Improved Risk Adjustment for Comparison of Surgical Site Infection Rates
Objective. To develop prognostic models for improved risk adjustment in surgical site infection surveillance for 5 surgical procedures and to compare these models with the National Nosocomial Infection Surveillance system (NNIS) risk index.
Design. In a multicenter cohort study, prospective assessment of surgical site infection and risk factors was performed from 1996 to 2000. In addition, risk factors abstracted from patient files, available in a national medical register, were used. The c‐index was used to measure the ability of procedure‐specific logistic regression models to predict surgical site infection and to compare these models with models based on the NNIS risk index. A c‐index of 0.5 indicates no predictive power, and 1.0 indicates perfect predictive power.
Setting. Sixty‐two acute care hospitals in the Dutch national surveillance network for nosocomial infections.
Participants. Patients who underwent 1 of 5 procedures for which the predictive ability of the NNIS risk index was moderate: reconstruction of the aorta (
), femoropopliteal or femorotibial bypass (
), colectomy (
), primary total hip prosthesis (
), and cesarean section (
).
Results. The predictive power of the new model versus the NNIS index was 0.75 versus 0.62 for reconstruction of the aorta (
), 0.78 versus 0.58 for femoropopliteal or femorotibial bypass (
), 0.69 versus 0.62 for colectomy (
), 0.64 versus 0.56 for primary total hip prosthesis arthroplasty (
), and 0.70 versus 0.54 for cesarean section (
).
Conclusion. Data available from hospital information systems can be used to develop models that are better at predicting the risk of surgical site infection than the NNIS risk index. Additional data collection may be indicated for certain procedures—for example, total hip prosthesis arthroplasty.
Received May 4, 2006; accepted May 8, 2006; electronically published November 17, 2006.
If comparison of surgical site infection (SSI) rates between hospitals, between surgical teams, or over time is to produce insight and lead to changes in infection‐prevention practices, the rates must be corrected as much as possible for the differences in the patients’ intrinsic—that is, nonmodifiable—risk of SSI. The remaining differences may be taken to reflect differences in quality of care and thus serve as an incentive to optimize care.
Currently, the risk index developed by the Centers for Disease Control and Prevention (CDC) National Nosocomial Infection Surveillance system (NNIS) is the most widely used for risk adjustment of SSI rates, both in and outside of the United States.1,2 The NNIS risk index is an adaptation of the Study on the Efficacy of Nosocomial Infection Control (SENIC) risk index3 and incorporates 3 risk factors with equal weight—wound contamination class,4 American Society of Anesthesiologists (ASA) score,5 and duration of surgery. Although the NNIS risk index performs well when predicting the risk of SSI for grouped operations,1 its performance is less satisfactory for certain specific types of surgery. One problem is that, for some procedures, the 3 variables in the NNIS risk index are not risk factors or are not equally important and that other, more important factors are not taken into account. This problem can be understood considering how the NNIS risk index was developed—that is, not in a procedure‐specific way and not by multivariate modeling to determine relative weights of the risk factors. A second problem is that the duration of surgery may partly reflect quality of care. If it is incorporated in an index or model for risk adjustment, this may mask, rather than disclose, a potential problem in quality of care.
Several authors have recognized that risk adjustment needs to be improved and tailored to be procedure specific6‐8; others have presented results of studies to identify procedure‐specific risk factors for SSI—for example, in cesarean sections9,10 and colorectal surgery.11 However, none of these studies deals specifically with the development of prognostic models aimed at risk adjustment of SSI rates. Furthermore, for most of these studies, hospitals had to collect information about a large number of risk factors that was not routinely available. Since this is quite labor intensive, we examined whether data that are routinely collected in the Dutch surveillance network for nosocomial infections (Preventie Ziekenhuisinfecties door Surveillance [PREZIES]) and in the Dutch National Medical Register (NMR) could be combined to improve the prediction of SSI. We selected 5 procedures from different specialties for which the predictive performance of the NNIS index on the data in our surveillance database was low to moderate and for which a sufficient number of infections (more than 60) had been registered. These were reconstruction of the aorta, femoropopliteal or femorotibial bypass, colectomy, primary total hip prosthesis arthroplasty, and cesarean section.
Methods
PREZIES Network Data
Data were collected within the PREZIES network for surveillance of SSIs, according to surveillance methods described elsewhere12 and summarized here. For each patient under surveillance, infection control practitioners prospectively collected data on the patient, procedure, and any SSI present. Postdischarge surveillance through a registration card, which is filled in for each patient by the surgeon, complemented by chart review by the infection control practitioner was recommended but not compulsory. For the assessment of SSI, CDC criteria were used in Dutch translation.13,14
From January 1, 1996, until December 31, 2000, 62 PREZIES hospitals collected data on 24,046 patients who underwent a reconstruction of the aorta, femoropopliteal or femorotibial bypass, colectomy, primary total hip prosthesis arthroplasty, or cesarean section. Postdischarge surveillance was performed for 10,980 of these patients. Information was collected about the potential predictive factors sex, age, ASA score, wound contamination class, duration of preoperative stay in the hospital, timing of surgery (elective versus emergent), and complexity of surgery (single procedure versus multiple procedures during 1 theatre session). The performance of postdischarge surveillance was also considered a predictive factor, because, with longer follow‐up of patients, the likelihood of detecting an SSI increases.15
NMR Data
The Prismant Institute holds the NMR, which is a database of both ambulatory and clinical hospital admissions in the Netherlands. All Dutch hospitals participate in this register, with coverage of 100% of all general acute care hospitals. In the hospitals, NMR data are collected from patient files by trained data extractors. For this study, the NMR provided information on the indication for surgery, the presence of diabetes, and the number of diagnoses at discharge. The last was taken as a proxy for the number of comorbidities existing at admission. Therefore, discharge diagnoses with International Classification of Diseases, Clinical Modification (ICD‐9‐CM) codes 996 to 999, referring to complications of medical treatment, were excluded from the number of discharge diagnoses. For cesarean sections, the presence of pregnancy‐induced hypertension, gestational diabetes, premature rupture of membranes, and long duration of labor were also extracted from the NMR data.
Data Analysis
SAS for Windows, release 6.12 (SAS Institute), was used for all analyses. First, the PREZIES data were matched to the NMR database. Since a unique and shared identification number was not available for both data sets, records were matched on the combination of the variables type of surgery, date of birth, admission date, discharge date, hospital, and operation date plus or minus 2 days. We allowed this difference of 2 days between the operation dates because we assumed that some operations might have been postponed without being noted correctly in both data sets.
Second, the predictive performance of the NNIS risk index was assessed for each type of operation by calculating the c‐index of separate logistic regression models. The c‐index is a measure of predictive performance and represents the proportion of instances in which a patient who acquires an SSI is assigned a higher probability of SSI than is a patient who does not acquire an SSI. At a c‐index of 0.5, the ability of the model to predict the correct outcome for a patient (SSI or no SSI) is as good as flipping a coin; at a c‐index of 1, the model is perfect in predicting SSI.16
Third, logistic regression analysis was used to develop a new prognostic model for each type of operation. If the number of potential factors considered for inclusion in a prognostic model is large in comparison with the number of SSIs, the model will be overfitted, resulting in low predictive performance in other populations.17 Therefore, the set of eligible prognostic factors was restricted to those factors with a P value of less than .2 for their univariate association with SSI. Stepwise inclusion of risk factors was performed. With each step, the factor that resulted in the largest increase in the c‐index was added to the model, with a required minimum increase in the c‐index of 0.005. Since the optimal cutoff values for both the preoperative hospital stay and the number of discharge diagnoses was unknown and was expected to vary among the 5 types of surgery, several cutoff values were tested. The differences in c‐index were tested using the method described by Hanley and McNeill.18
Fourth, internal validation was performed to give insight into the remaining amount of overfitting and thus to assess the expected performance of the models in future populations. To this end, the models were cross‐validated as follows.19 For each of the 5 procedures, the total data sample was randomly divided into 10 mutually exclusive subsamples. One subsample was dropped, and the remaining 9 samples were used to derive the parameter estimates of the prognostic model. These estimates were then used to calculate the predictive probabilities of SSI for the remaining subsample of 10% of the data. This process was repeated 10 times until the predicted probabilities of SSI were computed for all 10 subsamples, thus constituting the whole data set. Next, c‐indexes for these 5 cross‐validated data sets were calculated.
Last, since both the teaching status and the volume of surgery of a hospital may influence the risk of SSI, the models were applied to subsets of data from teaching versus nonteaching hospitals and hospitals with high versus low volume of surgery. High volume of surgery was defined according to the type of operation, as the number of operations exceeding the median of the distribution of the hospitals’ annual number of operations. The Hosmer‐Lemeshow statistic was used to assess the goodness of fit of the models on these subsets.20 For this statistic, a P value of greater than .05 indicates a good fit.
Results
Patient Population
The percentage of PREZIES records that could be matched to NMR records was 68.5% for reconstruction of the aorta, 69.2% for femoropopliteal or femorotibial bypass, 70.9% for colectomy, 82.7% for primary total hip prosthesis arthroplasty, and 84.0% for cesarean section. Tables 1‐5 describe the matched patient populations for each type of operation, give the observed SSI rate, and give the results of development of the prognostic models. The distributions of the patient population over the different operation types was similar to the overall distribution in PREZIES (data not shown).
)
)
)
)
)Prognostic Models
The c‐indexes for the 5 prognostic models containing only the NNIS risk index categories were 0.62 for reconstruction of the aorta, 0.58 for femoropopliteal or femorotibial bypass, 0.62 for colectomy, 0.56 for primary total hip prosthesis arthroplasty, and 0.54 for cesarean sections. The new prognostic models that were developed for these 5 types of surgery are presented in tables 1‐5. For reconstruction of the aorta, a model containing the ASA score, diabetes, preoperative hospital stay, indication for surgery, complexity of surgery, and the number of discharge diagnoses was best in predicting SSI. Its c‐index was 0.75, a difference of 0.13 from the NNIS‐index model (
). A model including sex, indication for surgery, complexity of surgery, number of discharge diagnoses, and performance of postdischarge surveillance best predicted SSIs occurring after femoropopliteal or femorotibial bypass, with a c‐index of 0.78. This was the type of surgery for which the largest improvement of predictive power could be obtained—that is, an increase of 0.20 over the NNIS risk index model (
). The smallest improvement was obtained for colectomy. Here, the combination of ASA score, indication for surgery, wound contamination class, and number of discharge diagnoses showed the highest predictive power (c‐index of 0.69, a difference of 0.07 from the NNIS model; 
). For primary total hip prosthesis arthroplasty, the predictive ability of the new model, containing age, preoperative hospital stay, number of discharge diagnoses, and postdischarge surveillance remained low (c‐index of 0.64, a difference of 0.08 from the NNIS model; 
). For cesarean section, however, a reasonable increase in predictive power of 0.16, to a c‐index value of 0.70, was achieved by use of ASA score, pregnancy‐induced hypertension, indication for surgery, and postdischarge surveillance as prognostic factors (
).
Evaluation
Cross‐validation of the 5 models resulted in a reduction of the c‐index of 0.07 for reconstruction of the aorta, 0.06 for femoropopliteal or femorotibial bypass, 0.03 for colectomy, 0.04 for cesarean section, and 0.02 for primary total hip prosthesis arthroplasty. Except for reconstruction of the aorta (P = .16) and colectomy (
), these cross‐validated models still showed statistically significantly better predictive performance than did the NNIS model (P values <.05)
Application of the models to subsets of hospitals with high and low volumes of surgery and of teaching and nonteaching hospitals resulted in satisfactory goodness of fit for 19 of 20 subsets. Insufficient goodness of fit was obtained for cesarean section in nonteaching hospitals, as indicated by a P value of .02 for the Hosmer‐Lemeshow statistic.
Discussion
The currently widely used NNIS risk index showed poor predictive performance for 5 commonly performed surgical procedures in our SSI surveillance network. We found that models developed with routinely available data markedly improve the prediction of SSI for these 5 types of surgery, although not to the same extent. With our method, enhanced risk adjustment of SSI rates will be possible for reconstruction of the aorta, femoropopliteal or femorotibial bypass, colectomy, and cesarean section. Additional data collection about prognostic factors may be indicated for certain operations—for example, primary total hip prosthesis arthroplasty—to achieve acceptable predictive capacity.
To appreciate these findings, a few remarks have to be made. The data used in this study have been collected by 62 different hospitals. Although this warrants the inclusion of a broad group of patients, it also creates an opportunity for differences in data collection, possibly resulting in loss of power to identify important prognostic factors. We think this is unlikely, however, because both in the PREZIES network and in the NMR, persons responsible for data collection are regularly trained in the application of definitions and coding. Also, on‐site validation of the data collection process and the use of criteria for the assessment of SSI is performed in PREZIES hospitals, indicating good quality of data.21
With the rather strict strategy we applied to match records from both data sets, only a relatively small percentage of PREZIES records could be matched to NMR information. For a PREZIES record to be matched to a NMR record, all dates except the operation date had to be exactly the same. Although we were able to obtain higher matching proportions when less strict matching criteria were applied, we considered it more important that the data be of good quality than to obtain as big a data set as possible.
Our evaluation showed that the models could be applied to subsets of teaching, nonteaching, high‐volume–surgery, and low‐volume–surgery hospitals. Internal validation indicated that the predictive ability in future populations might be slightly lower, thus warranting revalidation of the models in those populations.
The potential predictive factors that were included in this study have been identified as risk factors in other studies.11‐13,22‐28 However, to our knowledge, no other reports exist in the literature that deal specifically with the development of prognostic models aimed at risk adjustment of SSI rates for the 5 procedures in our study.
Therefore, we compared our models with the risk‐adjustment method that is currently most widely used in surveillance networks, the NNIS risk index. This index tries to capture the intrinsic risk of SSI attributable to the patient’s condition, by the ASA score, and the intrinsic risk associated with the procedure, by the wound contamination class and the duration of surgery.
Of the potential prognostic factors in our study, ASA score, the number of discharge diagnoses, diabetes, possibly the preoperative hospital stay, and, to a lesser extent, the indication for surgery are directly related to the patient’s condition. Our results show that the patient's underlying condition cannot easily be captured within one prognostic factor, since, for all procedures, 2 or more of these factors were included in the prognostic model. The number of discharge diagnoses seemed to be especially important, because it was a strong prognostic factor for all types of procedures except cesarean section. For pooled data for all procedures, this independent association was also observed in the SENIC study.29
The wound contamination class did not contribute much to prediction of SSI, except for colectomy. This can be understood from the fact that, for prognostic models, the change in predictive power of the model as indicated by the c‐index has to be the leading inclusion criterion for a factor, instead of the size of its parameter estimate or risk ratio. Because of this, strong risk factors are excluded if they are very uncommon. This was the case for wound contamination class in all procedure types except colectomy, with less than 2% of all operations being contaminated or dirty. For reasons explained in the introduction, we did not consider duration of surgery as a potential prognostic factor.
Whereas prediction of SSI improved considerably with our models, the resulting c‐indexes still remained relatively low. It must be emphasized, however, that this may result from the perspective chosen. The prediction rules are meant to adjust SSI rates only for intrinsic characteristics of the patient population, so potential risk factors reflecting the process of care were deliberately excluded from the models. If large variation exists in quality of care among hospitals, prediction of SSI solely on the basis of factors other than process‐of‐care characteristics cannot be perfect.
The models presented here can be used for risk adjustment of SSI rates, by calculation of the standardized infection ratio. This is the ratio of the observed SSI incidence divided by the expected SSI incidence. A ratio of greater than 1 indicates that more SSIs are observed than would be expected on the basis of the characteristics of the patient population. To calculate the expected SSI incidence for a group of patients, the probabilities of SSI for each of the patients in the group need to be averaged. These individual probabilities can be calculated by filling in the characteristics of a patient as shown in tables 1‐5 into the prediction rule. For example, for a patient with no severe systemic disease who underwent a contaminated colectomy because of colitis and who had 3 diagnoses at discharge, the individual logit(p) would be −2.89 plus 0 (for low ASA score) plus 0.22 (for colitis) plus 0.60 (for contaminated surgery) plus 0 (for 3 discharge diagnoses), which equals −2.07. By solving the equation 
, the individual probability can be calculated: 
.
Our findings show that significantly improved prognostic models for SSI can be developed. Indication for surgery and existing comorbidities were important predictors of SSI for all procedures. For this study, no labor‐intensive collection of additional data was necessary, since information was obtained from an already‐existing morbidity database. But these national data are only available with a delay. However, the data are usually available from the medical registration department in the hospital, and thus it would be possible for infection control practitioners to obtain these data.
This extra effort required must be weighed against the potential benefits. SSI rates can only serve as an indicator of quality of care if they are accepted by the people involved in changing the day‐to‐day practices of prevention of SSI within the hospital—namely, physicians and nurses. Improved risk adjustment for procedures for which the NNIS risk index has little discriminatory power may be more convincing to clinicians and thus may be more fit to stimulate changes in infection‐prevention practices. Also, improved risk adjustment of SSI rates will become more important in cases of public disclosure of SSI rates.
Acknowledgments
We gratefully acknowledge all infection control practitioners, medical specialists, and nurses of the following hospitals, for their contribution to the data collection: 't Lange Land Hospital, Zoetermeer; Albert Schweitzer Hospital at Amstelwijck, Dordrecht; Albert Schweitzer Hospital at Dordwijck, Dordrecht; Albert Schweitzer Hospital at Zwijndrecht, Zwijndrecht; Amphia Hospital at Pasteurlaan, Oosterhout; Beatrix Hospital, Gorinchem; Bernhoven Hospital at St. Anna, Oss; Bernhoven Hospital at St. Joseph, Veghel; Boven IJ Hospital, Amsterdam; Canisius Wilhelmina Hospital, Nijmegen; Carolus‐Liduina‐Lindenlust Hospitals Group at Carolus, Den Bosch; Catharina Hospital, Eindhoven; Delfzicht Hospital, Delfzijl; Deventer Hospitals Group, Deventer; Diaconessen Hospital, Meppel; Elkerliek Hospital, Helmond; Erasmus Medical Center at Daniël den Hoed Clinic, Rotterdam; Gelre Hospitals, Apeldoorn; Groene Hart Hospital, Gouda; Harbour Hospital, Rotterdam; Hofpoort Hospital, Woerden; Hospital De Gelderse Vallei at Bennekom, Bennekom; Hospital Hilversum, Hilversum; Hospital Nij Smellinghe, Drachten; Hospital Rivierenland, Tiel; Hospital Walcheren, Vlissingen; Isala Hospitals at Weezenlanden, Zwolle; Isala Hospitals at Sophia, Zwolle; Jeroen Bosch Hospital, Den Bosch; Kennemer Hospital at Deo, Haarlem; Kennemer Hospital at Elisabeth, Haarlem; Leiden University Hospital, Leiden; Leyenburg Hospital, The Hague; Martini Hospital, Groningen; Medical Center Alkmaar, Alkmaar; Medical Center Haaglanden at Antoniushove, Leidschendam; Medical Center Molendael, Baarn; Medical Center Leeuwarden South, Leeuwarden; Medical Spectrum Twente at Oldenzaal, Oldenzaal; Mesos Medical Center at Overvecht, Utrecht; Northern Limburg Hospitals at Venlo, Venlo; Oosterschelde Hospitals Group, Goes; Red Cross Hospital, Beverwijk; Regional Hospital Coevorden‐Hardenberg, Coevorden; Regional Hospital Queen Beatrix, Winterswijk; Regional Hospital Zevenaar, Zevenaar; Rijnstate Hospital, Arnhem; Slingeland Hospital, Doetinchem; Spaarne Hospital, Heemstede; St. Antonius Hospital, Nieuwegein; St. Elisabeth Hospital, Tilburg; St. Franciscus Hospital, Rotterdam; St. Jans Hospital, Weert; St. Lucas Andreas at Lucas, Amsterdam; University Hospital Groningen, Groningen; University Hospital Maastricht, Maastricht; Vlietland Hospital at Schieland, Schiedam; Vlietland Hospital at Holy, Vlaardingen; Vrije Universiteit Medical Center, Amsterdam; Westeinde Hospital, The Hague; Wilhelmina Hospital, Assen; and Zuider Hospital, Rotterdam.
We thank the Ministry of Health, Welfare, and Sports for their financial support of the PREZIES network.
References
- 1. Culver DH, Horan TC, Gaynes RP, et al. Surgical wound infection rates by wound class, operative procedure, and patient risk index. National Nosocomial Infections Surveillance System. Am J Med 1991; 91:152S‐157S.
- 2. Coello R, Gastmeier P, de Boer AS. Surveillance of hospital‐acquired infection in England, Germany, and The Netherlands: will international comparison of rates be possible? Infect Control Hosp Epidemiol 2001; 22:393‐397.
- 3. Haley RW, Culver DH, Morgan WM, White JW, Emori TG, Hooton TM. Identifying patients at high risk of surgical wound infection: a simple multivariate index of patient susceptibility and wound contamination. Am J Epidemiol 1985; 121:206‐215.
- 4. Altemeier WA, Burke JF, Pruitt BA, Sandusky WR. Manual on control of infection in surgical patients. 2nd ed. Philadelphia, PA: JB Lippincott, 1984.
- 5. American Society of Anesthesiologists. New classification of physical status. Anesthesiology 1963; 24:11.
- 6. Gaynes RP. Surgical‐site infections and the NNIS SSI Risk Index: room for improvement. Infect Control Hosp Epidemiol 2000; 21:184‐185.
- 7. Gaynes RP. Surgical‐site infections (SSI) and the NNIS Basic SSI Risk Index, part II: room for improvement. Infect Control Hosp Epidemiol 2001; 22:266‐267.
- 8. Roy MC, Herwaldt LA, Embrey R, Kuhns K, Wenzel RP, Perl TM. Does the Centers for Disease Control’s NNIS system risk index stratify patients undergoing cardiothoracic operations by their risk of surgical‐site infection? Infect Control Hosp Epidemiol 2000; 21:186‐190.
- 9. Killian CA, Graffunder EM, Vinciguerra TJ, Venezia RA. Risk factors for surgical‐site infections following cesarean section. Infect Control Hosp Epidemiol 2001; 22:613‐617.
- 10. Tran TS, Jamulitrat S, Chongsuvivatwong V, Geater A. Risk factors for postcesarean surgical site infection. Obstet Gynecol 2000; 95:367‐371.
- 11. Tang R, Chen HH, Wang YL, et al. Risk factors for surgical site infection after elective resection of the colon and rectum: a single‐center prospective study of 2,809 consecutive patients. Ann Surg 2001; 234:181‐189.
- 12. Geubbels ELPE. Prevention of surgical site infections through surveillance. Utrecht, The Netherlands: Utrecht University, 2002.
- 13. Garner JS, Jarvis WR, Emori TG, Horan TC, Hughes JM. CDC definitions for nosocomial infections, 1988. Am J Infect Control 1988; 16:128‐140.
- 14. Horan TC, Gaynes RP, Martone WJ, Jarvis WR, Emori TG. CDC definitions of nosocomial surgical site infections, 1992: a modification of CDC definitions of surgical wound infections. Infect Control Hosp Epidemiol 1992; 13:606‐608.
- 15. Sands K, Vineyard G, Platt R. Surgical site infections occurring after hospital discharge. J Infect Dis 1996; 173:963‐970.
- 16. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 1982; 143:29‐36.
- 17. Harrell FE Jr, Lee KL, Mark DB. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med 1996; 15:361‐387.
- 18. Hanley JA, McNeil BJ. A method of comparing the areas under receiver operating characteristic curves derived from the same cases. Radiology 1983; 148:839‐843.
- 19. Ash AS, Shwartz M. Evaluating the performance of risk‐adjustment methods: dichotomous outcomes. In: Iezzoni LI, ed. Risk‐adjustment for measuring healthcare outcomes. 2nd ed. Chicago: Health Administration Press, 1997:460.
- 20. Hosmer DW, Lemeshow S. Applied logistic regression. 1st ed. New York: John Wiley & Sons, 1989.
- 21. Manniën J, van der Zeeuw AE, van den Hof S, Wille JC. Validation of surgical site infection surveillance in the Netherlands. Infect Control Hosp Epidemiol 2005; 27:XXX‐XX (in this issue).
- 22. Bunt TJ. Vascular graft infections: an update. Cardiovasc Surg 2001; 9:225‐233.
- 23. Richet HM, Chidiac C, Prat A, et al. Analysis of risk factors for surgical wound infections following vascular surgery. Am J Med 1991; 91:170S–172S.
- 24. Edwards WH Jr, Kaiser AB, Tapper S, et al. Cefamandole versus cefazolin in vascular surgical wound infection prophylaxis: cost‐effectiveness and risk factors. J Vasc Surg 1993; 18:470‐475; discussion 475‐476.
- 25. Miransky J, Ruo L, Nicoletta S, et al. Impact of a surgeon‐trained observer on accuracy of colorectal surgical site infection rates. Dis Colon Rectum 2001; 44:1100‐1105.
- 26. Siegman‐Igra Y, Rozin R, Simchen E. Determinants of wound infection in gastrointestinal operations: the Israeli study of surgical infections. J Clin Epidemiol 1993; 46:133‐140.
- 27. Poss R, Thornhill TS, Ewald FC, Thomas WH, Batte NJ, Sledge CB. Factors influencing the incidence and outcome of infection following total joint arthroplasty. Clin Orthop 1984; 182:117‐126.
- 28. Wong ES. Surgical site infections. In: Mayhall CG, ed. Hospital epidemiology and infection control. Baltimore: Williams & Wilkins, 1996:154‐175.
- 29. Haley RW, Culver DH, White JW, et al. The efficacy of infection surveillance and control programs in preventing nosocomial infections in US hospitals. Am J Epidemiol 1985; 121:182‐205.
-
Presented in part: 12th Annual Meeting of the Society for Healthcare Epidemiology of America, Salt Lake City, Utah, April 6‐9, 2002.