| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
a Division of Thoracic Surgery, Massachusetts General Hospital, Boston, Massachusetts
b Duke Clinical Research Institute, Duke University, Durham, North Carolina
c Division of General Thoracic Surgery, Mayo Clinic School of Medicine, Minneapolis, Minnesota
Accepted for publication March 10, 2008.
* Address correspondence to Dr Wright, Division of Thoracic Surgery, Blake 1570, Massachusetts General Hospital, Boston MA 02114 (Email: cdwright{at}partners.org).
Presented at the Forty-fourth Annual Meeting of The Society of Thoracic Surgeons, Fort Lauderdale, FL, Jan 28–30, 2008.
 |
Abstract |
|---|
 Top Abstract IntroductionPatients and MethodsResultsCommentDiscussionReferences |
|---|
Methods: The Society of Thoracic Surgeons (STS) General Thoracic Surgery Database was queried for patients with lobectomy for lung cancer. A model of preoperative risk factors was developed by multivariate stepwise logistic regression setting the threshold for PLOS at 14 days. Morbidity was measured as postoperative events as defined in the STS database. Risk-adjusted results were reported to participating sites.
Results: From January 2002 to June 2006, 4979 lobectomies were performed for lung cancer at 56 STS sites, and 351 (7%) had a PLOS. They had more postoperative events than patients without PLOS (3.4 vs 1.2; p < 0.0001). Patients with PLOS also had higher mortality than those with normal LOS, at 10.8% (38 of 351) vs 0.7% (33 of 4628; p < 0.0001). Significant predictors of PLOS included age per 10 years (odds ratio [OR], 1.30, p < 0.001), Zubrod score (OR, 1.51; p < 0.001), male sex (OR, 1.45; p = 0.002), American Society of Anesthesiology score (OR, 1.54; p < 0.001), insulin-dependent diabetes (OR. 1.71; p = 0.037), renal dysfunction (OR, 1.79; p = 0.004), induction therapy (OR, 1.65; p = 0.001), percentage predicted forced expiratory volume in 1 second in 10% increments (OR, 0.88; p < 0.001), and smoking (OR, 1.33; p = 0.095). After risk adjustment, twofold interhospital variability existed in PLOS among STS sites
Conclusions: We identified significant predictors of PLOS, a surrogate morbidity marker after lobectomy for lung cancer. This model may be used to provide meaningful risk-adjusted outcome comparisons to STS sites for quality improvement purposes.
|
Introduction |
|---|
|
TopAbstractIntroductionPatients and MethodsResultsCommentDiscussionReferences |
|---|
For pulmonary resection, such work to date has been limited to single-institution retrospective studies that have identified risk factors for morbidity and mortality after lung resection [1] and a few multicenter risk models that represent limited patient populations or procedures, or both [2, 3]. As such, we sought to use the Society of Thoracic Surgeons (STS) General Thoracic Surgery Database (GTSD) to identify risk factors for adverse outcomes after surgical lobectomy. We also sought to develop a risk model that could provide STS sites with risk-adjusted outcomes comparisons for quality improvement purposes.
|
Patients and Methods |
|---|
|
TopAbstractIntroductionPatients and MethodsResultsCommentDiscussionReferences |
|---|
Patient Population
Between January 2002 and June 2006, 4979 lobectomies for primary lung cancer were entered in the STS GTSD. Excluded from analysis were patients who were missing data for age, gender, surgery date, or discharge date. Also excluded were patients who had a surgery date before January 2002, an emergency operation such as for massive hemoptysis and lung cancer, or length of stay exceeding 1 year. The length of stay, which was the outcome variable, was indeterminate in more than 10% of the records from 2 of the 58 participants. All of the records from these 2 participants were excluded from further consideration before imputation methods were applied.
Outcome Definitions
Postoperative events were those defined by the STS GTSD guidelines. Hospital mortality was defined as death during the same hospitalization or within 30 days of the procedure.
The previously reported operative mortality rate of 1.6% after lobectomy proved too low to establish a risk model, and individual complications were relatively uncommon [6]. We decided therefore to model prolonged length of stay (PLOS) as a surrogate marker for important postoperative complications. The threshold of extended hospitalization ensured that postoperative events had a deleterious impact on recovery from lung resection. We selected a hospitalization beyond 14 days as clinically relevant and statistically valid, and only 351 of 4979 patients (7%) passed this mark in a procedure, with a median stay of 6 days and a 25th percentile of 8 days.
Selected STS GTSD variables were considered for inclusion in the modeling process. Clinical staging information and certain other variables were excluded on the basis of a combination of clinical judgment and data analysis, particularly with regard to rate of missing information.
The final PLOS model included the covariates of age, gender, Zubrod score (0 to 4), American Society of Anesthesiology (ASA) class (1 to 5), insulin-dependent diabetes (yes/no), renal dysfunction (none/serum creatinine level >2 mg/dL/dialysis), preoperative therapy (chemotherapy or thoracic radiation/none), forced expiratory volume in 1 second (FEV1; % predicted), smoking (ever/never), year of operation (0 for first year in data set, 1, 2, etc for subsequent years). Three additional covariates excluded due to lack of association with the outcome in initial logistic modeling were operative status (elective/urgent), prior cardiothoracic operation, and cardiovascular disease, defined as having one or more of coronary artery disease, peripheral vascular disease, irreversible cerebrovascular disease, or congestive heart failure.
Data were adjusted to account for the limited maturity of the data set and to maximize use of available data points. When records with missing values of the model covariates (with exception of age, gender, year of operation, and %FEV1) were encountered, these values were imputed to the most common value of the covariate among the remaining eligible cases. Of all yes/no variables, only smoking status was imputed to yes because more than 80% of this lung cancer population had a history of smoking.
For the continuous variable %FEV1, several imputation schemes were considered because the rate of missing data for this variable was high (22%) compared with the other adjustor variables. Linear regression was attempted to predict %FEV1 according to the values of other spirometric variables but was discarded because correlation between %FEV1 and all other variables was low and the resulting regression fit was poor. Missing values for %FEV1 were finally imputed to the median value, based on adjustment to the remaining study population.
A sensitivity analysis using multiple imputations to simultaneously impute missing values for all variables provided results nearly identical to the simple imputation using the most common value or median; for example, the mean for %FEV1 using multiple imputation was 80, compared with a value of 81 using median imputation. The model was also run without the patients who were missing %FEV1 data, and the results of the model did not change.
After imputation of missing values, the process of modeling the PLOS began. On the basis of our small sample of data and the potential clustering of similar patients within hospitals (STS participants are typically single hospitals), a Bayesian hierarchic (random effects) logistic regression model was selected. The severity of illness in patients at different hospitals may vary, and some participants may treat a greater number of patients with comorbidities that favor a prolonged hospital stay. A hierarchic approach allows adjustment for both patient-level risk factors and participant-level case-mix. The Bayesian approach to hierarchic modeling takes into account the greater susceptibility to chance variation in the event rate associated with smaller sample sizes. Accounting for small sample sizes is crucial in this analysis because of the limited number of lobectomy procedures in the STS GTSD to date.
Statistical Model
Our model adjusted for the 10 patient factors listed above and included a separate random effect parameter for each of the 56 participants in the analysis.
|
| (1) |
Parameters of the random effects logistic model were estimated in a Bayesian framework using WinBUGS software (MRC Biostatistics Unit, Cambridge, United Kingdom, and Imperial College of Science, Technology and Medicine at St. Mary's, London). The Bayesian approach involves producing a posterior probability distribution for each parameter of interest; so, unlike conventional statistical methods, the results of Bayesian analyses are expressed in terms of probabilities. Thus, odds ratio estimates were computed by taking the mean of the antilogarithm of the posterior probabilities. Also, variation in these estimates is not described by typical 95% confidence intervals, but rather by 95% credible intervals with the following interpretation: Given the observed data, it is 95% likely that the true odds ratio lies in the indicated interval.
Bayesian methods allow computation of parameter estimates and 95% credible intervals; however, the ubiquitous p value that accompanies non-Bayesian 95% confidence intervals is not well defined. Instead, to assess significance, we define a Bayesian probability, defined as the probability that the true association between predictor and outcome is on the opposite side of the null hypothesis value from the estimated value (Table 1).
|
|
| (2) |
An SIR value of more than 1.0 implies that the participant's rate of PLOS is higher than the rate that would be projected for an average participant that treated the same mix of patients. Conversely, an SIR value of less than 1.0 implies that the participant's rate of PLOS is lower than would be projected for an average participant treating the same mix of patients.
For both of these quantities, accompanying 95% credible intervals were determined to account for uncertainty in the estimation of RAR and SIR. The 95% credible interval indicates the range of RAR and SIR values that are plausible in light of the observed data. If the 95% credible interval for a participant's SIR includes the null value 1.0, then we cannot reliably distinguish this participant's performance from the STS average: Either the participant's performance was close to average or else the participant's sample size was too small to make a reliable determination.
Analysis was performed using S-Plus 6 (Insightful Corp, Seattle, WA), SAS 9.1 (SAS Institute, Cary, NC), and WinBUGS 1.4.1 software.
|
Results |
|---|
|
TopAbstractIntroductionPatients and MethodsResultsCommentDiscussionReferences |
|---|
|
Postoperative events occurred in 1745 patients (35%), with a median of 1 and a mean of 1.6 (SD, 1.5). The hospital discharge mortality rate was 1.4% (71 of 4979). The median LOS was 6 days, and the mean LOS was 7.4 (SD, 7.4).
The LOS exceeded 14 days (PLOS) in 351 patients (7%). The mean LOS was 25.7 days for patients with PLOS and 6 days for those without PLOS (p < 0.0001; Fig 1 and Fig 2). Patients with PLOS had more postoperative events than those without PLOS (3.4 vs 1.2 events, p < 0.0001). Patients with PLOS had a higher discharge mortality rate than those without PLOS: 10.8% (38 of 351) vs 0.7% (33 of 4595; p < 0.0001). The characteristics of patients with and without PLOS are summarized in Table 3. PLOS was associated with older age and a greater number of comorbidities.
|
|
|
|
|
|
|
|
Comment |
|---|
|
TopAbstractIntroductionPatients and MethodsResultsCommentDiscussionReferences |
|---|
Other factors with impact on recovery from surgery were therefore considered. Operative complications, captured in the STS database as postoperative events, may impair quality of life, delay return to regular activities, increase cost, and raise the risk of death. However, individual adverse events do not occur with sufficient frequency to allow a separate analysis. We chose to model prolonged hospital LOS as a marker for those adverse events that either individually or collectively cause a delay of discharge after an operation. We acknowledge the selection of 14 days is somewhat arbitrary and that a case can be made for looking at perhaps a set of important complications and then modeling such a composite measure. Another reason for choosing 14 days was that it represents a true statistical outlier (ie, 95% chance the difference is not due to chance) with no quibbling about the severity of the adverse event(s). Another model could be constructed with a lesser LOS but with less confidence that the PLOS was different than that of the non-PLOS patients.
In contrast with our model based on LOS, risk prediction established for a Veterans Administration (VA) patient population measured the incidence of morbid events and the likelihood of their occurrence in the presence of certain risk factors [2]. The advantage of a time-related model for nonlethal events lies in the weight attributed to the actual delay of recovery, relevant to added cost and reduced quality of life, and to the collective impact of multiple complications on individual patients. Further, the identified risk factors, age, Zubrod score, male gender, ASA score, diabetes, renal dysfunction, induction therapy, and %FEV1, have relevance outside of this patient population. The choice of this model was born of necessity due to the small number of deaths and individual complications, and certain drawbacks are recognized. The model excludes early deaths (albeit rare) from analysis.
The STS GTDB data set for lobectomy for lung cancer is remarkable for the equal proportions of men and women, notably different than in previous reports from Europe and the VA system [1–3]. Presumably this reflects the smoking prevalence and lung cancer mortality rates in women that have been documented to rise to the levels seen in men in the United States [7]. The patients in the STS GTDB have many serious medical comorbidities and most were ASA score 3. Almost all were current or previous smokers. Despite their smoking, their pulmonary function was not that impaired, with a median FEV1 of 81% (interquartile range, 71% to 90%). Only 11% had induction therapy for locally advanced lung cancer.
Patients who had a PLOS had a mean LOS more than four times that of those who had a LOS of less than 14 days (26 vs 6, p < 0.0001; Fig 1 and Fig 2). This is quite remarkable and has obvious implications for the patient and for health care resource consumption. Patients with a normal LOS had only 0.4 complications per patient, whereas those with a PLOS had 3.2 complications per patient (p < 0.0001).
As expected, most of the complications were pulmonary in origin, including prolonged air leak, major atelectasis, pneumonia, adult respiratory distress syndrome, pulmonary embolism, prolonged ventilation, and need for tracheostomy. Atrial arrhythmias were about three times more common in those with a PLOS than those with a normal LOS. Myocardial infarction was rather rare, with an incidence of only 3%, suggesting both adequate preoperative evaluation and postoperative management of patients with coronary disease, which was present in 20%. Thromboembolism was also relatively uncommon, with only 0.38% pulmonary embolism and 0.42% deep venous thrombosis in the overall cohort, suggesting adequate prophylaxis of these high-risk patients. Reoperation for bleeding was also quite rare, with an incidence of only 0.66% attesting to the quality of surgery. Sepsis was rare overall (0.8%) but was, not surprisingly, present in 7% of patients with a PLOS. Deterioration in renal function occurred in 9% of patients with a PLOS, indicating the severity of the other complications sustained by these patients. A VATS approach was seemingly associated with fewer complications than a thoracotomy approach for lobectomy, but these patients were not matched in any way and thus there probably was some element of selection bias present that prohibits drawing firm conclusions.
The multivariate model largely confirmed what we, and others, have thought to be important risk factors for morbidity after lung resection. Age is an intuitive risk factor leading to the careful consideration of physiologic age when offering a resection to an older patient. Performance status is more commonly precisely assessed by our oncology colleagues but is also assessed, albeit usually not in a quantitative sense, by surgeons in their preoperative decision making. On the basis of this report, we should consider recording performance status in a formal, quantitative manner to both risk-adjust and to remind ourselves just how affected the patients are by their comorbidities.
The finding that male gender is a risk factor is puzzling and will bear more careful scrutiny of the database. Others have also reported male gender as a risk factor [1, 3]. Possible reasons include less pain tolerance than women and more comorbidities that were not captured in our database.
The ASA score, a validated marker of medical comorbidities for patients undergoing anesthesia, is also a good index of case complexity for lung resection and is readily assessed and available in the perioperative medical record. Insulin-dependent diabetes mellitus is associated with cardiovascular disease as well as with an increased risk of postoperative infection. It makes sense that diabetes would be a risk factor for morbidity in a relatively old cohort of patients undergoing a major operative procedure. Tight control of postoperative glucose levels with the aid of an endocrinologist might mitigate some of this risk.
Renal dysfunction, also a marker for significant cardiovascular disease was, not surprisingly, a risk factor for morbidity. In addition to the cardiovascular risk, the perioperative fluid shifts that normally occur would be expected to be more problematic in those with impaired excretory function, leading to more pulmonary complications.
Previous reports from smaller data sets have given conflicting information about whether induction therapy promotes postoperative morbidity [8, 9]. With the large STS data set it appears that indeed it is a risk factor.
Impaired pulmonary function has been the most studied risk factor for lung resection because it is indeed logical to assume that it is the dominant factor to consider when a surgeon weighs the risk of a lung resection [1–3]. Unfortunately, we were missing data on the %FEV1 in 22% of patients, which hampered our modeling. We were unsuccessful in developing a predictive model to estimate FEV1 from our data set and finally imputed it to the median value. We ran our model with and without the imputed values with no significant change in our entire model, indicating a satisfactory resolution of this missing data issue. As we expected, %FEV1 was predictive of postoperative morbidity. Preoperative pulmonary preparation of the high-risk patient with chronic obstructive pulmonary disease is an obvious strategy to mitigate this risk [10].
We would have liked to investigate the effect of diffusion capacity of the lung for carbon monoxide on morbidity because several reports indicate it is a reliable marker of poor lung reserve that leads to greater perioperative risk in single institution reports [11]. We were missing 37% of this data and hence decided to not include it in our model even though by univariate analysis it was highly predictive of elevated risk (p < 0.0001). It is unclear if many sites did not measure it at all or if it was mistakenly omitted. We plan to encourage the sites to obtain and record this important information to make our next model more complete.
There is a twofold difference between the best and worst sites, suggesting clinically significant interhospital variability in performance. However, a review of Figures 3 and 4 indicates that although some sites seem to perform better than others, the 95% credible intervals overlap. It is likely as the data set matures, these differences will become statistically significant. The importance of risk adjustment is seen in Figure 5. Several sites have unadjusted rates of PLOS that are twice the adjusted rates. Alternatively, there are sites with risk-adjusted rates of PLOS that are more than the unadjusted rates. It appears that participating STS surgeons are relatively uniform in their care for lung cancer patients treated by lobectomy, with relatively little variation in morbidity outcome. It is likely that one reason for the relative clustering of outcomes is that the current participants in the database are dedicated to quality improvement and are a self-selected group. These results are clearly better than what other national databases have reported and are not likely reflective of what is happening in the entire United States [6]. As we continue to grow the database, it is probable the results will more closely approximate the entire patient population.
Our analysis has several limitations. Many submissions had an incomplete data set, and the missing data mars the analysis. The GTSD has not yet been audited by an external review, which raises concerns about the quality of the data. It is, however, audited electronically by the Duke Clinical Research Center for missing and nonsensical data, which is in turn relayed back to the submitting site. The Adult Cardiac Database has recently started an external audit; thus far, the results thus far are rewarding, with very few errors. It is unlikely that our database has major intentional errors because there is currently no reward for such behavior.
Another potential confounding factor is the degree of conformity of entering real patients into the database among the various data managers. This cohort is not fully representative of the entire country and is likely to represent the "best" of thoracic surgeons, because participation in the database is an altruistic act that demonstrates commitment to quality improvement.
In conclusion, we identified a surrogate marker for significant complications after lobectomy for lung cancer, PLOS (exceeding 14 days). Patients with a PLOS had a markedly increased mortality rate of 10.8% vs 0.7% (p < 0.0001). Risk factors for a PLOS were age, Zubrod score, male gender, ASA score, insulin-dependent diabetes mellitus, renal dysfunction, induction therapy, and the percentage predicted FEV1. This risk model will allow STS surgeons to risk-adjust their outcomes for lobectomy for lung cancer to foster quality improvement.
|
Discussion |
|---|
|
TopAbstractIntroductionPatients and MethodsResultsCommentDiscussionReferences |
|---|
DR ALLEN: Once we did the risk adjustment, volume was not a risk factor. You saw the graph that demonstrated the risk for prolonged length of stay was not much different at each center after risk adjustment, so for this database we did not find volume a significant predictor. However, there were only 58 participants in our analysis.
DR DANIEL P. RAYMOND (Rochester, NY): These analyses require that we are dependent upon the variables that are entered into the analysis, and if you miss one significant variable, it can disrupt the quality of the analysis. Considering that, and I think the STS database does an outstanding job of trying to pick up the variables, there are things that I think we all can think of that may be additional predictors that are not included; for instance, nutritional parameters. We all know that malnutrition is a very significant predictor of poor outcomes, yet maybe the surrogate would be weight loss, which I don't know if it is a very reliable predictor, that is included. Are there efforts to look into other variables that may be impacting these models and trying to make the model better? Such as alcohol abuse, narcotic abuse, obesity and other variables that may have an impact on length of stay?
DR ALLEN: There are a variety of risk factors that on the initial pass did not show any relationship, so those are omitted; but you are right, there are other factors that may affect the results. The problem with collecting every conceivable variable is that it makes the database difficult to maintain and more expensive. The other problem we have is the completeness of the data. For example, the DLCO, which we probably would think would be a good predictor, had 37% of the values missing. Whether they weren't recorded or not done, we don't know. They are just not in the database. So completeness of the database is another problem, not just what specific factors we pick.
DR JOE B. PUTNAM (Nashville, TN): Mark, thank you for sharing this data with us. I just have one comment. The National Quality Forum has established a clinician-level quality outcome measure for perioperative services. The Steering Committee exists to evaluate models to improve quality of care. The Society of Thoracic Surgeons has submitted this prolonged length of stay as a risk-adjusted model as a quality measure, along with lung cancer staging and pulmonary function studies. The Steering Committee of the National Quality Forum views these as important indicators, and certainly appreciates the efforts of the Society in making these measures available. To the individuals in this audience and to the Society, this presentation represents the application of the General Thoracic Surgery Database to establish quality outcomes for our patients in the United States. I look forward to more of this risk-adjusted information coming forward in the future. Thank you.
DR ALLEN: Thank you, Bill.
DR PHILIP A. LINDEN (Cleveland, OH): Mark, I saw that you didn't really make a differentiation between VATS and open surgery. A prolonged length of stay of 14 days may be appropriate for open, but for VATS, typically if anyone stays longer than 7 days, many would consider that a very prolonged length of stay. Do you think that the predictors of length of stay may be different for a VATS type lobectomy than that for open lobectomy?
DR ALLEN: Well, since the data are not randomized and it is a highly selected group that undergoes a VATS lobectomy, we didn't do the analysis comparing length of stay for VATS vs open lobectomy since it would be biased because of a selection bias.
|
References |
|---|
|
TopAbstractIntroductionPatients and MethodsResultsCommentDiscussionReferences |
|---|
This article has been cited by other articles:
|
|
 |
|
  P. H. Schipper, B. S. Diggs, R. M. Ungerleider, and K. F. Welke The influence of surgeon specialty on outcomes in general thoracic surgery: a national sample 1996 to 2005. Ann. Thorac. Surg., November 1, 2009; 88(5): 1566 - 1573. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
M. F. Berry, J. Hanna, B. C. Tong, W. R. Burfeind Jr, D. H. Harpole, T. A. D'Amico, and M. W. Onaitis Risk factors for morbidity after lobectomy for lung cancer in elderly patients. Ann. Thorac. Surg., October 1, 2009; 88(4): 1093 - 1099. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
D. P. Mason, S. Subramanian, E. R. Nowicki, J. D. Grab, S. C. Murthy, T. W. Rice, and E. H. Blackstone Impact of smoking cessation before resection of lung cancer: a Society of Thoracic Surgeons General Thoracic Surgery Database study. Ann. Thorac. Surg., August 1, 2009; 88(2): 362 - 370. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
F. Farjah, D. R. Flum, T. K. Varghese Jr, R. G. Symons, and D. E. Wood Surgeon Specialty and Long-Term Survival After Pulmonary Resection for Lung Cancer Ann. Thorac. Surg., April 1, 2009; 87(4): 995 - 1006. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| ANN THORAC SURG | ASIAN CARDIOVASC THORAC ANN | EUR J CARDIOTHORAC SURG |
| J THORAC CARDIOVASC SURG | ICVTS | ALL CTSNet JOURNALS |
