Ann Thorac Surg 2009;88:S23-S42. doi:10.1016/j.athoracsur.2009.05.056
© 2009 The Society of Thoracic Surgeons
Report of STS Quality Measurement Task Force
The Society of Thoracic Surgeons 2008 Cardiac Surgery Risk Models: Part 2—Isolated Valve Surgery
Sean M. O'Brien, PhDa,
David M. Shahian, MDb,*,
Giovanni Filardo, PhD, MPHc,
Victor A. Ferraris, MDd,
Constance K. Haan, MDe,
Jeffrey B. Rich, MDf,
Sharon-Lise T. Normand, PhDg,
Elizabeth R. DeLong, PhDa,
Cynthia M. Shewan, PhDh,
Rachel S. Dokholyan, MPHa,
Eric D. Peterson, MD, MPHa,
Fred H. Edwards, MDe,
Richard P. Anderson, MDi,
a Duke Clinical Research Institute, Durham, North Carolina
b Massachusetts General Hospital, Boston, Massachusetts
c Institute for Health Care Research and Improvement, Baylor Health Care System, Dallas, Texas
d University of Kentucky Chandler Medical Center, Division of Cardiovascular and Thoracic Surgery, Lexington, Kentucky
e University of Florida, Division of Cardiothoracic Surgery, Jacksonville, Florida
f Sentara Cardiovascular Research Institute, Norfolk, Virginia
g Department of Health Care Policy, Harvard Medical School, and Department of Biostatistics, Harvard School of Public Health, Boston, Massachusetts
h The Society of Thoracic Surgeons, Chicago, Illinois
i Seattle, Washington
Abbreviations and Acronyms AVR = aortic valve replacement; CABG = coronary artery bypass graft surgery; CI = confidence interval; MI = myocardial infarction; MVR = mitral valve replacement; MVRepair = mitral valve repair; NCD = National Adult Cardiac Surgery Database; QMTF = Quality Measurement Task Force; STS = The Society for Thoracic Surgeons
* Address correspondence to Dr Shahian, Massachusetts General Hospital, 55 Fruit St, Boston, MA 02114 (Email: dshahian{at}partners.org).
| Drs O'Brien, Shahian, Filardo, Ferraris, Haan, Rich, Normand, DeLong, Shewan, Peterson, Edwards, Anderson, and Ms Dokholyan, have no conflicts of interest to declare regarding this work.
|
 |
Abstract
|
|---|
Background: Adjustment for case-mix is essential when using observational data to compare surgical techniques or providers. That is most often accomplished through the use of risk models that account for preoperative patient factors that may impact outcomes. The Society of Thoracic Surgeons (STS) uses such risk models to create risk-adjusted performance reports for participants in the STS National Adult Cardiac Surgery Database (NCD). Although risk models were initially developed for coronary artery bypass surgery, similar models have now been developed for use with heart valve surgery, particularly as the proportion of such procedures has increased. The last published STS model for isolated valve surgery was based on data from 1994 to 1997 and did not include patients undergoing mitral valve repair. STS has developed new valve surgery models using contemporary data that include both valve repair as well as replacement. Expanding upon existing valve models, the new STS models include several nonfatal complications in addition to mortality.
Methods: Using STS data from 2002 to 2006, isolated valve surgery risk models were developed for operative mortality, permanent stroke, renal failure, prolonged ventilation (> 24 hours), deep sternal wound infection, reoperation for any reason, a major morbidity or mortality composite endpoint, prolonged postoperative length of stay, and short postoperative length of stay. The study population consisted of adult patients who underwent one of three types of valve surgery: isolated aortic valve replacement (n = 67,292), isolated mitral valve replacement (n = 21,229), or isolated mitral valve repair (n = 21,238). The population was divided into a 60% development sample and a 40% validation sample. After an initial empirical investigation, the three surgery groups were combined into a single logistic regression model with numerous interactions to allow the covariate effects to differ across these groups. Variables were selected based on a combination of automated stepwise selection and expert panel review.
Results: Unadjusted operative mortality (in-hospital regardless of timing, and 30-day regardless of venue) for all isolated valve procedures was 3.4%, and unadjusted in-hospital morbidity rates ranged from 0.3% for deep sternal wound infection to 11.8% for prolonged ventilation. The number of predictors in each model ranged from 10 covariates in the sternal infection model to 24 covariates in the composite mortality plus morbidity model. Discrimination as measured by the c-index ranged from 0.639 for reoperation to 0.799 for mortality. When patients in the validation sample were grouped into 10 categories based on deciles of predicted risk, the average absolute difference between observed versus predicted events within these groups ranged from 0.06% for deep sternal wound infection to 1.06% for prolonged postoperative stay.
Conclusions: The new STS risk models for valve surgery include mitral valve repair as well as multiple endpoints other than mortality. Model coefficients are provided and an online risk calculator is publicly available from The Society of Thoracic Surgeons website.
Models for predicting surgical outcomes on the basis of patient preoperative characteristics are valuable tools for research, quality improvement, and clinical practice. Such models are used by The Society of Thoracic Surgeons (STS) to produce risk-adjusted performance reports for providers participating in the STS National Adult Cardiac Surgery Database (NCD). They are also used by STS surgeons and other physicians for counseling patients about the risk of surgery.
The earliest STS risk models were developed nearly 2 decades ago for isolated coronary artery bypass graft surgery (CABG). Subsequently, similar models have been developed for isolated valve replacement and combined CABG plus valve replacement. Because surgical practice and outcomes are changing rapidly, these models are updated periodically to reflect contemporary experience.
The last published STS model for isolated valve surgery was based on STS data from 1994 to 1997. The reference population included aortic and mitral valve replacements but excluded mitral valve repair, and the endpoint was operative mortality. In the decade since this model was published, many aspects of heart surgery have changed. First, as CABG volumes have decreased with the introduction of coronary stents, valve surgery as a proportion of overall heart surgery volume has increased in most practices. Between 2000 and 2006, the percentage of isolated CABG procedures decreased from 73% to 60% and the percentage of isolated valve procedures increased from 18% to 22%. Thus, in assessing provider performance, it is no longer sufficient only to consider isolated CABG surgery. Second, the frequency of mitral repair as a percentage of all isolated mitral operations in the STS NCD increased from 35% in 2000 to 53% in 2006. Third, during the same time period, the average mortality rate for isolated aortic or mitral surgery also decreased. Finally, efforts to measure and compare surgical performance have intensified and expanded. In addition to measuring operative mortality, performance reports increasingly focus on nonfatal complications as well as resource utilization and efficiency. Such outcomes have not historically been risk-adjusted for valve surgery.
The STS Quality Measurement Task Force (QMTF) has undertaken a complete revision of all STS risk models for adult cardiac surgery, and these new models were implemented in January 2008. This report, Part 2 of 3, describes the new STS models for isolated valve surgery (Part 1 describes the STS isolated CABG models, and Part 3 describes the models for CABG plus valve surgery). Authors of this report are the QMTF members who were involved in this initiative.
Two important features have been incorporated into these new models. First, the population includes mitral valve repair as well as aortic and mitral valve replacement. Second, in addition to operative mortality, the new models include six nonfatal in-hospital morbidity endpoints and two length-of-stay endpoints. In comparison with several other valve models that have recently been published [1–6], the STS models are distinguished by the large size of the development population and the broad spectrum of endpoints included.
|
Study Population and Endpoints
|
|---|
The population for this analysis consisted of operations on adult patients aged 20 to 100 years who underwent isolated single aortic or mitral valve surgery between January 1, 2002, and December 31, 2006. Only patients undergoing one of the following procedures were included: (1) isolated aortic valve replacement (AVR); (2) isolated mitral valve replacement (MVR); and (3) isolated mitral valve repair (MVRepair).
Because of the relatively small number of pulmonic, tricuspid, multiple valve procedures, and aortic repairs, these cases were not included in the current models. Patients undergoing concomitant CABG were excluded from the current analysis, but these were included in the separate STS valve plus CABG models described in Part 3 of this series. Records with missing data on sex (n = 44) were excluded because missing sex is not allowed in the analysis dataset used for creating STS database participant feedback reports. This left a final study population of 109,759 patient operations performed at 809 STS NCD participating groups. Patients on dialysis preoperatively (n = 2,699) were not included when developing the risk model for prediction of postoperative renal failure.
Patient characteristics in the study population are presented in Table 1.
Training and Validation Samples
The study population was randomly divided into a 60% training (development) sample and a 40% test (validation) sample. The development sample was used to identify predictor variables and estimate model coefficients. Data from the validation sample were used to assess model fit, discrimination, and calibration. After choosing variables and assessing model fit, the development and validation samples were subsequently combined, and the final model coefficients were estimated using the combined (development plus validation) data.
Endpoints
Risk models were developed for nine endpoints, identical to those in the STS CABG models. In contrast with the definition of operative mortality, which includes hospital deaths as well as deaths that occur after discharge within 30 days of surgery, the morbidity endpoints only include events that occurred before discharge. However, beginning with version 2.61, sternal infection data will be recorded for as long as 30 days postoperatively. The nine endpoints are as follows: (1) operative mortality: death during the same hospitalization as surgery, regardless of timing, or within 30 days of surgery regardless of venue; (2) permanent stroke (cerebrovascular accident [CVA]): a central neurologic deficit persisting longer than 72 hours; (3) renal failure: a new requirement for dialysis or an increase of the serum creatinine to greater than 2.0 mg/dL and double the most recent preoperative creatinine level; (4) prolonged ventilation (longer than 24 hours); (5) deep sternal wound infection; (6) reoperation for any reason; (7) major morbidity or mortality: a composite defined as the occurrence of any of the above endpoints; (8) prolonged postoperative length of stay (PLOS): length of stay (LOS) more than 14 days (alive or dead); and (9) short postoperative LOS (SLOS): LOS less than 6 days and patient alive at discharge.
Table 2
summarizes the endpoint frequencies in the study population.
|
Single Versus Multiple Models
|
|---|
Two issues required particularly careful consideration: whether to construct separate models for the AVR and MVR populations, and how best to further subdivide the mitral population into repair versus replacement.
Because of the large size of the STS NCD, separate models for AVR, MVR, and MVRepair initially seemed both feasible and appropriate. However, because the endpoints of interest are rare events, we recognized the possibility that the number of such events would be too small to support reliable estimation of the model coefficients.
To assess this tradeoff, we conducted a pilot study to compare two alternative strategies for developing risk models for isolated valve surgery. The first strategy involved developing models separately for three subpopulations (AVR, MVR, and MVRepair). The second strategy involved modeling all three subpopulations together in a single model; several interaction terms were included to allow the effect of selected risk factors to differ across the subpopulations. Both strategies were pilot tested by developing risk models for two endpoints: operative mortality and permanent stroke. These pilot models were developed in a 60% development sample and tested in a separate 40% validation sample. Each model was assessed by calculating the c-index and the generalized R2 index of Nagelkerke [7] in the validation sample for each combination of subpopulation and endpoint (3 subpopulations x 2 endpoints = 6 combinations). With the exception of AVR operative mortality, the combined model with interactions resulted in better discrimination. With the exception of MVR and MVrepair operative mortality, the combined model also captured more variation as measured by the generalized R2 statistic.
Because the combined model strategy performed better in the majority of cases, and because a single combined model was consistent with the previous STS valve model, the combined model strategy was selected. To avoid assuming that the weighting of each risk factor was exactly constant across the three populations, we included interactions between surgery type and several key predictor variables. In principle, fitting a single model with several interactions is advantageous because it allows for pooling information across related groups without making an a priori assumption that all of the covariate effects are exactly constant across groups.
|
Selection of Candidate Predictor Variables
|
|---|
Our general approach to variable selection is discussed in Part 1 of this series describing the development of the 2008 STS isolated CABG risk models. Briefly, we initially identified potential candidate variables by reviewing four versions of the STS data collection instrument (data versions 2.35, 2.41, 2.52.1, and 2.61) as well as previously published STS and similar cardiac risk models [1–6]. A panel of cardiac surgeons and health policy experts reviewed the initial variables for face validity and to be certain that no important predictor variables available in (or mappable to) to STS NCD data version 2.61 had been excluded.
Final candidate explanatory variables and their coding are summarized in Table 3. The variables were identical to the CABG model candidate variables with the following differences: (1) percutaneous coronary intervention conducted within 6 hours or less of surgery was not a candidate variable because it was present in only 122 patients (0.1%) in the valve model population; (2) infectious endocarditis was included. This risk factor was rarely present among isolated CABG patients (0.09%), but was not uncommon (7.7%) among patients undergoing valve surgery; (3) mitral stenosis was included; this risk factor was rarely present among isolated CABG patients (0.35%) but was common (7.4%) among patients undergoing valve surgery; and (4) an indicator for surgery type (AVR, MVR, MVRepair) was included in the valve models.
Coding of Explanatory Variables
The coding of continuous and categorical variables was identical to the CABG models, except for the following differences: (1) age was modeled as a linear spline truncated from below at 50 years and with a change of slope at 75; (2) creatinine was modeled as a linear term with values less than 0.5 and greater than 5.0 mapped to those values respectively (approximately the 1st and 99th percentiles of the empirical distribution); (3) previous myocardial infarction (MI) was modeled as three categories (< 24 hours, 1 to 21 days, and > 21 days or no MI); the first two categories were subsequently combined after expert panel review; (4) race was modeled as three categories: black, Hispanic, Caucasian/other; and (5) chronic lung disease was modeled as linear across four categories (none, mild, moderate, severe).
In general, these differences reflect a slightly simpler coding scheme (fewer parameters) for the valve models compared with the isolated CABG models.
Repair Versus Replacement
In addition to a number of variables whose inclusion or coding were noted to be problematic during development of the 2008 STS isolated CABG models (Part 1 of this series), the approach to modeling mitral valve repair versus replacement was of some concern in the valve models. From a methodologic perspective, models used for risk-adjustment should include all patient preoperative risk factors that vary in prevalence between institutions and that substantially impact the probability of an adverse outcome. Such models should include variables that reflect the patient's baseline condition but should not include intraoperative events (eg, unexpected hemorrhage) or discretionary care processes (eg, use of a mechanical versus bioprosthetic valve). Adjusting for intraoperative events is not appropriate because these may be a reflection of the surgeon's performance. Adjusting for discretionary care processes may likewise mask differences in performance if the surgeon's choice of procedures has a substantial impact on outcomes. The same patient may receive valve repair if treated by one surgeon and replacement if treated by another. Adjusting for repair versus replacement will potentially conceal the outcomes of surgeons who achieve excellent results by repairing technically challenging valves that might otherwise be replaced if treated by a surgeon with less skill or tenacity. Importantly, there is considerable evidence to suggest the superiority of valve repair whenever feasible.
However, in addition to such discretionary factors, the decision to repair rather than replace the mitral valve is also dependent upon the patient's preoperative valve disease etiology, anatomy, and pathophysiology. On average, patients amenable to valve repair have less extensive valve pathology and a relatively favorable postoperative prognosis (the mortality rate for valve repair is 1.6% compared with 5.7% for replacement). Ignoring these anatomical differences can introduce bias when comparing institutions, especially because these variables are not captured elsewhere on the STS data collection form.
A related difficulty in adjusting for repair versus replacement is that the former approach may sometimes be abandoned intraoperatively by the surgeon and converted to MVR. That may sometimes occur because of unforeseen technical problems that would prevent most surgeons from completing the repair, but in other instances, a more skilled surgeon might persist and achieve successful valve repair. Effectively separating these two scenarios is problematic from available data.
Ultimately, it was elected to include an indicator for mitral valve repair versus replacement in the valve risk models, consistent with the approach in a number of existing valve surgery models. We acknowledge that available data make it impossible to determine whether patient differences or surgical skill and judgment are the most important factors in determining between-provider variation in the proportion of valves repaired.
Recognizing the potential limitations of this modeling approach, the decision to adjust for repair versus replacement may be reassessed in future versions of the STS risk models. Beginning with data in version 2.61, the database will capture whether or not repair was attempted, and repair versus replacement may be analyzed based on an intention-to-treat principle.
|
Missing Data
|
|---|
Model variables with more than 1% missing data in the study sample were ejection fraction (8.9%), NYHA class (4.9%), tricuspid insufficiency (2.7%), aortic insufficiency (2.0%), mitral insufficiency (1.6%), left main disease (1.5%), creatinine/dialysis (1.6%), and number of diseased vessels (1.4%). The method of imputing missing data was identical to that employed in the isolated CABG models and described in Part 1 of this series. Briefly, binary risk factors were modeled as yes versus no or missing (ie, missing values were analyzed as if the endpoint did not occur). Missing data on categorical variables were imputed to the lowest risk value, typically the mode, and outcomes were typically similar for missing data and lowest risk patients. Missing data on continuous variables were imputed by grouping patients into strata and assigning the stratum-specific median value. For example, ejection fraction was imputed by grouping on sex and congestive heart failure and calculating the median ejection fraction among patients with nonmissing ejection fraction in each group.
Although multiple imputation is generally preferable to single imputation [8], single imputation was chosen for this analysis mainly because of practical considerations. Furthermore, because of the small fraction of missing data, the impact of single versus multiple imputation was considered to be inconsequential. Subsequent sensitivity analyses confirmed that the choice between single versus multiple imputation had little impact on the final regression coefficients, risk estimates, and confidence intervals. A summary of these sensitivity analyses, including coefficients and covariance matrices, is available at www.sts.org/riskmodels.
|
Final Variable Selection Procedure
|
|---|
Variables were initially selected using an automated stepwise model selection algorithm. The stepwise procedure began with a model that included all of the final candidate variables except for interaction terms. Age, sex, body surface area, and month of surgery were forced into each model. Other variables were selected in a stepwise fashion using a significance criterion of 0.05 for entry and removal. This criterion was less stringent than that employed in development of the CABG models, because the sample size in the former was so much larger than that which was used for the valve models. The stepwise procedure was performed separately for each endpoint. The results were then reviewed by an expert panel of surgeons, and the following changes were made based on their feedback: (1) "MI less than 24 hours" and "MI 1 to 21 days" were collapsed into a single category; (2) preoperative atrial fibrillation was forced into the model for stroke (CVA); and (3) an indicator variable for dialysis was forced into any model that included creatinine level.
Interaction Terms
In addition to including main effects, we tested the interaction between surgery group (AVR, MVR, MVRepair) and each of the following variables: age, diabetes mellitus, dialysis, creatinine, reoperation, endocarditis, emergent status, chronic lung disease, congestive heart failure, ejection fraction, sex, shock, intra-aortic balloon pump/inotropes, mitral insufficiency, aortic insufficiency, mitral stenosis, and aortic stenosis. These interaction terms allowed the effect of these selected risk factors to differ across the surgery populations.
Four additional sets of interactions were also included in the models: (1) sex by body surface area (BSA); (2) sex by BSA2; (3) age by reoperation; and (4) age by emergent status. These interaction terms were preselected and were not tested as part of the backward selection algorithm. Additional technical details are provided in the Appendix. For reasons described in Part 1 of this series (isolated CABG risk models), an extensive automated search for additional interaction terms was not conducted.
Adjustment for Time Trends
Surgery date was included in each model to adjust for changes in the frequency of adverse outcomes over the 5-year study period. Although surgery date is not itself a variable of interest, we adjusted for it to reduce potential confounding by time trends when estimating regression coefficients for the variables that are of primary interest (ie, patient preoperative risk factors). An example is provided in Part 1 of this series.
Surgery date was categorized into 6-month intervals (corresponding to the biannual STS data harvests) and modeled as a linear trend across the ordinal categories. Because it is a nuisance variable, surgery date is not included in the final risk prediction algorithm. Thus, a patient's predicted risk does not depend on the patient's surgery date. As described in the Appendix, the published intercept parameter has been adjusted to incorporate the time trend. The adjusted intercept reflects the baseline risk for a reference period of July to December 2006.
|
Results
|
|---|
Assessment of Model Fit and Discrimination
Because of the relatively large size of our sample, the Hosmer-Lemeshow test is uninformative and would invariably result in a significant p value [9]. As an alternative, model fit was assessed graphically by plotting observed versus predicted rates of each endpoint across deciles of predicted risk in the development and validation samples. This was done in the overall population and in subgroups based on surgery type (AVR, MVR, MVRepair); age (< 60, 60 to 79, 
80 years); sex (male, female); diabetes mellitus (yes/no); status (elective, nonelective); and ejection fraction (
40, > 40). Calibration plots (observed versus expected) based on the overall validation sample are presented in Figure 1. The average absolute difference between observed versus predicted event rates within deciles of predicted risk ranged from 0.06% for deep sternal wound infection to 1.06% for prolonged postoperative stay. Analogous figures were produced for specific valve procedures and numerous subgroups, and these are available at www.sts.org/riskmodels.
Model fit appeared to be adequate for each endpoint with the possible exception of deep sternal wound infection, which revealed some overfitting within certain subgroups. A modest degree of overfitting was expected for this endpoint given the relatively small number of infections and large number of candidate predictors.
Discrimination was assessed by the c-statistic, also known as the area under the receiver operating characteristic (ROC) curve. Table 4
presents the discrimination of each model in the development and validation samples for all patients combined and for subgroups consisting of AVR, MVR, and MVRepair. In the validation sample, c-statistics for the operative mortality model were 0.799 (overall), 0.759 (AVR), 0.802 (MVR), and 0.844 (MVRepair). C-statistics in the validation sample for other endpoints ranged from 0.619 for reoperation in the AVR subgroup to 0.800 for prolonged length of stay in the MVRepair subgroup.
Final Models
After validating the models in the 40% validation sample, the development and validation samples were then combined, and the final model coefficients were estimated using the overall 100% combined sample. The final logistic regressions were estimated using generalized estimating equations with empirical (sandwich) standard error estimates to account for clustering of patients within institutions [10]. An independence working correlation matrix was used to apply the generalized estimating equations methodology. With this approach, the estimated regression coefficients were identical to those obtained using ordinary logistic regression, but the standard errors were adjusted to account for the clustered data structure.
Odds Ratios
Odds ratios and 95% confidence intervals (CI) for the final selected models are presented in Table 5. "Not applicable" indicates that the specific predictor was not included in a particular risk model. Because several variables interact with surgery type, the odds ratios for these variables differ depending on the type of surgery (AVR, MVR, MVRepair). For example, in the operative mortality model, the odds ratio for emergent status is 3.77 (95% CI: 2.75, 5.16) for AVR, 2.74 (95% CI: 1.99, 3.78) for MVR, and 8.73 (95% CI: 4.84, 15.74) for MVRepair. Odds ratios that do not interact with surgery type are summarized in Table 5, Part A. Odds ratios that differ by surgery type for at least one endpoint are presented in Table 5, Parts B, C, and D.
Final Model Intercept and Coefficients
The final risk prediction algorithms, including all coefficients and intercepts, are presented in the Appendix.
|
Limitations
|
|---|
The limitations for these valve models are similar to those for the CABG models and are thoroughly discussed in Part 1 of this series (2008 STS CABG risk models).
|
Conclusion
|
|---|
The STS Quality Measurement Task Force has developed and tested nine new risk-adjustment models for isolated valve surgery using the STS NCD. This report includes a detailed exposition of the model development process, including not only statistical issues but also the many clinical and pragmatic judgments that were required. An online risk calculator is also available through a link from the STS website.
|
Appendix
|
|---|
Regression Coefficients and Variable Definitions for STS 2008 Valve Models
For each endpoint, the formula for calculating a patient's predicted risk of the endpoint has the form:
where x
1, x
2, ... , xn denote patient preoperative risk factors (eg, quantitative variables such as age, and comorbidities coded as 1 = present, 0 = absent); and β0, β1, ... , βn denote regression coefficients (numerical constants). Regression coefficients for each endpoint are presented in Appendix Table 1. The variables x
1, x
2, ... , xn are the same for each endpoint and are defined in Appendix Table 2. The regression coefficient for the time trend is not presented. Instead, the intercept has been adjusted to incorporate the time trend. This adjusted intercept reflects the baseline risk for a reference period of July to December 2006.
|
Footnotes
|
|---|
This author is deceased. Former Chair, Quality, Research and Patient Safety Council, The Society of Thoracic Surgeons, Chicago, IL. 
|
References
|
|---|
- Ambler G, Omar RZ, Royston P, Kinsman R, Keogh BE, Taylor KM. Generic, simple risk stratification model for heart valve surgery Circulation 2005;112:224-231.[Abstract/Free Full Text]
- Edwards FH, Peterson ED, Coombs LP, et al. Prediction of operative mortality after valve replacement surgery J Am Coll Cardiol 2001;37:885-892.[Abstract/Free Full Text]
- Hannan EL, Wu C, Bennett EV, et al. Risk index for predicting in-hospital mortality for cardiac valve surgery Ann Thorac Surg 2007;83:921-929.[Abstract/Free Full Text]
- Jin R, Grunkemeier GL, Starr A. Validation and refinement of mortality risk models for heart valve surgery Ann Thorac Surg 2005;80:471-479.[Abstract/Free Full Text]
- Kuduvalli M, Grayson AD, Au J, Grotte G, Bridgewater B, Fabri BM. A multi-centre additive and logistic risk model for in-hospital mortality following aortic valve replacement Eur J Cardiothorac Surg 2007;31:607-613.[Abstract/Free Full Text]
- Nowicki ER, Birkmeyer NJ, Weintraub RW, et al. Multivariable prediction of in-hospital mortality associated with aortic and mitral valve surgery in Northern New England Ann Thorac Surg 2004;77:1966-1977.[Abstract/Free Full Text]
- Nagelkerke NJD. A note on a general definition of the coefficient of determination Biometrika 1991;78:691-692.[Abstract/Free Full Text]
- Little RJA, Rubin DB. Statistical analysis with missing data2nd ed.. Hoboken, NJ: Wiley-Interscience; 2002.
- Marcin JP, Romano PS. Size matters to a model's fit Crit Care Med 2007;35:2212-2213.[Medline]
- Liang KY, Zeger SL. Longitudinal data-analysis using generalized linear-models Biometrika 1986;73:13-22.[Abstract/Free Full Text]
Related Article
-
The Society of Thoracic Surgeons 2008 Cardiac Surgery Risk Models: Introduction
- David M. Shahian and Fred H. Edwards
Ann. Thorac. Surg. 2009 88: S1.
[Extract]
[Full Text]
[PDF]
Top
Abstract
Introduction
