Clinical Predictors of Major Infections After Cardiac Surgery
Background— Major infections are infrequent but important complications of cardiac surgery. Predicting their occurrence is essential for future prevention. The objective of the current investigation was to create and validate a bedside scoring system to estimate patient risk for major infection (mediastinitis, thoracotomy or vein harvest site infection, or septicemia) after coronary artery bypass grafting.
Methods and Results— Using the Society of Thoracic Surgeons National Cardiac Database, we analyzed 331 429 coronary artery bypass grafting cases from January 1, 2002, to December 31, 2003, to identify risk factors for major infection. Using logistic regression, 2 models were generated and validated using split-sample validation: (1) One limited to preoperative characteristics (preop model) and (2) one model including both preoperative and intraoperative characteristics (combined model). Major infection occurred in 11 636 patients (3.51%) (25.1% mediastinitis, 32.6% saphenous harvest site, 35.0% septicemia, 0.5% thoracotomy, 6.8% multiple sites). Patients with major infection had significantly higher mortality (17.3% versus 3.0%, P<0.0001) and postoperative length of stay >14 days (47.0% versus 5.9%, P<0.0001) than patients without major infection. Both the preop model (c-index 0.697) and combined model (c-index: 0.708) successfully discriminated between high- and low-risk patients. A simplified risk scoring system of 12 variables accurately predicted risk for major infection.
Conclusions– We identified and validated a model that can identify patients undergoing cardiac surgery who are at high risk for major infection. These high-risk patients may be targeted for perioperative intervention strategies to reduce rates of major infection.
Infection in the setting of cardiac surgery increases morbidity, mortality, and cost. These infections can require prolonged treatment with antibiotics, additional surgery, or both. The proportion of coronary artery bypass patients at high-risk for infection is increasing because of the aging US population, a growing number of patients undergoing “redo” procedures, and the frequency of conditions conferring both cardiovascular and infectious risks (obesity, diabetes mellitus) among this population. Thus, there is a critical need to identify patients undergoing cardiac surgery who are at risk for major infections and to develop effective interventions to prevent these infections.
Although prior studies have identified risk factors for postoperative mediastinitis,1–6 many of these characteristics could not be determined before surgery, the time of greatest opportunity for intervention strategies. Further, no study to date has developed a simplified scoring system to estimate an individual patient’s risk for major infection after coronary artery bypass grafting (CABG).
We used the Society of Thoracic Surgeons (STS) National Cardiac Database to evaluate major infection in more than 300 000 patients who had undergone CABG procedures. The objectives of this investigation were to identify the frequency of major infection after CABG, identify determinants of major infection among patients undergoing CABG, and convert these determinants into a bedside scoring system to estimate a patient’s risk of major infection after CABG.
The STS National Cardiac Database was established in 1989 to report surgical outcomes after cardiothoracic surgical procedures.7,8 The database currently captures clinical information from nearly two thirds of all US bypass procedures from more than half of all centers performing adult cardiac surgery. Sites enter patient data using uniform definitions (available online at http://www.sts.org) and certified software systems. This information is sent semiannually to the STS Data Warehouse and Analysis Center at the Duke Clinical Research Institute. There, a series of data quality checks are performed before a site’s data are aggregated into the national sample. Data completeness in the STS database is high and highly accurate.9
All patients in the STS adult cardiac database undergoing CABG surgery (isolated or in combination) from January 1, 2002, to December 31, 2003, were considered for inclusion in the study. Records collected under previous data versions, patients undergoing cardiac surgery for infectious endocarditis or heart transplantation, or patients with records not collected on the STS standard data collection form (STS version 2.41) were excluded from analysis.
Two logistic regression models were developed to estimate the risk of major infection after CABG. The first model (preop model) included only preoperative patient characteristics. The second model (combined model) included both preoperative patient characteristics and intraoperative variables. Predictor variables were identified on the basis of clinical judgment and previous literature. Full models were fit using all available explanatory variables. Simplified models were subsequently developed by removing predictors from the model and rounding regression coefficients to integers.
Clinical End Point
Our primary end point, major infection, was defined as one or more of the following: (1) Surgical site infection (deep sternal wound, thoracotomy, or leg vein harvest site) or septicemia before discharge, or (2) readmission within 30 days of surgery for deep sternal wound infection, leg wound infection, or septicemia. Infection was documented by at least 1 of the following: (1) Wound opened with excision of tissue (incision and drainage); (2) positive culture; or (3) treated with antibiotics. Additional data definitions are available at www.sts.org.
Data were randomly split 50/50 into a training (development) sample and validation (test) sample. Only data in the training sample were used for developing the models described here. All of the explanatory variables were retained in the initial preop only and combined models. The statistical significance of each predictor variable was measured by the Wald χ2 statistic.
Both models overestimated the risk of infection among high-risk patients (predicted risk >10%) in the training sample. To improve model calibration, risk scores were obtained for each subject by assigning the predicted log odds from the original model shown. Then, for each model, the log odds of infection was modeled as a 2-phase linear (spline) function of the risk score. The join point for the 2-phase linear function was placed at the location corresponding to 9% predicted risk. This recalibration greatly improved the fit of both models. The recalibrated models were used for subsequent development of a simplified risk scoring system.
Risk Scoring System
The 2 methods of simplifying the model considered were omitting predictor variables and rounding regression coefficients to integers. The loss of information due to simplifying the model was measured by computing the Pearson correlation coefficient between the predicted log odds based on the original versus simplified models. Following standard practice, regression coefficients were multiplied by a scaling factor before rounding them. The loss of information due to rounding coefficients can be made arbitrarily small by increasing the scaling factor; however, this increases the complexity of the resulting risk scores. The number of variables and the scaling factor were subjectively chosen to optimize the trade-off between ease-of-use versus accuracy (measured by the Pearson correlation).
Risk scores were assigned to patients by summing the rounded regression coefficients across risk factors. An estimate of risk was obtained for each patient by averaging the predicted values from the original full model among all patients having the same risk score. The heterogeneity of risk among patients having the same risk score was conveyed by plotting the 5th and 95th percentiles of the distribution of predicted risk on the basis of the original full model.
Validation of Predictive Accuracy
Calibration was assessed in the validation sample by comparing observed versus predicted rates of infection within subgroups of patients defined by predicted risk. The ability to discriminate between high-risk and low-risk patients was assessed by computing the area under the receiver operating characteristic curve (c-index) using patients in the validation sample. To increase precision, the training and validation samples were combined after model validation, and estimates were obtained using the entire sample.
All 370 133 patients in the STS adult cardiac database undergoing CABG surgery during the study period were considered for inclusion. Records collected under previous data versions (n=37 348), from patients with infectious endocarditis (n=1,345), or coded as “CABG + heart transplant surgery” (n=11) were excluded from the analysis, leaving a final study population of 331 429. Of these, 11 636 patients (3.51%) were identified as having major infection during the study period (25.1% mediastinitis, 32.6% vein harvest site infection, 35.0% septicemia, 0.5% thoracotomy site infection, 6.8% multiple sites).
Patients with major infection had significantly higher mortality (17.3% versus 3.0%, P<0.0001) and postoperative length of stay >14 days (47.0% versus 5.9%, P<0.0001) than patients without major infection.
Demographic Characteristics and Bivariate Analyses
A number of factors were associated with major infection by bivariate analysis (Table 1). The most common risk factors included a body mass index (BMI) of 30 to 40 kg/m2 (32.03%; odds ratio [OR]: 1.44; 95% confidence interval [CI]: 1.39 to 1.50), diabetes mellitus (34.85%; OR: 1.78; 95% CI: 1.71 to 1.85), previous myocardial infarction (43.69%; OR: 1.40; 95% CI: 1.35 to 1.46), urgent operative status (43.35%; OR: 1.27; 95% CI: 1.22 to 1.32), and hypertension (75.72%; OR: 1.38; 95% CI: 1.32 to 1.45). Several factors were uncommon but strongly associated with major infection; they included cardiogenic shock (2.07%; OR: 3.11; 95% CI: 2.86 to 3.37), dialysis dependence (1.55%; OR: 2.99; 95% CI: 2.71 to 3.29); perfusion time of 200 to 300 minutes (3.70%; OR: 3.56; 95% CI: 3.28 to 3.87); and immunosuppressive therapy (2.02%; OR: 2.03; 95% CI: 1.84 to 2.23).
Multivariate Model Limited to Preoperative Characteristics
The full preop model provided robust predictive ability in the derivation cohort (n=166 043; c-index 0.697) (Table 2). Risk factors associated with major infection in the preop model included BMI >40 kg/m2 (OR: 2.95; 95% CI: 2.75 to 3.17); hemodialysis (OR: 2.12; 95% CI: 1.91 to 2.35); cardiogenic shock (OR: 2.11; 95% CI: 1.93 to 2.31); age >85 years (OR: 1.97; 95% CI: 1.73 to 2.23); immunosuppressive treatment (OR: 1.51; 95% CI: 1.36 to 1.66); and diabetes mellitus (OR: 1.44; 95% CI: 1.38 to 1.49). Risk for major infection was distributed asymmetrically throughout the cohort, with only 19.49% of patients in the preop model having an estimated risk for major infection of >5%.
Multivariate Model of Both Preoperative and Intraoperative Characteristics
Next, characteristics identified during the time of CABG (eg, intraoperative characteristics) were added to the preop model. Although only a small incremental increase in predictive ability was provided by adding intraoperative characteristics to the preop model (combined c-index 0.711), several additional risk factors were identified. These additional factors included perfusion time 200 to 300 minutes (OR: 2.08; 95% CI: 1.89 to 2.29); perfusion time >300 minutes (OR: 1.96; 95% CI: 1.65 to 2.34); placement of an intra-aortic balloon pump (OR: 1.53: 95% CI: 1.44 to 1.63); and the presence of 3 or more distal anastomoses (OR: 1.28; 95% CI: 1.19 to 1.39) (Table 2).
Validation of Model Accuracy
The recalibrated preop and combined models were assessed by examining calibration and discrimination in the validation sample (n =165 386). Overall, there was good agreement between observed versus predicted rates of infection, as well as good discrimination (c-index: 0.697 for the preop model, 0.708 for the combined model).
Creation of Risk Score Models
Simplified risk scores ranged from 1 point (for each 5 years above 55 years of age) to 9 points (for a BMI ≥40 in the preop model) (Table 3). The estimated risk of major infection for a future patient was obtained by summing the total number of risk points present and identifying the corresponding probability of infection from Table 4. Eighty percent of the study population had a risk score of ≤14 (<5% estimated risk) (Figure 1).
Validation of Risk Score Models
The preop and combined risk score models both exhibited robust predictive ability in the derivation set, with only minimal additional predictive ability provided by inclusion of the intraoperative variables. In the validation sample, there was good agreement between observed and predicted rates of infection for both models (Figure 2). The c-index was 0.686 for the preop model and 0.696 for the combined model.
Accuracy of Model Excluding Leg Infection From End Point
Because the severity of vein harvest site infections could be disputed, the analyses were repeated using a modified end point that excluded leg infection. Following the same analysis protocol, the same sets of variables were identified in the risk score models, and the external validation of the risk score models remained robust (c-index 0.697 preop, 0.703 combined). The original risk scoring system predicted the modified end point well (externally validated c-index 0.691 preop, 0.700 combined). Thus, although the model was developed using an end point that included leg infection, it had slightly better discrimination ability for an end point that excluded leg infection.
Major infection is an uncommon but potentially devastating complication of cardiac surgery. In the current investigation, we used clinical information from almost two thirds of all US bypass procedures performed during the study period to develop and validate a simple scoring system that can be used at the bedside to estimate an individual patient’s risk of developing major infection. This investigation has demonstrated several key findings.
Major infection was relatively uncommon in the STS Database, with ≈3.5% of patients developing this complication after CABG. Rates of leg wound infection and septicemia in the current study were similar to prior reports.5,10 Rates of mediastinitis, ≈0.6% in the STS Database, are consistent with some reports2,11 but lower than others.1,6,12 The difference in rates of mediastinitis may reflect in part the large, geographically diverse population within the STS database, which increases generalizability. Alternately, lower rates of mediastinitis in the STS Database may reflect differences in identification of infectious complications of CABG. Rates of mediastinitis may be higher in centers with active infection control surveillance rather than voluntary reporting, as occurs with STS Database. The lower rate of mediastinitis may also reflect the fact that the STS Database primarily captures acute events, whereas some infections may only become apparent weeks after surgery.13
Several variables in this investigation were particularly important risk factors for infectious complications after cardiac surgery. Obesity was a significant risk factor in both the preop and combined model. The presence of BMI >40 kg/m2 alone conferred a 2.6% risk for major infection in the preop model, a finding that agrees with some,2,14,15 but not all,16 prior reports. Potential explanations for the impact of obesity include inadequate serum levels of prophylactic antibiotics, technical difficulties in maintaining sterility of tissue folds, and poor perfusion of adipose tissue. The importance of diabetes mellitus as an independent risk factor for major infection was again demonstrated.2,12,14,17,18 Although characteristics such as cardiogenic shock, perfusion time of 200 to 300 minutes, and placement of an intraarterial balloon pump were associated with the highest risk for major infection, their overall impact on the models were limited by their infrequent occurrence. For example, all 3 variables occurred in less than 10% of the study population. By contrast, the impact of obesity, diabetes, and congestive heart failure was significant because of the frequency of several conditions: Over one third of study patients had a BMI >30 kg/m2, one third had diabetes mellitus, and ≈18% had congestive heart failure. Given the epidemic proportions of obesity in the US, it seems likely that the importance of these risk factors for major infection after CABG will only increase.
This investigation has limitations. First, the model did not identify infections according to the specific causative pathogen and included infectious complications, such as sepsis, other than mediastinitis. These broader definitions, however, reflect the overall clinical objective. The potential for underreporting infectious complications exists, particularly with infections that are less serious, occur after discharge or more than 30 days after surgery, or present to hospitals other than where the CABG was performed. The study had limited ability to evaluate detailed preoperative or intraoperative process data such as the type, timing, and dosage of antibiotic prophylaxis, the use of intravenous insulin infusions for diabetic patients, or preoperative preparations. Finally, an ideal prediction model requires prospective validation.19
This investigation has important clinical implications. Before this study, it was difficult to preoperatively estimate an individual patient’s risk for infectious complications. This investigation suggests that a relatively small proportion of patients undergoing CABG bear the majority of risk for major infection, that high-risk patients are identifiable before surgery, and that several of these risks are potentially modifiable. For example, continuous intravenous insulin therapy designed to maintain blood glucose levels <150 mg/dL reduces the risk of deep sternal wound infection by 66% among diabetic patients.18 Other strategies could include clinical process improvements,20 weight loss and/or smoking cessation efforts, and interventions targeting Staphylococcus aureus, including nasal decolonization21 and vaccines.22 Finally, preoperative risk assessments could be used to identify at-risk populations in whom clinical trials are cost-effective.23
In conclusion, the current investigation has identified and validated several risk factors available to the clinician in the preoperative setting that may be used to identify patients at risk for major infection. This risk score accurately identifies high-risk patients who may benefit from targeted interventions to reduce this devastating complication of cardiac surgery.
This work was supported by grant AI-059111 (Dr Fowler) from National Institutes of Health.
Fowler VG Jr, Kaye KS, Simel DL, Cabell CH, McClachlan D, Smith PK, Levin S, Sexton DJ, Reller LB, Corey GR, Oddone EZ. Staphylococcus aureus bacteremia after median sternotomy: clinical utility of blood culture results in the identification of postoperative mediastinitis. Circulation. 2003; 108: 73–78.
Athanasiou T, Aziz O, Skapinakis P, Perunovic B, Hart J, Crossman MC, Gorgoulis V, Glenville B, Casula R. Leg wound infection after coronary artery bypass grafting: a meta-analysis comparing minimally invasive versus conventional vein harvesting. Ann Thorac Surg. 2003; 76: 2141–2146.
Jonkers D, Elenbaas T, Terporten P, Nieman F, Stobberingh E. Prevalence of 90-days postoperative wound infections after cardiac surgery. Eur J Cardio-Thorac Surg. 2003; 23: 97–102.
Milano CA, Kesler K, Archibald N, Sexton DJ, Jones RH. Mediastinitis after coronary artery bypass graft surgery: risk factors and long-term survival. Circulation. 1995; 92: 2245–2251.
Moulton MJ, Creswell LL, Mackey ME, Cox JL, Rosenbloom M. Obesity is not a risk factor for significant adverse outcomes after cardiac surgery. Circulation. 1996; 94 (suppl II): II-87–II-92.
Furnary AP, Wu Y, Bookin SO. Effect of hyperglycemia and continuous intravenous insulin infusions on outcomes of cardiac surgical procedures: the Portland Diabetic Project. Endocrine Practice. 2004; 10 (suppl 2): 21–33.
Rao N, Schilling D, Rice J, Ridenour M, Mook W, Santa E. Prevention of postoperative mediastinitis: a clinical process improvement model. J Healthcare Qual. 2004; 26: 22–27.
Shinefield H, Black S, Fattom A, Horwith G, Rasgon S, Ordonez J, Yeoh H, Law D, Robbins JB, Schneerson R, Muenz L, Fuller S, Johnson J, Fireman B, Alcorn H, Naso R. Use of a Staphylococcus aureus conjugate vaccine in patients receiving hemodialysis. N Engl J Med. 2002; 346: 491–496.