other hand, our data were collected in four different centers working independently and with different local standards. Thus, results from our analysis can be better transferred to other institutions than a score derived from a single center.
Another potential criticism of our study might be the preprocessing of the data before entering the logistic regression analysis. While combining and dichotomizing the data, several simplifications were made that could have affected the final results. For example, all IV induction drugs used during the survey were merged into one dichotomous variable and grouped against all other children in whom anesthesia was induced by inhaled halothane or sevoflurane. On one hand, this procedure eliminates potentially valuable information but, on the other hand, it is the only possibility to create variables with adequate statistical power to “survive” in the logistic regression model. Another example for this dilemma—grouping all the heterogeneous techniques of local, regional, and central regional blockades in one variable—has been discussed previously.
Another problem in this context is that even variables that seem to be perfectly defined are not homogenous. For example, “strabismus surgery” summarizes different surgical approaches (with or without myopexy) that might all have different emetogenic potential (24).
Other problematic aspects of the statistical modeling are potential interactions between the individual factors. Logistic regression is not the optimal statistical tool for complex interaction analysis. Furthermore, it has been recommended that such interactions should only be examined when there is a biological rationale for a potential interaction. As an example, an interaction term “age older than 11 years” and “female gender” provides more information than each of the single variables alone.
Comparison of the Predictive Properties of the Initial Basic Risk Model and the Finally Adjusted Simplified Risk Score
When comparing the initial risk model with its 7 items and the final simplified risk score with 4 variables, there was a statistically significant decrease in the predictive properties of the model (measured by the area under the ROC curve) from 0.76 to 0.72. However, from the clinical point of view, this decrease is of minor importance (16).
As in previous risk models to predict the incidence of nausea and vomiting in adults (8,25), it is of great importance that risk models can be used easily in clinical practice to guarantee widespread use and acceptance by clinicians. In this context, two criteria are important: First, there should be as few factors as possible that must be remembered when the score is applied to a patient. Second, the recorded risk factors must be easily transferable to the patient’s individual risk. This is the case only if no complicated calculations are needed. The score that used this approach was presented by Koivuranta et al. (16), who reported interesting details about their approach to simplify the predictive model. For example, they did not observe a relevant reduction in the discriminating power of their model when the number of factors was reduced from 8 (prediction of PONV) or 10 (prediction of PV) to the final 5 factors. Furthermore, eliminating the coefficients for each factor and simply summing up their numbers resulted in nearly identical AUC values for the model.
We conclude that the occurrence of PV in children can be predicted using a duration of surgery longer than 30 minutes, age ≥3 years, strabismus surgery, and a positive history of PV in the children or PV/PONV in relatives (mother, father, or siblings) as major predictors.
1. Kapur PA. Editorial: the big “little problem,,”. Anesth Analg 1991;73:243–5.
2. Cieslak GD, Watcha MF, Phillips MB, Pennant JH. The dose-response relation and cost-effectiveness of granisetron for the prophylaxis of pediatric postoperative emesis. Anesthesiology 1996;85:1076–85.
3. Patel RI, Davis PJ, Orr RJ, et al. Single-dose ondansetron prevents postoperative vomiting in pediatric outpatients. Anesth Analg 1997;85:538–45.
4. Splinter WM, Rhine EJ, Roberts DJ. Vomiting after strabismus surgery in children: ondansetron vs propofol. Can J Anaesth 1997;44:825–9.
5. Watcha MF, Bras PJ, Cieslak GD, Pennant JH. The dose-response relationship of ondansetron in preventing postoperative emesis in pediatric patients undergoing ambulatory surgery. Anesthesiology 1995;82:47–52.
6. Patel RI, Hannallah RS. Anesthetic complications following pediatric ambulatory surgery: a 3-year study. Anesthesiology 1988;69:1009–12.
7. Scuderi PE, James RL, Harris L, Mims GR. Antiemetic prophylaxis does not improve outcomes after outpatient surgery when compared to symptomatic treatment. Anesthesiology 1999;90:360–71.
8. Eberhart LHJ, Högel J, Seeling W, et al. Evaluation of three risk scores to predict postoperative nausea and vomiting. Acta Anaesthesiol Scand 2000;44:480–8.
9. Rose JB, Watcha MF. Postoperative nausea and vomiting in paediatric patients. Br J Anaesth 1999;83:104–17.
10. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 1982;143:29–36.
11. Eberhart L, Morin A, Guber D, et al. Applicability of risk scores for postoperative nausea and vomiting in paediatric patients. Eur J Anaesthesiol 2004;21(suppl 32):A591.
12. Apfel CC, Greim CA, Haubitz I, et al. A risk score to predict the probability of postoperative vomiting in adults. Acta Anaesthesiol Scand 1998;42:495–501.
13. Cohen MM, Cameron CB, Duncan PG. Pediatric anesthesia morbidity and mortality in the perioperative period. Anesth Analg 1990;70:160–7.
14. Sinclair DR, Chung F, Mezei G. Can postoperative nausea and vomiting be predicted? Anesthesiology 1999;91:109–18.
15. Tramèr M, Moore A, McQuay H. Prevention of vomiting after pediatric strabismus surgery. Br J Anaesth 1996;76:473–7.
16. Koivuranta M, Läärä E, Snåre L, Alahuhta S. A survey of postoperative nausea and vomiting. Anaesthesia 1997;52:443–9.
17. Apfel CC, Läärä E, Koivuranta M, et al. A simplified risk score for predicting postoperative nausea and vomiting. Anesthesiology 1999;91:693–700.
18. Palazzo M, Evans R. Logistic regression analysis of fixed patient factors for postoperative sickness: a model for risk assessment. Br J Anaesth 1993;70:135–40.
19. Apfel CC, Kranke P, Katz MH, et al. Volatile anaesthetics may be the main cause for early but not delayed postoperative vomiting: a randomized controlled trial of factorial design. Br J Anaesth 2002;88:659–68.
20. Toner CC, Broomhead CJ, Littlejohn IH, et al. Prediction of postoperative nausea and vomiting using a logistic regression model. Br J Anaesth 1996;76:347–51.
21. Rowley MP, Brown TCK. Postoperative vomiting in children. Anaesth Intensive Care 1982;10:309–13.
22. Tramèr M, Moore A, McQuay H. Prevention of vomiting after pediatric strabismus surgery: a systematic review using the numbers-needed-to-treat method. Br J Anaesth 1995;75:556–61.
23. Tramèr MR, Moore RA, McQuay HJ. Propofol anaesthesia and postoperative nausea and vomiting: quantitative systematic review of randomized controlled studies. Br J Anaesth 1997;78:247–55.
24. Saiah M, Borgeat A, Ruetsch YA, et al. Myopexy (Faden) results in more postoperative vomiting after strabismus surgery in children. Acta Anaesthesiol Scand 2001;45:59–64.
25. Apfel CC, Kranke P, Eberhart LHJ, et al. Comparison of predictive models for postoperative nausea and vomiting. Br J Anaesth 2002;88:234–40.