LOGISTIC REGRESSION ANALYSIS OF FIXED PATIENT FACTORS FOR POSTOPERATIVE SICKNESS: A MODEL FOR RISK ASSESSMENT

Transcription

1 British Journal of Anaesthesia 1993; 70: LOGISTIC REGRESSION ANALYSIS OF FIXED PATIENT FACTORS FOR POSTOPERATIVE SICKNESS: A MODEL FOR RISK ASSESSMENT M. PALAZZO AND R. EVANS SUMMARY One hundred and forty-seven patients undergoing minor orthopaedic surgery were studied prospectively by logistic regression analysis to determine the association of independent fixed patient factors with the incidence of postoperative sickness (nausea, retching or vomiting). Gender, history of previous postoperative sicknness, postoperative opioids and interaction between gender and previous history of sickness were significant independent factors for postoperative sickness; history of motion sickness was weakly associated. The probability of postoperative sickness in the first h after surgery may be estimated from the equation: log it postoperative sickness = (postoperative opioids) (previous sickness history) + 2.4(gender) (motion sickness) 3.2(gender x previous sickness history). (Log likelihood ratio test for 5 degrees of freedom for the coefficients, chi-square = 53.5 (P < 0.001).) It is suggested that the calculated probability for sickness may be a useful addition for balancing patient treatment groups and allowing between-study comparisons. (Br. J. Anaesth. 1993; 70: ) KEY WORDS Vomiting: nausea, postoperative, preoperative factors. Risk. Postoperative sickness (nausea, retching and vomiting) occurs frequently after anaesthesia and surgery. Clinical impressions of a decrease in the incidence and severity of sickness are not supported by evidence published over the past 35 years [1-9] (table I). This may reflect, in part, variability in the method of investigation. There are many alleged causes of postoperative sickness, some patientrelated, others technique-related [10, 11]; these factors make adequately controlled clinical trials difficult to conduct. Although most studies control technique factors they usually fail to establish strictly matched patient groups. There may therefore be a need to establish some well-defined criteria by which patients can be considered comparable, particularly for the evaluation of antiemetics. The purpose of this preliminary study was to see if it is possible to provide a basis upon which to define comparable patients. PATIENTS AND METHODS Study design The factors that may influence postoperative sickness may be classified into those related to the patient and those related to anaesthetic technique or management. Patient or technique factors may be classified further as "fixed" or "variable" factors. Fixed factors are those which cannot be altered, such as gender or previous history of sickness. Variable factors are those which may be modified by the anaesthetist. Most patient factors are fixed. In contrast, technique factors may be fixed, or may be allowed to vary depending on experimental design. The primary aim of this study was to isolate and quantify the relative effect offixed patient factors on the incidence of postoperative sickness. An important aspect of the design of this study, therefore, was to ensure strict control of all possible technique factors while allowing the fixed patient factors to distribute freely without experimental bias among those who vomit and those who do not. Additionally, an attempt was made to reduce the confounding effect of controlled technique factors on the incidence of sickness by using the least emetic anaesthetic technique available. It was hoped that this would highlight die contribution of the fixed factors to postoperative sickness. The study design and analysis assumed that fixed patient factors did not have equal influences on the incidence of postoperative sickness and, furthermore, that they may have interacted to introduce additional independent variables. For this reason, statistical analysis of individual factor effects was assessed by logistic regression. Patients We studied 147 ASA I adult patients (80 female) undergoing minor peripheral orthopaedic surgery, MARK PALAZZO, M.B., CH.B., M.R.C.P., F.R.C.ANAES., M.D., Department of Anaesthetics, Charing Cross Hospital, Fulham Palace Road, London W6. RHYS EVANS, B.SC,, M.B., B.S., F.R.C.ANAES., D.PHIL., Nuffield Department of Anaesthetics, Oxford. Accepted for Publication: September 4, Correspondence to M.P.

2 136 BRITISH JOURNAL OF ANAESTHESIA TABLE I. Some previous studies examining postoperative sickness: nausea (N), retching (i?) or vomiting (V) included in the study. * Retrospective study. NA = Not available Reference Sample size Gender of subjects Sickness Observation period (h) [9] [2] [1] [31 [4] [5] [7] [6] [8] (1936)* (1955) (1957) (1958) (1960) (1961) (1970) (1979) (1982) F F F NA VR RV TABLE II. Some of the factors that may be associated with postoperative sickness. These can be classified into factors associated with the patient or those associated with technique. Most patient factors are fixed. Variable factors, either patient or technique may be fixed by study design Factors examined Age Weight Sex Duration of preoperative fast Previous postoperative emetic history History of motion sickness History of sickness with pregnancy Heartburn History of sickness with pain History of sickness with alcohol Anxiety before induction Nausea before induction Thirst before induction Hunger before induction Duration of anaesthesia Opioid administration Factors controlled Preoperative assessment Premedication Induction agent Maintenance agents Oropharyngeal suction Guedel airway Mask ventilation I.v. fluids Journey to theatre from ward Journey from theatre to recovery Recovery position Duration of postoperative observation Nurses observing patients Anaesthetist Time of first ambulation Time to first oral intake presenting sequentially over a 3-year period. Patients who were pregnant or lactating were not included. Hospital Ethics Committee approval and patient agreement to the study were obtained before the study. In order not to alert patients to our special interest in postoperative sickness, their consent was sought for the investigation of a variety of postoperative complications. All patients were admitted to the same ward the evening before surgery. Each patient was interviewed by one of the investigators as part of the routine preoperative assessment. A standard questionnaire was used to determine the fixed factors from the patient's history. Questions were distributed equally between those which were thought relevant to postoperative sickness and those which were not; the questions were asked in random order, to minimize bias. Patients were allocated randomly to receive either a light meal or no meal 6 h before anaesthesia. All patients received oral premedication with diazepam 10 mg 90 min before anaesthesia. Patients, lying supine, were taken by trolley from the ward to the anaesthetic room, so that all experienced the same journey. In the anaesthetic room, the patients were asked additional questions which sought information on whether or not the patient was hungry, thirsty, sleepy, anxious or feeling sick before induction of anaesthesia. Anaesthesia was induced with thiopentone 4 mg kg" 1, followed by oxygen 3 litre min" 1, nitrous oxide 6 litre min" 1 and increasing concentrations of enflurane via a Bain system. Care was taken to avoid apnoea, thus obviating the need for ventilatory assistance by mask. During induction, insertion of a Guedel airway and oropharyngeal suction were avoided. Patients who had ventilatory assistance, suction or insertion of an oropharyngeal airway at any time during the anaesthetic management were excluded from the analysis. No patient received supplementary local or regional anaesthesia. Anaesthesia was maintained by spontaneous ventilation of a mixture of 2% enflurane and 65% nitrous oxide in oxygen via a face mask and Bain system. Patients were monitored with non-invasive arterial pressure measurements, pulse oximetry and ECG. All patients received an i.v. infusion of Hartmann's solution during anaesthesia. The volume administered was calculated to compensate for the period of fasting and included any additional loss during surgery. The infusion was completed before the end of surgery. The volume for the fasting period was calculated as 1.5 ml kg" 1 li for example, a 70- kg patient who had fasted for 7 h received 735 ml. At the end of anaesthesia, a standard routine was followed for each patient. All patients were placed in the left lateral position, received supplementary oxygen therapy and were sent to a -h recovery

3 ASSESSING RISK OF POSTOPERATIVE SICKNESS 137 room with attention paid to the airway, avoiding oropharyngeal suction or insertion of a Guedel airway. If there was a need for additional airway intervention at this stage, the patient was withdrawn from the study. A total of 11 patients were excluded, for reasons of loss of postoperative data (six patients), variation from the study design (four patients) and inadvertent inclusion of one patient twice. Patients were allowed papaveretum, paracetamol and antiemetics as required. Consequently, postoperative opioid analgesia was considered a fixed factor for the purposes of this study and was not controlled. The total dose of analgesics administered was recorded over h. All episodes of postoperative sickness were recorded for h. Postoperative sickness was denned as vomiting (an integrated reflex which results in forceful expulsion of stomach contents), retching (a non-productive vomit) and nausea (the sensation of wishing to vomit or retch). No attempt was made to grade the severity of postoperative sickness. A "vomiter" was any patient who had at least one episode of nausea, retching or vomiting. Episodes of nausea, retching and vomiting were either volunteered, witnessed or elicited as the result of direct questioning by the nursing staff. Concomitant changes in heart rate, arterial pressure and ventilatory frequency were recorded with any obvious provocative episodes such as early ambulation. Oral fluids were not given until 4 h after anaesthesia, but were freely available thereafter. The same nurses observed all the patients (maximum four patients cared for by one recovery nurse) throughout the study. These nurses were aware of a study on postoperative complications but were not aware of its design or of the patients' responses to previous questioning. Table II shows a list of all the factors that were examined. Statistical analysis The data obtained from this study were either categorical, such as patient gender, or continuous, such as age, weight and duration of anaesthetic. The anaesthetic technique factors were strictly controlled, leaving the patient factors to distribute freely. All the continuous data were examined by Mann- Whitney U test between vomiters and non-vomiters. Significant statistical differences were assumed at P < All the categorical data that were not controlled by study design (fixed patient factors) were also examined initially by a univariate method (2x2 contingency tables) between vomiters and non-vomiters for rapid identification of possible important differences. The uncontrolled categorical data (fixed patient factors) and continuous data (age, weight, duration of fast and duration of operation) were examined by logistic regression (logit module, Systat 5.1 for the Macintosh, Systat, Illinois, U.S.A.). Logistic regression makes no assumption about distribution. The independent factors were indicated as 1 if present and 0 if absent (for gender 1 was female and 0 male). The inclusion of factors was determined by a forward stepwise regression of the main factors, followed by interactions. The threshold for inclusion of a factor was a probability of 0.15 or less based on log likelihood test ratio G{G = 2 (log likelihood test ratio for the model without the additional factor) (log likelihood test ratio for the model with the additional factor)) [12]. Maximum likelihood estimators (coefficients) were calculated for each factor from which odds ratios were determined. The odds ratios were expressed as probabilities for postoperative sickness. The probabilities of postoperative sickness for baseline and various factor combinations were calculated with their 95% confidence limits [12, 13]. Examples of these calculations are shown in the Appendix. RESULTS The uncontrolled factors (continuous and categorical) examined are listed in table II. The continuous variables, age, weight, duration of operation and duration of fast were distributed evenly between vomiters and non-vomiters (table III). These factors also failed to reach the inclusion criteria for the logistic regression model. Their coefficients were less than 0.01 and log likelihood test ratios gave P values between 0.4 and 0.6. The uncontrolled categorical factors (fixed patient factors) were also analysed initially by 2 x 2 contingency tables between vomiters and non-vomiters (table IV). Among these factors, only four appeared to have any association with the incidence of postoperative sickness when examined by univariate analysis: patient gender, a previous history of postoperative sickness, postoperative opioid analgesics, and a history of motion sickness (table IV). When forward stepwise logistic regression analysis was applied to the uncontrolled categorical data, the following independent factors met the criteria for inclusion in the model for postoperative sickness: postoperative opioid analgesia, gender, previous history of postoperative sickness, motion sickness and an independent interaction between gender and previous sickness history (table V). Motion sickness had the weakest association and unlike the other factors did not reach statistical significance at the 0.05 level, possibly because of low power, estimated at Assuming the same trend in the distribution of motion sickness among vomiters and nonvomiters, 800 patients would have been required to reach statistical significance. The relative effects of these factors may be summarized by the logistic regression equation: p) that is: Logit postoperative sickness = (postoperative opioids) (previous sickness history)+ 2.4(gender) (history of motion sickness) 3.2(genderx previous sickness history) where p = proportion of patients with sickness and 1 p = proportion without sickness. The presence of postoperative opioids, previous sickness history, female gender, and history of motion sickness are coded as 1 in the model; absence of these factors is coded as 0.

4 138 BRITISH JOURNAL OF ANAESTHESIA TABLE III. Patient data {median (range)) for vomiters and non-vomiters. No significant differences within gender groups Males Females Vomiters Non-vomiters Vomiters Non-vomiters Age (yr) Wt(kg) Duration of operation (min) Duration of fast (min) Total No. patients 41 (18-63) 75(66-86) 35(30-60) 660 ( ) 9 36 (16-78) 76 (54-128) 47.5(17-115) 540 ( ) 58 50(14-72) 62(41-98) 42(10-73) 620( ) (15-77) 65 (45-82) 36(20-115) 715( ) 49 TABLE IV. Univariate analysis of the categorical data Vomiter (No.) Non-vomiter (No.) Vomiter Non-vomiter Opioid analgesic (papaveretum) h postop. No opioid analgesic (papaveretum) h postop. Previous history of P No previous history of P Males Females Morning sickness No morning sickness History of motion sickness No history of motion sickness Breakfast Overnight fast, no breakfast Suffer with heartburn No heartburn Nervous at time of induction Not nervous at time of induction Hungry at time of induction Not hungry at time of induction Thirst at time of induction No thirst at time of induction Nausea in anaesthetic room before induction No nausea in anaesthetic room before induction Antiemetics postop. No antiemetics postop. Alcohol does not normally make feel sick Alcohol normally makes feel sick Pain does not normally make feel sick Pain normally makes feel sick TABLE V. Logistic regression analysis offactors associated with postoperative sickness during the first h after orthopaedic surgery (peripheral limb surgery). The overall odds ratio for emesis is given by the solution to: logit postoperative sickness = (postoperativc opioids) +3.97(previous sickness history)+ 2.4 (gender) (history of motion sickness) 3.2(genderx previous sickness history) where 5.03 is the constant. The presence of opioids, previous sickness history, motion sickness and female gender are coded as 1 in the above equation; absence of these factors or male gender is coded as 0 Independent factors influencing postoperative sickness Coefficient X SE (coeff.) Odds ratio e 95 % CI (odds ratio) P Constant Postoperative opioids Previous history of sickness after anaesthesia Effect of gender History of morion sickness Interaction of gender x previous sickness history < < < < < 0.2 > 0.1 <0.01 The log likelihood ratio test for the model produced a chi-square value of 53.5, P < (5 degrees of freedom). The most surprising finding was that, although males have an intrinsically lower probability of postoperative sickness than females, the effect of a previous history of sickness in a male increases the risk to greater than that for females (table VI).

5 ASSESSING RISK OF POSTOPERATIVE SICKNESS 139 TABLE VI. List of probability calculations for postoperative sickness depending on the combination of independent fixed patient factors. Baseline probability represents a male with no previous history of sickness or motion sickness who did not receive opioid analgesics. Baseline odds ratio for postoperative sickness is the solution to <* where x = The odds ratios (OR), 0.01, may be converted to a probability from P = OR// + OR, that iip = 0.01/1.01 = approximately 1 % probability for postoperative sickness Combination of independent factors influencing postoperative sickness Males Males with a history of motion sickness Males receiving postop. opioids Females Males with a history of motion sickness receiving postoperative opioids Females with a history of previous sickness Females with a history of motion sickness Females with previous sickness and history of motion sickness Males with a previous history of sickness Females receiving postoperative opioids Males with a previous history of sickness and motion sickness Females with previous history of sickness receiving postoperative opioids Females with history of motion sickness receiving postoperative opioids Females with a previous history of sickness and motion sickness receiving opioids Males with previous history of sickness receiving opioids Males with a previous history of sickness and motion sickness receiving opioids Probability of postop. sickness % CI of the probability of postop. sickness ^ ^ DISCUSSION Progress in identifying important factors for postoperative sickness may have become confounded by methodology that does not take account of fixed patient factors or their distribution, thereby introducing sample bias. The majority of studies have evaluated the merits of antiemetics in groups usually matched only by gender and anaesthetic technique. Further progress in the prevention of sickness may be facilitated if some method is adopted for assessing the inadvertent bias resulting from patient factors. The method we propose, based on logistic regression analysis, allows individual patient probability for postoperative sickness and the mean probability for a heterogenous group of patients to be calculated. The mean probability for sickness for groups of patients provides an additional comparison for within and between studies by examining the predicted and observed vomiting rates. Similar calculations might be preferable to historical controls for single group studies whereas, for very large studies, patients with similar probabilities may be risk-stratified into homogenous groups. Commonly, groups of patients with many influencing factors are described as "balanced" if the frequency of factors is the same in each study group. This can be conveniently described as global balancing. However, it would appear to be inappropriate to apply global balancing to groups made up of heterogenous patients if the probabilities for sickness caused by the factors are different and logarithmically related. This is the case for probabilities derived by logistic regression. This may be illustrated by an example. Consider two groups of 10 patients, each group hasfivewomen and five men; in addition, five of these patients (males, females or both) in each group have a previous history of sickness. These two groups would be globally balanced. However, using our own model, they would not be balanced by risk if, in one group, the previous history of sickness was only present in the five males whereas, in the other group, previous history of sickness was associated with the five females. The mean probability of sickness for the group in which previous sickness is associated with males would be 16.2% whereas, in the group where it is associated with females, the mean probability for the group would be This study has identified previous history of postoperative sickness, female gender, history of motion sickness, an interaction between gender and previous history and postoperative opioids as fixed patient risk factors. Postoperative opioid administration was included as a fixed factor: for ethical reasons it was not controlled, and therefore it distributed according to patient requirement. There is a practical difficulty in deciding if sickness in patients who have received opioids is caused by pain, or opioids, or if it would have occurred irrespective of these two factors. Even if sickness always occurred after administration of opioids, it is still difficult to decide the time that has to elapse before a bout of sickness can be ascribed to another cause. In this study, the analysis has shown an association between sickness and receipt of opioids at some time during a -h period. Fortunately, 55 patients did not receive opioids by choice, and this provided a control group. Therefore it is possible to be more confident of opioids as a cause of sickness. A criticism of this study may be the small sample from which coefficients were derived for the model. This decreases the power of the study and produces wide confidence limits, in particular when several factors are combined. We attempted to reduce our requirements for large numbers and yet maintain

6 140 BRITISH JOURNAL OF ANAESTHESIA power by strict control of anaesthetic technique. Additionally, a surgical procedure and anaesthetic method were chosen in which the potential for confounding emetic effects related to these procedures was kept to a minimum. In spite of the small numbers, the model obtained highly significant effects from postoperative opioids, gender, previous emetic history and the interaction between gender and previous emetic history. Motion sickness failed to reach statistical significance at 0.05, possibly because of low power (0.65), but met the standard inclusion criteria for the model suggested by Hosmer and Lemeshow [12]. The reason for the lack of power could, in part, be ascribed to the study design, which allowed opioids to be administered only after the journey to the -h observation area; the suggested vestibular sensitization by opioids may therefore have been averted [14]. Alternatively, the effect of motion sickness may be small when compared with other factors. The proposed model can only be considered useful if, when tested prospectively with patients other than those used in the study, it produces high correct classification rates for vomiters and non-vomiters. Therefore the statistical techniques of bootstrapping or jack-knifing which use the original data set to evaluate the ability of the model to classify patients correctly were not used. In evaluating the model, it should be noted that this study was limited to peripheral orthopaedic surgery and therefore the derived coefficients for fixed factors may need to be interpreted with caution when the model is applied to types of surgery having strong associations with postoperative sickness. APPENDIX The solution to the equation: Log. (p/1 p) that is: Logit postoperative sickness = (postoperative opioids)+ 3.97(previous sickness history) + 2.4(gender) (history of motion sickness) 3.2(gender x previous sickness history) (where p = proportion of patients with sickness; 1 p = proportion without sickness; presence of postoperative opioids, previous sickness history, female gender, and history of motion sickness are coded as 1 in the model, absence of these factors coded aso) provides the basis for calculating the probability of sickness within h in any particular patient. For calculating the risk of any group of patients, the individual patient probabilities should be summed and divided by the number of patients, to provide a mean value and so. Calculation of individual risks is shown in examples 1 and 2. Example 1 Baseline probability of sickness is calculated with all independent variables coded as 0: Logit postoperative sickness = constant = 5.03 From this the odds ratio can be calculated: Odds ratio = e"' 0> = 0.01 (The solution to e* where AT = 5.03 may be found in exponential tables.) Derivation of the probability (P) of sickness from the odds ratio is given by: P = odds ratio/1 +odds ratio In the above example, P is approximately 0.01 or 1 %. Example 2 Male with a previous history of sickness and motion sickness who also received opioids. Logit postoperative sickness = (l) (l) + 2.4(0) (l)-3.2(Oxl) = = 1.96 e 1 ' = 7.09 probability of sickness in h = 7.09/8.09 = or 87.6% The 95 % confidence limits for the probability of sickness was derived from the coefficient estimate il.%\/variance of the coefficient estimate. The upper and lower coefficient estimates thus calculated were converted to odds ratios using exponential tables, which in turn were used to calculate the 95 % confidence limits for the probability of sickness. The variance for a combination of risk factors is calculated from the sum of the independent factor estimate variances plus twice the covariances of the combined independent factor estimates. The 95 % confidence limits for the probability of sickness for various independent factors are shown in table VI. REFERENCES 1. Burtles R, Peckert BW. Postoperative vomiting. British Journal of Anaesthesia 1957; 29: Dent S, Ramachandra V, Stephen CR. Postoperative vomiting; incidence analysis and therapeutic measures in 3000 patients. Anesthesiology 1955; 16: Bonica JJ, Crepps W, Monk B, Bennett B. Postoperative nausea, retching and vomiting. Anesthesiology 1958; 19: Bellville JW, Bross IDJ, Howlands WS. Postoperative nausea and vomiting. IV, Factors related to postoperative nausea and vomiting. Anestheswlogy 1960; 21: Adriani J, Summers FW, Anthony SO. Is the prophylactic use of antiemetics in surgical patients justified? Journal of the American Medical Association 1961; 175: Kortrjla, K, Kauste A, Auvinen J. Comparison of domperidone, droperidol, and metoclopramide in the prevention and treatment of nausea and vomiting after balanced general anesthesia. Anesthesia and Analgesia 1979; 58: 396-^ McKie BD. Postoperative nausea and vomiting: A review of their incidence, causes and effects. Australian and New Zealand Journal of Surgery 1970; 39: Mortensen PT. Droperidol postoperative antiemetic effect when given intravenously to gynaecological patients. Acta Anaesthesiologica Scandinavica 1982; 26: Waters RM. The present state of cyclopropane. British Medical Journal 1936; 2: Palazzo MGA, Strunin L. Anaesthesia and emesis I. Etiology. Canadian Society of Anaesthetists Journal 1984; 31: Watcha M, White P. Postoperative nausea and vomiting. Its etiology, treatment and prevention. Anesthesiology 1992; 77: Hosmer D, Lemeshow S. Applied Logistic Regression. Chichester: John Wiley and Sons, Airman D. Practical Statistics for Medical Research. London: Chapman and Hall, Comroe J, Dripps, R. Reactions to morphine in ambulatory and bed patients. Surgery, Gynecology and Obstetrics 1948; 87: 1-2.