- Email: [email protected]
Analyzing Risk Factors for Morbidity and Mortality after Lung Resection for Lung Cancer Using the NSQIP Database
Analyzing Risk Factors for Morbidity and Mortality after Lung Resection for Lung Cancer Using the NSQIP Database Raymond A Jean, MD, Matthew R DeLuzio, DO, Alexander I Kraev, MD, Gongyi Wang, Daniel J Boffa, MD, FACS, Frank C Detterbeck, MD, FACS, Zuoheng Wang, PhD, Anthony W Kim, MD, FACS
MS,
Our goal was to develop a predictive model that identifies how preoperative risk factors and perioperative complications lead to mortality after anatomic pulmonary resections. STUDY DESIGN: This was a retrospective cohort study. The American College of Surgeons NSQIP database was examined for all patients undergoing elective lobectomies for cancer from 2005 through 2012. Fifty-eight pre- and intraoperative risk factors and 13 complications were considered for their impact on perioperative mortality within 30 days of surgery. Multivariate logistic regression and a logistic regression model using least absolute shrinkage and selection operator (LASSO) selection methods were used to identify preoperative risk factors that were significant for predicting mortality, either through or independent of complications. Only factors that were significant under both the multivariate logistic regression and LASSO-selected models were considered to be validated for the final model. RESULTS: There were 6,435 lobectomies identified. After multivariate logistic regression modeling, 28 risk factors and 5 complications were found to be predictors for mortality. This was then tested against the LASSO method. There were 7 factors shared between the LASSO and multivariate logistic regressions that predicted mortality based on comorbidity: age (p ¼ 0.007), male sex (p ¼ 0.011), open lobectomy (p ¼ 0.001), preoperative dyspnea at rest (p < 0.001), preoperative dyspnea on exertion (p ¼ 0.003), preoperative dysnatremia (serum sodium <135 mEq/L or >145 mEq/L) (p ¼ 0.011), and preoperative anemia (p ¼ 0.002). Of these, 3 variables predicted mortality independent of any complications: dyspnea at rest, dyspnea on exertion, and dysnatremia. CONCLUSIONS: The clinical factors that predict postoperative complications and mortality are multiple and not necessarily aligned. Efforts to improve quality after anatomic pulmonary resections should focus on mechanisms to address both types of adverse outcomes. (J Am Coll Surg 2016;222:992e1000. 2016 by the American College of Surgeons. Published by Elsevier Inc. All rights reserved.)
BACKGROUND:
Several models have been designed to detect and quantify elevated risk for perioperative mortality in patients undergoing primary anatomic lobar resection.1-6 However, no accepted model currently exists that is considered completely accurate. In primary lung resections, there have been several studies investigating factors that impact postoperative mortality.7-9 There are data to suggest that the development of pulmonary complications is the primary driver of mortality after pulmonary resections.10 However, the contribution of complications to postoperative mortality remains incompletely understood. We sought to use predictive models to find associations and correlations between perioperative risk,
Disclosure Information: Nothing to disclose. Disclaimer: The American College of Surgeons NSQIP and the hospitals participating in the American College of Surgeons NSQIP are the source of the data used herein; they have not verified and are not responsible for the statistical validity of the data analysis or the conclusions derived by the authors. Presented at the 96th Annual Meeting of the New England Surgical Society, Newport, RI, September 2015. Received October 21, 2015; Revised February 19, 2016; Accepted February 22, 2016. From the Departments of Surgery (Jean, DeLuzio) and Biostatistics (G Wang, Z Wang), and Section of Thoracic Surgery (Boffa, Detterbeck, Kim), Yale School of Medicine, New Haven, CT, and Cardiovascular Surgery, Billings Clinic, Billings, MT (Kraev). Correspondence address: Anthony W Kim, MD, FACS, Section of Thoracic Surgery, Yale School of Medicine, 330 Cedar Street, BB 205, New Haven, CT, 06520. email: [email protected]
ª 2016 by the American College of Surgeons. Published by Elsevier Inc. All rights reserved.
992
http://dx.doi.org/10.1016/j.jamcollsurg.2016.02.020 ISSN 1072-7515/16
Vol. 222, No. 6, June 2016
complications, and mortality. In describing the interplay between risk factors, complications, and mortality, there can be 3 possibilities: first, that risk factors predict complications, which in turn can lead to perioperative mortality. Second, that risk factors simultaneously lead to complications and to mortality in a way that is not solely due to those complications. Finally, it is possible for risk factors to predict for perioperative mortality completely independent of any complications. We hypothesize that perioperative risk factors can be categorized with regard to their ability to predict mortality through their association with postoperative complications. The objective of this study was to investigate these 3 possible pathways to mortality and thereby gain insight into the association of these entities with perioperative mortality in patients undergoing anatomic pulmonary resections for lung cancer by using the large, nationally validated NSQIP database.
METHODS Data source The American College of Surgeons (ACS) NSQIP data from 2005 through 2012 was used to perform the analysis. The ACS NSQIP is a database of inpatient outcomes and quality measures collected from hundreds of participating hospitals that use the data for feedback about quality improvement.11 The NSQIP data were queried to identify all patients who underwent anatomic pulmonary resections. The NSQIP definitions of preoperative, intraoperative, and postoperative factors were incorporated into this analysis.12 This study was approved by the Yale University Human Investigation Committee. Patients were screened for inclusion based on the CPT codes that are recorded in the NSQIP dataset. The NSQIP database provides information on up to 10 procedures performed for each patient. Patients were selected for inclusion in this study if the CPT code for an anatomic lobar lung resection (either open or minimally invasive) was listed as their first, second, or third procedure. The ICD-9 codes were used then to filter the data to capture only those anatomic resections performed for oncologic pathology. All CPT and ICD-9 codes used are listed in Appendix 1 (Available online). All emergency cases were excluded. Variable selection The primary end point of this investigation was perioperative mortality, defined as 30-day mortality according to the NSQIP. There were 58 preoperative and intraoperative factors included as risk factors, and 13 postoperative complications. To allow for ease of interpretation,
Jean et al
Analyzing Mortality after Lung Resection
993
continuous variables were either dichotomized to “yes” or “no” binomial variables, or were subdivided into several defined groups when possible. Statistical analysis Significant predictors for mortality were identified by 2 methods; first by the use of traditional multivariate logistic regression modeling and second, by the use of the least absolute selection and shrinkage operator (LASSO) variable selection and logistic regression modeling. In the traditional multivariate logistic model, each risk factor was analyzed for its ability to be prognostic for mortality using forward stepwise selection. Additionally, each risk factor was also assessed for its ability to be prognostic for the set of complications that significantly predicted mortality. This was then repeated using the LASSO-selected variable model for validation. This model selection operator has been used in several outcomes studies, in cases where the large number of potential variables are available for model fitting. The LASSO method is used in this setting for its ability to filter through large numbers of variables and to shrink effect coefficients. The LASSO selection operator applies a penalty to large coefficients, which for large numbers of variables results in more interpretable regression models of manageable size and additionally allows for a minimization of errors in stepwise model selection.13 The LASSO involves penalizing the absolute size of the regression coefficients, which is convenient when dealing with highly correlated predictors. In classical multivariate logistic regression, it is often necessary to check for multicollinearity among predictors before performing an automatic selection procedure. In the case of this study, the LASSO selection models are designed to better deal with overfitting that is often found in standard logistic regression models. For both the standard multivariate logistic regression and LASSO, to sort out the relationship among these identified risk factors, complications, and mortality, a regression model, with mortality as a response variable and both risk factors and complications as predictors, was created. By adding complications into the model, the residual effect of risk factors on mortality, accounting for the contribution of complications to mortality, could be determined. This approach could be thought of as modeling the impact of risk factors on mortality adjusted by the risk of causing postoperative complications. Those risk factors that predicted for mortality directly would remain significant, and those that predicted complication-related mortality only would not be significant and would then be removed from the model.
994
Jean et al
J Am Coll Surg
Analyzing Mortality after Lung Resection
Complications that remained significant were described as “high-risk complications.” In the final step of model building, to determine which risk factors were independently predictive for mortality without predicting for a high-risk complication, regression analysis was performed with each of the high-risk complications as a dependent variable, and the significant risk factors as potential predictors. This model then identified those risk factors that were not significant predictors for a complication, but had been shown to be predictive for mortality. All statistical analysis was done using the R Foundation for Statistical Computing software version 3.11 using the “Penalized” statistical package (www.Rproject.org). Risk model scenarios In categorizing the relationship between risk factors, complications and perioperative mortality, there were 3 possible scenarios. In scenario 1, a risk factor was found to be a significant predictor for a postoperative complication. In scenario 2, a risk factor is not only an independent predictor for mortality, but also predicts for the development of a high-risk complication. Finally, in Scenario 3, a risk factor is an independent predictor for mortality, but does not predict for the development of a highrisk complication.
RESULTS A total of 6,435 patients were filtered from the NSQIP data as having undergone anatomic pulmonary resections for lung cancer and constituted the study cohort; 38% (2,451 of 6,435) were video-assisted thoracic lobectomies and 62% (3,984 of 6,435) were open lobectomies. The population was 53% female (3,414 of 6,435), 81% white (5,212 of 6,435), with 30% smoking at the time of surgery and 30% with a 30 or more pack-year history of tobacco use. The descriptive statistics of the study are summarized in Table 1. Complications that significantly predicted for perioperative mortality In a multivariate logistic model using complications as predictors for perioperative mortality, 5 postoperative complications were found to be significant predictors for mortality: reintubation (p < 0.001), progressive renal insufficiency (p ¼ 0.006), CVA/stroke (p ¼ 0.002), cardiac arrest requiring CPR (p < 0.001), and septic shock (p < 0.001). These 5 were used as outcomes variables for perioperative risk factors in the traditional logistic and LASSO.
Table 1. Demographic Characteristics of the Dataset Characteristic
Patients Sex Male Female Race White Black Asian Other Unknown Age, y, mean (SD) Younger than 50 y 50 to 69 y 70 to 79 y 80þ y Operation Video-assisted thoracic lobectomy Open lobectomy Deaths
Data
6,435 (100.0) 3,021 (46.9) 3,414 (53.1) 5,212 350 211 76 586 66.8 390 3,235 2,125 685
(81.0) (5.4) (3.3) (1.2) (9.1) (10.7) (6.1) (50.3) (33.0) (10.6)
2,451 (38.1) 3,984 (61.9) 120 (1.9)
Data are presented as n (%) unless otherwise noted.
Comparison of results of least absolute selection and shrinkage operator and standard multivariate logistic model Significant predictors for each of the significant complications, using the logistic regression model and the LASSO selected models are listed in Tables 2 and 3. In detecting significant predictors for mortality and complications, there was a strong degree of agreement between the standard multivariate logistic and LASSO modeling methods (Table 4). For the most part, the LASSO resulted in a model with fewer significant predictors than the standard multivariate logistic. In addition, it is notable that for both mortality and complications, the odds ratios for each predictor were similar between models, expressing some uniformity between the 2 in the effect size represented by the presence or absence of these risk factors. For 30-day perioperative mortality, shared predictors were age, male sex, dyspnea at rest, dyspnea on exertion, dysnatremia (serum sodium <135 mEq/L or >145 mEq/L), anemia, and open lobectomy. Of these, only dyspnea at rest and dyspnea on exertion predicted for mortality, but not complications. Five risk factors were classified under scenario 1 because they predicted for a high-risk complication but did not predict for mortality. These included American Society of Anesthesiologists classification, COPD, current smoking, operative time, and hypertension. Four risk factors were independent predictors for mortality and
Significant Predictors of Complications in Multivariate Model Mortality
Predictor Age, y
p Value
OR
95% CI
0.002
1.03
1.013e1.056
ASA classification 3 ASA classification 4/5 Contaminated/ dirty wound Current smoker Dependent functional status Male sex Open lobectomy Operative time
<0.001
1.03
0.028
1.62
0.015
1.96
0.003
1.53
4.19
<0.001 0.005
4.33 1.83
1.990e9.415 1.199e2.789
<0.001
0.002
2.08
1.300e3.335
<0.001 <0.001
0.044
1.53
1.80 1.00
1.015 e1.045 1.054 e2.504 1.141 e3.362
4.24
1.969e9.131
0.001
2.03
1.348e3.047
0.003
2.57
1.379e4.808
<0.001
4.63
2.276e9.414
<0.001
2.84
OR
95% CI
0.016
1.05
1.009e1.087
Cardiac arrest p Value OR 95% CI
0.028
4.94
1.187e20.572
0.044
5.24
1.047e26.241
1.339 e2.423 1.002 e1.005
0.024
1.00
0.014
4.72
1.365 e16.293
0.003
3.93
1.617 e9.551
0.001
3.02
1.558e5.857
0.035
2.21
1.057e4.621
OR
95% CI
0.010 0.016
9.34 1.79
1.726e50.500 1.117e2.878
<0.001
6.96
3.169e15.273
0.001
2.20
1.369e3.538
0.011
2.89
1.274e6.572
0.003
2.33
1.336e34.074
<0.001
8.39
4.035e17.435 (Continued)
1.000 e1.007
1.353 e2.329
1.601 e5.018
Septic shock p Value
0.010
3.98
1.387e11.400
Analyzing Mortality after Lung Resection
<0.001
1.78
CVA/stroke p Value
1.158 e2.018 2.300 e7.620
1.011e2.322
<0.001
Renal insufficiency p Value OR 95% CI
Jean et al
Preoperative EtOH use, >2 drinks/d Preoperative Hct <35% (female) or 40% (male) Preoperative history of COPD Preoperative history of CVA Preoperative history of dyspnea at rest Preoperative history of dyspnea on exertion Preoperative history of HTN Preoperative history of insulin-dependent DM Preoperative history of radiation therapy
Reintubation p Value OR 95% CI
Vol. 222, No. 6, June 2016
Table 2.
995
2.35 5.50 0.022
95% CI OR
3.327 e275.760
1.186 e206.651
0.002 Preoperative sodium <135 mEq/L or >145 mEq/L Preoperative transfusion >4 U packed RBC
30.29
15.65 0.037 1.59 0.040
0.017
OR
1.88
95% CI
1.118e3.167
1.82 0.023
1.087 e3.055 1.021 e2.465 Preoperative history of weight loss >10% Preoperative history of PCI Preoperative open wound/ wound infection Preoperative sepsis
Predictor
Preoperative pneumonia was included in the analysis and did not achieve significance. ASA, American Society of Anesthesiologists; DM, diabetes mellitus; HTN, hypertension; OR, odds ratio; PCI, percutaneous coronary intervention.
1.278e23.684
0.004
95% CI OR
Septic shock
p Value Cardiac arrest p Value OR 95% CI Renal insufficiency p Value OR 95% CI Mortality
Reintubation p Value OR 95% CI p Value
Table 2. Continued
J Am Coll Surg
development of a high-risk complication, and fell under scenario 2: age, male sex, anemia, and open lobectomy. Only 3 risk factors predicted for mortality without predicting for a high-risk complication: dyspnea at rest, dyspnea on exertion, and dysnatremia. These results are summarized in Figure 1.
p Value
CVA/stroke
Analyzing Mortality after Lung Resection
1.306e4.222
Jean et al
996
DISCUSSION There has been a strong interest in the prediction of complications and mortality associated with pulmonary resections.1,14,15 In addition, there are additional studies that also have examined individual factors that can influence mortality, such as age or type of resection.10,16,17 Interestingly, there is a paucity of data examining the intermediary end point of complications as a pathway toward mortality. In an attempt to better understand the relationship among risk factors, complications, and mortality, we used multivariate regression analysis to identify significant predictors for mortality and complications, using a standard multivariate logistic regression, and validating it with a LASSO selection model. It is generally assumed that the development of postoperative complications directly contributes to perioperative mortality, and evidence shows that postoperative complications after lung resection are both prevalent and affect postoperative mortality. Watanabe and colleagues10 reported that the development of pulmonary complications, such as pneumonia and bronchopulmonary fistula, were associated with nearly 60% of the mortalities after lung resection. In addition, this group also found that 17% of mortalities were due to postoperative CVAs or cardiac events.10 Allen and colleagues18 demonstrated an overall complication rate of 38% with pulmonary resections for lung cancer, with bilobectomies and segmentectomies having complication rates >45%. Although the operative mortality in this study was 1%, nearly all causes of death were the result of complications, the most frequent of which included multisystem organ failure (29%), pneumonia (21%), and CVA (14%).18 The importance of this analysis lies in its attempt to provide additional stratification to the understanding of perioperative mortality and morbidity. In an era in which surgical risk stratification has become increasingly scrutinized, these factors are important because they examine the shared contribution between development of complications and direct risk of death. Studies that have examined predictors for perioperative mortality and morbidity have demonstrated the complex nature by which several factors predict for only mortality, only morbidity, or both.3,19-22 Although ostensibly morbidity and mortality are linked together, subtle nuances in
Mortality Predictor Age, y
p Value
OR
95% CI
0.007
1.03
1.008e1.050
ASA classification 3 Current smoker Male sex Open lobectomy
<0.001
1.03
1.016 e1.045
0.001
1.59
1.207 e2.087
<0.001
1.96
<0.001
1.00
1.459 e2.622 1.002 e1.005
Renal insufficiency p Value OR 95% CI
CVA/stroke p Value
OR
0.009
1.05
95% CI
Cardiac arrest p Value OR 95% CI
0.011 0.001
1.71 2.20
1.129e2.587 1.378e3.506
1.90
1.00
0.006
3.55
1.007 e1.007
1.315e5.165 1.079e4.696
0.006
1.98
1.221e3.208
0.008
2.08
1.212e3.552
0.015
1.00
1.001e1.005
0.001
2.15
1.348e3.427
0.006
2.19
1.252e3.834
0.027
1.69
1.061e2.685
1.446 e8.706
4.30
2.007e9.218
0.003
1.89
1.240e2.875
1.92
1.471 e2.510
0.003
1.163e 3.201
0.006
1.45
26.07
3.121e217.659
1.114 e1.897
ASA classification 4/5, contaminated/dirty wound, dependent functional status, preoperative history of CVA, preoperative history of insulin dependent DM, preoperative history of radiation therapy, preoperative history of weight loss >10%, preoperative history of PCI, preoperative open wound/wound infection, and preoperative transfusion >4 U packed RBC were included in the analysis and did not achieve significance. ASA, American Society of Anesthesiologists; DM, diabetes mellitus; HTN, hypertension; OR, odds ratio; PCI, percutaneous coronary intervention.
Analyzing Mortality after Lung Resection
<0.001
1.93
2.61 2.25
1.021e17.904
1.274e2.832 <0.001
0.011
4.28
Jean et al
0.002
0.006 0.031 0.016
Septic shock p Value OR 95% CI
1.013e1.091 0.047
Operative time Preoperative EtOH use >2 drinks/d Preoperative Hct <35% (female) or 40% (male) Preoperative history of COPD Preoperative history of dyspnea at rest Preoperative history of dyspnea on exertion Preoperative history of HTN Preoperative pneumonia Preoperative sepsis Preoperative sodium <135 mEq/L or >145 mEq/L
Reintubation p Value OR 95% CI
Vol. 222, No. 6, June 2016
Table 3. Significant Predictors of Complications for Least Absolute Shrinkage and Selection Operator Model
997
998
Jean et al
Analyzing Mortality after Lung Resection
Table 4. Shared Risk Factors Shared risk factors for predicting mortality Age in years Male sex Preoperative history of dyspnea at rest Preoperative history of dyspnea on exertion Preoperative sodium <135 mEq/L or >145 mEq/L Preoperative Hct <35% (female) or 40% (male) Open lobectomy Shared risk factors for predicting reintubation Age in years Current smoker Preoperative history of COPD Open lobectomy Operative time Shared risk factors for predicting renal insufficiency Variable Preoperative Hct <35% (female) or 40% (male) Operative time Shared risk factors for predicting CVA/stroke Age in years Shared risk factors for predicting cardiac arrest requiring CPR Male sex ASA classification 3 Open lobectomy Shared risk factors for predicting septic shock Current smoker Preoperative history of COPD Preoperative history of HTN ASA, American Society of Anesthesiologists; HTN, hypertension.
patient’s health status are at play and are most likely responsible for any variability. In our study, only 3 risk factors, dyspnea at rest, dyspnea on exertion, and dysnatremia, were associated with mortality without predicting for a high-risk complication. Given that our study is limited to the exploration of associations rather than definitive causality, we are unable to definitively state why these 3 risk factors were categorized as such. It is notable that these risk factors are associated strongly with severe derangements in multiple organ systems, including respiratory, metabolic, cardiovascular, renal, and neurologic processes. It is possible that these risk factors are proxies for overall severe critical illness rather than simply predictors for a subsequent complication. Therefore, the clinical implications of these results have a complex relationship with preoperative care for patients undergoing lobectomies for cancer. Although dysnatremia and dyspnea should not be taken to be absolute contraindications to surgical therapy, these results allow clinicians to view these risk factors as representative of
J Am Coll Surg
an increased perioperative risk. Therefore, physicians might be warranted in taking a guarded approach to patients with risk factors highly predictive of mortality and complications. It is notable that although there is good agreement between the standard multivariate and LASSO-selected models, the primary reason for mismatch between risk factors predicting mortality versus complications might be due to the observation of associations rather than causative relationships; a phenomenon seen in many retrospective database studies. A secondary reason for any disconnect might be due to the fact that advances in the processes of care might be able to atone for the occurrence of complications. Sukumar and colleagues23 found that although complications after oncologic operations have increased during the past decade, this increase was met with a concomitant decrease in failure to rescue rates and mortality, possibly due to improved recognition and management. Similarly, Downey and colleagues24 examined the trends of several complications and surgical quality indicators and demonstrated that several indicators, including postoperative pulmonary embolism, postoperative sepsis, and postoperative respiratory failure, have increased in reported incidence, despite the fact that in-hospital mortality after major postoperative complication (failure to rescue) had shown a significant decreasing trend. Additional study into detecting and appropriately managing the development of complications might mitigate against its consequences on overall mortality. Ultimately, it might be that dealing with complications can be just as important as assessing upfront risk. There are a number of limitations to this study. First and foremost, data used in study were from a national registry and, therefore, are associated with the limitations associated with the use of the NSQIP database. As a result, our analysis is limited to a retrospective data analysis, and might not represent direct causality as opposed to correlations. In addition, all assessments of mortality are limited to 30 days postoperatively. As such, the full impact of postoperative comorbidity after this period could not be studied directly. Another important limitation, related to this issue, is the fact that the NSQIP is not a dedicated thoracic surgery database and, therefore, there might be risk of bias in detecting significant predictors for primary lung resections. Also, because many disease and treatment patterns are associated with particular hospitals or regions, a hierarchical regression model was considered for this analysis; however, the NSQIP does not provide hospital-specific identifiers and we were unable to do so. Along this same vein, a significant limitation is that we were not able to perform external
Vol. 222, No. 6, June 2016
Jean et al
Analyzing Mortality after Lung Resection
999
Figure 1. Relationship among shared risk factors, complications, and mortality. Risk factors are listed on the left, with arrows to the left identifying the significant correlation to mortality. Arrows to the right identify a significant correlation of a given risk factor with postoperative complications. Those risk factors that only associate with complications are classified in scenario 1. Those that associate to both complications and mortality are classified in scenario 2. Those that only associate with mortality are classified as scenario 3. ASA, American Society of Anesthesiologists.
validation of our data using another validated data source. We did perform an internal validation by dividing the full set of data into validation sets of equal size, and the comparison of the multivariate logistic regression model and the logistic regression with LASSO selection were found to identify fewer significant risk factors (data not shown). These results were comparable with those of the whole sample. As such, however, there does exist significant potential for using this method with other datasets, or combined datasets, such as the Society of Thoracic Surgeons data. Additional research is warranted in the continuation of this analysis going forward, however, the nature of these results is such that it merits consideration of how thoracic surgeons can unify the reporting of data to best allow for retrospective analysis.
CONCLUSIONS Ultimately, this analysis demonstrates that the relationship between surgical risk factors, complications, and postoperative mortality is complex, and additional investigation into the mechanisms of accrual of mortality risk in the perioperative period is warranted.
Author Contributions Study conception and design: Jean, DeLuzio, Kraev, G Wang, Boffa, Detterbeck, Z Wang, Kim Acquisition of data: Kim Analysis and interpretation of data: Jean, DeLuzio, Kraev, G Wang, Z Wang, Kim Drafting of manuscript: Jean, DeLuzio, Kraev, G Wang, Boffa, Detterbeck, Z Wang, Kim Critical revision: Jean, DeLuzio, Kraev, G Wang, Boffa, Detterbeck, Z Wang, Kim REFERENCES 1. Pierce RJ, Copland JM, Sharpe K, Barter CE. Preoperative risk evaluation for lung cancer resection: predicted postoperative product as a predictor of surgical mortality. Am J Respir Crit Care Med 1994;150:947e955. 2. Khuri SF, Daley J, Henderson W, et al. Risk adjustment of the postoperative mortality rate for the comparative assessment of the quality of surgical care: results of the National Veterans Affairs Surgical Risk Study. J Am Coll Surg 1997;185:315e327. 3. Gibbs J, Cull W, Henderson W, et al. Preoperative serum albumin level as a predictor of operative mortality and morbidity: results from the National VA Surgical Risk Study. Arch Surg 1999;134:36e42.
1000
Jean et al
Analyzing Mortality after Lung Resection
4. Harpole DH, DeCamp MM, Daley J, et al. Prognostic models of thirty-day mortality and morbidity after major pulmonary resection. J Thorac Cardiovasc Surg 1999;117: 969e979. 5. Fink AS, Hutter MM, Campbell DC, et al. Comparison of risk-adjusted 30-day postoperative mortality and morbidity in Department of Veterans Affairs hospitals and selected university medical centers: general surgical operations in women. J Am Coll Surg 2007;204: 1127e1136. 6. Henderson WG, Khuri SF, Mosca C, et al. Comparison of risk-adjusted 30-day postoperative mortality and morbidity in Department of Veterans Affairs hospitals and selected university medical centers: general surgical operations in men. J Am Coll Surg 2007;204:1103e1114. 7. Kozower BD, Sheng S, O’Brien SM, et al. STS database risk models: predictors of mortality and major morbidity for lung cancer resection. Ann Thorac Surg 2010;90:875e881; discussion 881883. 8. Bernard A, Rivera C, Pages PB, et al. Risk model of inhospital mortality after pulmonary resection for cancer: a national database of the French Society of Thoracic and Cardiovascular Surgery (Epithor). J Thorac Cardiovasc Surg 2011;141:449e458. 9. Poullis M, McShane J, Shaw M, et al. Prediction of in-hospital mortality following pulmonary resections: improving on current risk models. Eur J Cardiothorac Surg 2013;44: 238e242; discussion 242243. 10. Watanabe S, Asamura H, Suzuki K, Tsuchiya R. Recent results of postoperative mortality for surgical resections in lung cancer. Ann Thorac Surg 2004;78:999e1002. 11. American College of Surgeons. About ACS NSQIP. Available at: http://site.acsnsqip.org/about/. Accessed November 12, 2013. 12. American College of Surgeons. User guide for the 2012 ACS NSQIP Participant Use Data File. https://www. facs.org/w/media/files/quality%20programs/nsqip/ug12.ashx. Accessed November 14, 2013.
J Am Coll Surg
13. Tibshirani R. Regression shrinkage and selection via the lasso. J R Stat Soc Ser B 1996;58:267e288. 14. Damhuis RA, Schu¨tte PR. Resection rates and postoperative mortality in 7,899 patients with lung cancer. Eur Respir J 1996;9:7e10. 15. Wyser C, Stulz P, Sole`r M, et al. Prospective evaluation of an algorithm for the functional assessment of lung resection candidates. Am J Respir Crit Care Med 1999;159:1450e1456. 16. McGuire HH. Measuring and managing quality of surgery. Arch Surg 1992;127:733. 17. Nagasaki F. Complications of surgery in the treatment of carcinoma of the lung. Chest J 1982;82:25. 18. Allen MS, Darling GE, Pechet TTV, et al. Morbidity and mortality of major pulmonary resections in patients with early-stage lung cancer: initial results of the randomized, prospective ACOSOG Z0030 trial. Ann Thorac Surg 2006;81: 1013e1020. 19. Bernard A, Ferrand L, Hagry O, et al. Identification of prognostic factors determining risk groups for lung resection. Ann Thorac Surg 2000;70:1161e1167. 20. Hamel MB, Henderson WG, Khuri SF, Daley J. Surgical outcomes for patients aged 80 and older: morbidity and mortality from major noncardiac surgery. J Am Geriatr Soc 2005;53: 424e429. 21. Alam N, Park BJ, Wilton A, et al. Incidence and risk factors for lung injury after lung cancer resection. Ann Thorac Surg 2007;84:1085e1091; discussion 1091. 22. Linden PA, Yeap BY, Chang MY, et al. Morbidity of lung resection after prior lobectomy: results from the Veterans Affairs National Surgical Quality Improvement Program. Ann Thorac Surg 2007;83:425e431; discussion 432. 23. Sukumar S, Roghmann F, Trinh VQ, et al. National trends in hospital-acquired preventable adverse events after major cancer surgery in the USA. BMJ Open 2013;3:e002843. 24. Downey JR, Hernandez-Boussard T, Banka G, Morton JM. Is patient safety improving? National trends in patient safety indicators: 1998-2007. Health Serv Res 2012;47: 414e430.
Vol. 222, No. 6, June 2016
Jean et al
Appendix 1. CPT and ICD-9 Codes Used in Selection of the Study Cohort CPT codes: open and thoracoscopic anatomical lung resection
32440 32445 32484 32442 32480 32663 ICD-9 codes to identify anatomic resections for oncologic pathology 031.0 170.3 492 114.3 171.4 492.0 114.4 171.9 492.8 114.5 193 494.0 117.3 195.1 494.1 117.7 197 510.0 150.8 197.0 510.9 151.0 209 512.0 158.9 209.21 512.8 162 209.6 513.0 162.2 209.61 515 162.3 212.3 518 162.4 212.4 518.89 162.5 212.6 747.3 162.8 235.7 748.5 162.9 235.8 748.8 163 239.1 759.6 163.9 417.8 786.6 164 446.4 793.1 164.3 448.0 844.9 164.9 485 V10.42
Analyzing Mortality after Lung Resection
1000.e1
MS,
Our goal was to develop a predictive model that identifies how preoperative risk factors and perioperative complications lead to mortality after anatomic pulmonary resections. STUDY DESIGN: This was a retrospective cohort study. The American College of Surgeons NSQIP database was examined for all patients undergoing elective lobectomies for cancer from 2005 through 2012. Fifty-eight pre- and intraoperative risk factors and 13 complications were considered for their impact on perioperative mortality within 30 days of surgery. Multivariate logistic regression and a logistic regression model using least absolute shrinkage and selection operator (LASSO) selection methods were used to identify preoperative risk factors that were significant for predicting mortality, either through or independent of complications. Only factors that were significant under both the multivariate logistic regression and LASSO-selected models were considered to be validated for the final model. RESULTS: There were 6,435 lobectomies identified. After multivariate logistic regression modeling, 28 risk factors and 5 complications were found to be predictors for mortality. This was then tested against the LASSO method. There were 7 factors shared between the LASSO and multivariate logistic regressions that predicted mortality based on comorbidity: age (p ¼ 0.007), male sex (p ¼ 0.011), open lobectomy (p ¼ 0.001), preoperative dyspnea at rest (p < 0.001), preoperative dyspnea on exertion (p ¼ 0.003), preoperative dysnatremia (serum sodium <135 mEq/L or >145 mEq/L) (p ¼ 0.011), and preoperative anemia (p ¼ 0.002). Of these, 3 variables predicted mortality independent of any complications: dyspnea at rest, dyspnea on exertion, and dysnatremia. CONCLUSIONS: The clinical factors that predict postoperative complications and mortality are multiple and not necessarily aligned. Efforts to improve quality after anatomic pulmonary resections should focus on mechanisms to address both types of adverse outcomes. (J Am Coll Surg 2016;222:992e1000. 2016 by the American College of Surgeons. Published by Elsevier Inc. All rights reserved.)
BACKGROUND:
Several models have been designed to detect and quantify elevated risk for perioperative mortality in patients undergoing primary anatomic lobar resection.1-6 However, no accepted model currently exists that is considered completely accurate. In primary lung resections, there have been several studies investigating factors that impact postoperative mortality.7-9 There are data to suggest that the development of pulmonary complications is the primary driver of mortality after pulmonary resections.10 However, the contribution of complications to postoperative mortality remains incompletely understood. We sought to use predictive models to find associations and correlations between perioperative risk,
Disclosure Information: Nothing to disclose. Disclaimer: The American College of Surgeons NSQIP and the hospitals participating in the American College of Surgeons NSQIP are the source of the data used herein; they have not verified and are not responsible for the statistical validity of the data analysis or the conclusions derived by the authors. Presented at the 96th Annual Meeting of the New England Surgical Society, Newport, RI, September 2015. Received October 21, 2015; Revised February 19, 2016; Accepted February 22, 2016. From the Departments of Surgery (Jean, DeLuzio) and Biostatistics (G Wang, Z Wang), and Section of Thoracic Surgery (Boffa, Detterbeck, Kim), Yale School of Medicine, New Haven, CT, and Cardiovascular Surgery, Billings Clinic, Billings, MT (Kraev). Correspondence address: Anthony W Kim, MD, FACS, Section of Thoracic Surgery, Yale School of Medicine, 330 Cedar Street, BB 205, New Haven, CT, 06520. email: [email protected]
ª 2016 by the American College of Surgeons. Published by Elsevier Inc. All rights reserved.
992
http://dx.doi.org/10.1016/j.jamcollsurg.2016.02.020 ISSN 1072-7515/16
Vol. 222, No. 6, June 2016
complications, and mortality. In describing the interplay between risk factors, complications, and mortality, there can be 3 possibilities: first, that risk factors predict complications, which in turn can lead to perioperative mortality. Second, that risk factors simultaneously lead to complications and to mortality in a way that is not solely due to those complications. Finally, it is possible for risk factors to predict for perioperative mortality completely independent of any complications. We hypothesize that perioperative risk factors can be categorized with regard to their ability to predict mortality through their association with postoperative complications. The objective of this study was to investigate these 3 possible pathways to mortality and thereby gain insight into the association of these entities with perioperative mortality in patients undergoing anatomic pulmonary resections for lung cancer by using the large, nationally validated NSQIP database.
METHODS Data source The American College of Surgeons (ACS) NSQIP data from 2005 through 2012 was used to perform the analysis. The ACS NSQIP is a database of inpatient outcomes and quality measures collected from hundreds of participating hospitals that use the data for feedback about quality improvement.11 The NSQIP data were queried to identify all patients who underwent anatomic pulmonary resections. The NSQIP definitions of preoperative, intraoperative, and postoperative factors were incorporated into this analysis.12 This study was approved by the Yale University Human Investigation Committee. Patients were screened for inclusion based on the CPT codes that are recorded in the NSQIP dataset. The NSQIP database provides information on up to 10 procedures performed for each patient. Patients were selected for inclusion in this study if the CPT code for an anatomic lobar lung resection (either open or minimally invasive) was listed as their first, second, or third procedure. The ICD-9 codes were used then to filter the data to capture only those anatomic resections performed for oncologic pathology. All CPT and ICD-9 codes used are listed in Appendix 1 (Available online). All emergency cases were excluded. Variable selection The primary end point of this investigation was perioperative mortality, defined as 30-day mortality according to the NSQIP. There were 58 preoperative and intraoperative factors included as risk factors, and 13 postoperative complications. To allow for ease of interpretation,
Jean et al
Analyzing Mortality after Lung Resection
993
continuous variables were either dichotomized to “yes” or “no” binomial variables, or were subdivided into several defined groups when possible. Statistical analysis Significant predictors for mortality were identified by 2 methods; first by the use of traditional multivariate logistic regression modeling and second, by the use of the least absolute selection and shrinkage operator (LASSO) variable selection and logistic regression modeling. In the traditional multivariate logistic model, each risk factor was analyzed for its ability to be prognostic for mortality using forward stepwise selection. Additionally, each risk factor was also assessed for its ability to be prognostic for the set of complications that significantly predicted mortality. This was then repeated using the LASSO-selected variable model for validation. This model selection operator has been used in several outcomes studies, in cases where the large number of potential variables are available for model fitting. The LASSO method is used in this setting for its ability to filter through large numbers of variables and to shrink effect coefficients. The LASSO selection operator applies a penalty to large coefficients, which for large numbers of variables results in more interpretable regression models of manageable size and additionally allows for a minimization of errors in stepwise model selection.13 The LASSO involves penalizing the absolute size of the regression coefficients, which is convenient when dealing with highly correlated predictors. In classical multivariate logistic regression, it is often necessary to check for multicollinearity among predictors before performing an automatic selection procedure. In the case of this study, the LASSO selection models are designed to better deal with overfitting that is often found in standard logistic regression models. For both the standard multivariate logistic regression and LASSO, to sort out the relationship among these identified risk factors, complications, and mortality, a regression model, with mortality as a response variable and both risk factors and complications as predictors, was created. By adding complications into the model, the residual effect of risk factors on mortality, accounting for the contribution of complications to mortality, could be determined. This approach could be thought of as modeling the impact of risk factors on mortality adjusted by the risk of causing postoperative complications. Those risk factors that predicted for mortality directly would remain significant, and those that predicted complication-related mortality only would not be significant and would then be removed from the model.
994
Jean et al
J Am Coll Surg
Analyzing Mortality after Lung Resection
Complications that remained significant were described as “high-risk complications.” In the final step of model building, to determine which risk factors were independently predictive for mortality without predicting for a high-risk complication, regression analysis was performed with each of the high-risk complications as a dependent variable, and the significant risk factors as potential predictors. This model then identified those risk factors that were not significant predictors for a complication, but had been shown to be predictive for mortality. All statistical analysis was done using the R Foundation for Statistical Computing software version 3.11 using the “Penalized” statistical package (www.Rproject.org). Risk model scenarios In categorizing the relationship between risk factors, complications and perioperative mortality, there were 3 possible scenarios. In scenario 1, a risk factor was found to be a significant predictor for a postoperative complication. In scenario 2, a risk factor is not only an independent predictor for mortality, but also predicts for the development of a high-risk complication. Finally, in Scenario 3, a risk factor is an independent predictor for mortality, but does not predict for the development of a highrisk complication.
RESULTS A total of 6,435 patients were filtered from the NSQIP data as having undergone anatomic pulmonary resections for lung cancer and constituted the study cohort; 38% (2,451 of 6,435) were video-assisted thoracic lobectomies and 62% (3,984 of 6,435) were open lobectomies. The population was 53% female (3,414 of 6,435), 81% white (5,212 of 6,435), with 30% smoking at the time of surgery and 30% with a 30 or more pack-year history of tobacco use. The descriptive statistics of the study are summarized in Table 1. Complications that significantly predicted for perioperative mortality In a multivariate logistic model using complications as predictors for perioperative mortality, 5 postoperative complications were found to be significant predictors for mortality: reintubation (p < 0.001), progressive renal insufficiency (p ¼ 0.006), CVA/stroke (p ¼ 0.002), cardiac arrest requiring CPR (p < 0.001), and septic shock (p < 0.001). These 5 were used as outcomes variables for perioperative risk factors in the traditional logistic and LASSO.
Table 1. Demographic Characteristics of the Dataset Characteristic
Patients Sex Male Female Race White Black Asian Other Unknown Age, y, mean (SD) Younger than 50 y 50 to 69 y 70 to 79 y 80þ y Operation Video-assisted thoracic lobectomy Open lobectomy Deaths
Data
6,435 (100.0) 3,021 (46.9) 3,414 (53.1) 5,212 350 211 76 586 66.8 390 3,235 2,125 685
(81.0) (5.4) (3.3) (1.2) (9.1) (10.7) (6.1) (50.3) (33.0) (10.6)
2,451 (38.1) 3,984 (61.9) 120 (1.9)
Data are presented as n (%) unless otherwise noted.
Comparison of results of least absolute selection and shrinkage operator and standard multivariate logistic model Significant predictors for each of the significant complications, using the logistic regression model and the LASSO selected models are listed in Tables 2 and 3. In detecting significant predictors for mortality and complications, there was a strong degree of agreement between the standard multivariate logistic and LASSO modeling methods (Table 4). For the most part, the LASSO resulted in a model with fewer significant predictors than the standard multivariate logistic. In addition, it is notable that for both mortality and complications, the odds ratios for each predictor were similar between models, expressing some uniformity between the 2 in the effect size represented by the presence or absence of these risk factors. For 30-day perioperative mortality, shared predictors were age, male sex, dyspnea at rest, dyspnea on exertion, dysnatremia (serum sodium <135 mEq/L or >145 mEq/L), anemia, and open lobectomy. Of these, only dyspnea at rest and dyspnea on exertion predicted for mortality, but not complications. Five risk factors were classified under scenario 1 because they predicted for a high-risk complication but did not predict for mortality. These included American Society of Anesthesiologists classification, COPD, current smoking, operative time, and hypertension. Four risk factors were independent predictors for mortality and
Significant Predictors of Complications in Multivariate Model Mortality
Predictor Age, y
p Value
OR
95% CI
0.002
1.03
1.013e1.056
ASA classification 3 ASA classification 4/5 Contaminated/ dirty wound Current smoker Dependent functional status Male sex Open lobectomy Operative time
<0.001
1.03
0.028
1.62
0.015
1.96
0.003
1.53
4.19
<0.001 0.005
4.33 1.83
1.990e9.415 1.199e2.789
<0.001
0.002
2.08
1.300e3.335
<0.001 <0.001
0.044
1.53
1.80 1.00
1.015 e1.045 1.054 e2.504 1.141 e3.362
4.24
1.969e9.131
0.001
2.03
1.348e3.047
0.003
2.57
1.379e4.808
<0.001
4.63
2.276e9.414
<0.001
2.84
OR
95% CI
0.016
1.05
1.009e1.087
Cardiac arrest p Value OR 95% CI
0.028
4.94
1.187e20.572
0.044
5.24
1.047e26.241
1.339 e2.423 1.002 e1.005
0.024
1.00
0.014
4.72
1.365 e16.293
0.003
3.93
1.617 e9.551
0.001
3.02
1.558e5.857
0.035
2.21
1.057e4.621
OR
95% CI
0.010 0.016
9.34 1.79
1.726e50.500 1.117e2.878
<0.001
6.96
3.169e15.273
0.001
2.20
1.369e3.538
0.011
2.89
1.274e6.572
0.003
2.33
1.336e34.074
<0.001
8.39
4.035e17.435 (Continued)
1.000 e1.007
1.353 e2.329
1.601 e5.018
Septic shock p Value
0.010
3.98
1.387e11.400
Analyzing Mortality after Lung Resection
<0.001
1.78
CVA/stroke p Value
1.158 e2.018 2.300 e7.620
1.011e2.322
<0.001
Renal insufficiency p Value OR 95% CI
Jean et al
Preoperative EtOH use, >2 drinks/d Preoperative Hct <35% (female) or 40% (male) Preoperative history of COPD Preoperative history of CVA Preoperative history of dyspnea at rest Preoperative history of dyspnea on exertion Preoperative history of HTN Preoperative history of insulin-dependent DM Preoperative history of radiation therapy
Reintubation p Value OR 95% CI
Vol. 222, No. 6, June 2016
Table 2.
995
2.35 5.50 0.022
95% CI OR
3.327 e275.760
1.186 e206.651
0.002 Preoperative sodium <135 mEq/L or >145 mEq/L Preoperative transfusion >4 U packed RBC
30.29
15.65 0.037 1.59 0.040
0.017
OR
1.88
95% CI
1.118e3.167
1.82 0.023
1.087 e3.055 1.021 e2.465 Preoperative history of weight loss >10% Preoperative history of PCI Preoperative open wound/ wound infection Preoperative sepsis
Predictor
Preoperative pneumonia was included in the analysis and did not achieve significance. ASA, American Society of Anesthesiologists; DM, diabetes mellitus; HTN, hypertension; OR, odds ratio; PCI, percutaneous coronary intervention.
1.278e23.684
0.004
95% CI OR
Septic shock
p Value Cardiac arrest p Value OR 95% CI Renal insufficiency p Value OR 95% CI Mortality
Reintubation p Value OR 95% CI p Value
Table 2. Continued
J Am Coll Surg
development of a high-risk complication, and fell under scenario 2: age, male sex, anemia, and open lobectomy. Only 3 risk factors predicted for mortality without predicting for a high-risk complication: dyspnea at rest, dyspnea on exertion, and dysnatremia. These results are summarized in Figure 1.
p Value
CVA/stroke
Analyzing Mortality after Lung Resection
1.306e4.222
Jean et al
996
DISCUSSION There has been a strong interest in the prediction of complications and mortality associated with pulmonary resections.1,14,15 In addition, there are additional studies that also have examined individual factors that can influence mortality, such as age or type of resection.10,16,17 Interestingly, there is a paucity of data examining the intermediary end point of complications as a pathway toward mortality. In an attempt to better understand the relationship among risk factors, complications, and mortality, we used multivariate regression analysis to identify significant predictors for mortality and complications, using a standard multivariate logistic regression, and validating it with a LASSO selection model. It is generally assumed that the development of postoperative complications directly contributes to perioperative mortality, and evidence shows that postoperative complications after lung resection are both prevalent and affect postoperative mortality. Watanabe and colleagues10 reported that the development of pulmonary complications, such as pneumonia and bronchopulmonary fistula, were associated with nearly 60% of the mortalities after lung resection. In addition, this group also found that 17% of mortalities were due to postoperative CVAs or cardiac events.10 Allen and colleagues18 demonstrated an overall complication rate of 38% with pulmonary resections for lung cancer, with bilobectomies and segmentectomies having complication rates >45%. Although the operative mortality in this study was 1%, nearly all causes of death were the result of complications, the most frequent of which included multisystem organ failure (29%), pneumonia (21%), and CVA (14%).18 The importance of this analysis lies in its attempt to provide additional stratification to the understanding of perioperative mortality and morbidity. In an era in which surgical risk stratification has become increasingly scrutinized, these factors are important because they examine the shared contribution between development of complications and direct risk of death. Studies that have examined predictors for perioperative mortality and morbidity have demonstrated the complex nature by which several factors predict for only mortality, only morbidity, or both.3,19-22 Although ostensibly morbidity and mortality are linked together, subtle nuances in
Mortality Predictor Age, y
p Value
OR
95% CI
0.007
1.03
1.008e1.050
ASA classification 3 Current smoker Male sex Open lobectomy
<0.001
1.03
1.016 e1.045
0.001
1.59
1.207 e2.087
<0.001
1.96
<0.001
1.00
1.459 e2.622 1.002 e1.005
Renal insufficiency p Value OR 95% CI
CVA/stroke p Value
OR
0.009
1.05
95% CI
Cardiac arrest p Value OR 95% CI
0.011 0.001
1.71 2.20
1.129e2.587 1.378e3.506
1.90
1.00
0.006
3.55
1.007 e1.007
1.315e5.165 1.079e4.696
0.006
1.98
1.221e3.208
0.008
2.08
1.212e3.552
0.015
1.00
1.001e1.005
0.001
2.15
1.348e3.427
0.006
2.19
1.252e3.834
0.027
1.69
1.061e2.685
1.446 e8.706
4.30
2.007e9.218
0.003
1.89
1.240e2.875
1.92
1.471 e2.510
0.003
1.163e 3.201
0.006
1.45
26.07
3.121e217.659
1.114 e1.897
ASA classification 4/5, contaminated/dirty wound, dependent functional status, preoperative history of CVA, preoperative history of insulin dependent DM, preoperative history of radiation therapy, preoperative history of weight loss >10%, preoperative history of PCI, preoperative open wound/wound infection, and preoperative transfusion >4 U packed RBC were included in the analysis and did not achieve significance. ASA, American Society of Anesthesiologists; DM, diabetes mellitus; HTN, hypertension; OR, odds ratio; PCI, percutaneous coronary intervention.
Analyzing Mortality after Lung Resection
<0.001
1.93
2.61 2.25
1.021e17.904
1.274e2.832 <0.001
0.011
4.28
Jean et al
0.002
0.006 0.031 0.016
Septic shock p Value OR 95% CI
1.013e1.091 0.047
Operative time Preoperative EtOH use >2 drinks/d Preoperative Hct <35% (female) or 40% (male) Preoperative history of COPD Preoperative history of dyspnea at rest Preoperative history of dyspnea on exertion Preoperative history of HTN Preoperative pneumonia Preoperative sepsis Preoperative sodium <135 mEq/L or >145 mEq/L
Reintubation p Value OR 95% CI
Vol. 222, No. 6, June 2016
Table 3. Significant Predictors of Complications for Least Absolute Shrinkage and Selection Operator Model
997
998
Jean et al
Analyzing Mortality after Lung Resection
Table 4. Shared Risk Factors Shared risk factors for predicting mortality Age in years Male sex Preoperative history of dyspnea at rest Preoperative history of dyspnea on exertion Preoperative sodium <135 mEq/L or >145 mEq/L Preoperative Hct <35% (female) or 40% (male) Open lobectomy Shared risk factors for predicting reintubation Age in years Current smoker Preoperative history of COPD Open lobectomy Operative time Shared risk factors for predicting renal insufficiency Variable Preoperative Hct <35% (female) or 40% (male) Operative time Shared risk factors for predicting CVA/stroke Age in years Shared risk factors for predicting cardiac arrest requiring CPR Male sex ASA classification 3 Open lobectomy Shared risk factors for predicting septic shock Current smoker Preoperative history of COPD Preoperative history of HTN ASA, American Society of Anesthesiologists; HTN, hypertension.
patient’s health status are at play and are most likely responsible for any variability. In our study, only 3 risk factors, dyspnea at rest, dyspnea on exertion, and dysnatremia, were associated with mortality without predicting for a high-risk complication. Given that our study is limited to the exploration of associations rather than definitive causality, we are unable to definitively state why these 3 risk factors were categorized as such. It is notable that these risk factors are associated strongly with severe derangements in multiple organ systems, including respiratory, metabolic, cardiovascular, renal, and neurologic processes. It is possible that these risk factors are proxies for overall severe critical illness rather than simply predictors for a subsequent complication. Therefore, the clinical implications of these results have a complex relationship with preoperative care for patients undergoing lobectomies for cancer. Although dysnatremia and dyspnea should not be taken to be absolute contraindications to surgical therapy, these results allow clinicians to view these risk factors as representative of
J Am Coll Surg
an increased perioperative risk. Therefore, physicians might be warranted in taking a guarded approach to patients with risk factors highly predictive of mortality and complications. It is notable that although there is good agreement between the standard multivariate and LASSO-selected models, the primary reason for mismatch between risk factors predicting mortality versus complications might be due to the observation of associations rather than causative relationships; a phenomenon seen in many retrospective database studies. A secondary reason for any disconnect might be due to the fact that advances in the processes of care might be able to atone for the occurrence of complications. Sukumar and colleagues23 found that although complications after oncologic operations have increased during the past decade, this increase was met with a concomitant decrease in failure to rescue rates and mortality, possibly due to improved recognition and management. Similarly, Downey and colleagues24 examined the trends of several complications and surgical quality indicators and demonstrated that several indicators, including postoperative pulmonary embolism, postoperative sepsis, and postoperative respiratory failure, have increased in reported incidence, despite the fact that in-hospital mortality after major postoperative complication (failure to rescue) had shown a significant decreasing trend. Additional study into detecting and appropriately managing the development of complications might mitigate against its consequences on overall mortality. Ultimately, it might be that dealing with complications can be just as important as assessing upfront risk. There are a number of limitations to this study. First and foremost, data used in study were from a national registry and, therefore, are associated with the limitations associated with the use of the NSQIP database. As a result, our analysis is limited to a retrospective data analysis, and might not represent direct causality as opposed to correlations. In addition, all assessments of mortality are limited to 30 days postoperatively. As such, the full impact of postoperative comorbidity after this period could not be studied directly. Another important limitation, related to this issue, is the fact that the NSQIP is not a dedicated thoracic surgery database and, therefore, there might be risk of bias in detecting significant predictors for primary lung resections. Also, because many disease and treatment patterns are associated with particular hospitals or regions, a hierarchical regression model was considered for this analysis; however, the NSQIP does not provide hospital-specific identifiers and we were unable to do so. Along this same vein, a significant limitation is that we were not able to perform external
Vol. 222, No. 6, June 2016
Jean et al
Analyzing Mortality after Lung Resection
999
Figure 1. Relationship among shared risk factors, complications, and mortality. Risk factors are listed on the left, with arrows to the left identifying the significant correlation to mortality. Arrows to the right identify a significant correlation of a given risk factor with postoperative complications. Those risk factors that only associate with complications are classified in scenario 1. Those that associate to both complications and mortality are classified in scenario 2. Those that only associate with mortality are classified as scenario 3. ASA, American Society of Anesthesiologists.
validation of our data using another validated data source. We did perform an internal validation by dividing the full set of data into validation sets of equal size, and the comparison of the multivariate logistic regression model and the logistic regression with LASSO selection were found to identify fewer significant risk factors (data not shown). These results were comparable with those of the whole sample. As such, however, there does exist significant potential for using this method with other datasets, or combined datasets, such as the Society of Thoracic Surgeons data. Additional research is warranted in the continuation of this analysis going forward, however, the nature of these results is such that it merits consideration of how thoracic surgeons can unify the reporting of data to best allow for retrospective analysis.
CONCLUSIONS Ultimately, this analysis demonstrates that the relationship between surgical risk factors, complications, and postoperative mortality is complex, and additional investigation into the mechanisms of accrual of mortality risk in the perioperative period is warranted.
Author Contributions Study conception and design: Jean, DeLuzio, Kraev, G Wang, Boffa, Detterbeck, Z Wang, Kim Acquisition of data: Kim Analysis and interpretation of data: Jean, DeLuzio, Kraev, G Wang, Z Wang, Kim Drafting of manuscript: Jean, DeLuzio, Kraev, G Wang, Boffa, Detterbeck, Z Wang, Kim Critical revision: Jean, DeLuzio, Kraev, G Wang, Boffa, Detterbeck, Z Wang, Kim REFERENCES 1. Pierce RJ, Copland JM, Sharpe K, Barter CE. Preoperative risk evaluation for lung cancer resection: predicted postoperative product as a predictor of surgical mortality. Am J Respir Crit Care Med 1994;150:947e955. 2. Khuri SF, Daley J, Henderson W, et al. Risk adjustment of the postoperative mortality rate for the comparative assessment of the quality of surgical care: results of the National Veterans Affairs Surgical Risk Study. J Am Coll Surg 1997;185:315e327. 3. Gibbs J, Cull W, Henderson W, et al. Preoperative serum albumin level as a predictor of operative mortality and morbidity: results from the National VA Surgical Risk Study. Arch Surg 1999;134:36e42.
1000
Jean et al
Analyzing Mortality after Lung Resection
4. Harpole DH, DeCamp MM, Daley J, et al. Prognostic models of thirty-day mortality and morbidity after major pulmonary resection. J Thorac Cardiovasc Surg 1999;117: 969e979. 5. Fink AS, Hutter MM, Campbell DC, et al. Comparison of risk-adjusted 30-day postoperative mortality and morbidity in Department of Veterans Affairs hospitals and selected university medical centers: general surgical operations in women. J Am Coll Surg 2007;204: 1127e1136. 6. Henderson WG, Khuri SF, Mosca C, et al. Comparison of risk-adjusted 30-day postoperative mortality and morbidity in Department of Veterans Affairs hospitals and selected university medical centers: general surgical operations in men. J Am Coll Surg 2007;204:1103e1114. 7. Kozower BD, Sheng S, O’Brien SM, et al. STS database risk models: predictors of mortality and major morbidity for lung cancer resection. Ann Thorac Surg 2010;90:875e881; discussion 881883. 8. Bernard A, Rivera C, Pages PB, et al. Risk model of inhospital mortality after pulmonary resection for cancer: a national database of the French Society of Thoracic and Cardiovascular Surgery (Epithor). J Thorac Cardiovasc Surg 2011;141:449e458. 9. Poullis M, McShane J, Shaw M, et al. Prediction of in-hospital mortality following pulmonary resections: improving on current risk models. Eur J Cardiothorac Surg 2013;44: 238e242; discussion 242243. 10. Watanabe S, Asamura H, Suzuki K, Tsuchiya R. Recent results of postoperative mortality for surgical resections in lung cancer. Ann Thorac Surg 2004;78:999e1002. 11. American College of Surgeons. About ACS NSQIP. Available at: http://site.acsnsqip.org/about/. Accessed November 12, 2013. 12. American College of Surgeons. User guide for the 2012 ACS NSQIP Participant Use Data File. https://www. facs.org/w/media/files/quality%20programs/nsqip/ug12.ashx. Accessed November 14, 2013.
J Am Coll Surg
13. Tibshirani R. Regression shrinkage and selection via the lasso. J R Stat Soc Ser B 1996;58:267e288. 14. Damhuis RA, Schu¨tte PR. Resection rates and postoperative mortality in 7,899 patients with lung cancer. Eur Respir J 1996;9:7e10. 15. Wyser C, Stulz P, Sole`r M, et al. Prospective evaluation of an algorithm for the functional assessment of lung resection candidates. Am J Respir Crit Care Med 1999;159:1450e1456. 16. McGuire HH. Measuring and managing quality of surgery. Arch Surg 1992;127:733. 17. Nagasaki F. Complications of surgery in the treatment of carcinoma of the lung. Chest J 1982;82:25. 18. Allen MS, Darling GE, Pechet TTV, et al. Morbidity and mortality of major pulmonary resections in patients with early-stage lung cancer: initial results of the randomized, prospective ACOSOG Z0030 trial. Ann Thorac Surg 2006;81: 1013e1020. 19. Bernard A, Ferrand L, Hagry O, et al. Identification of prognostic factors determining risk groups for lung resection. Ann Thorac Surg 2000;70:1161e1167. 20. Hamel MB, Henderson WG, Khuri SF, Daley J. Surgical outcomes for patients aged 80 and older: morbidity and mortality from major noncardiac surgery. J Am Geriatr Soc 2005;53: 424e429. 21. Alam N, Park BJ, Wilton A, et al. Incidence and risk factors for lung injury after lung cancer resection. Ann Thorac Surg 2007;84:1085e1091; discussion 1091. 22. Linden PA, Yeap BY, Chang MY, et al. Morbidity of lung resection after prior lobectomy: results from the Veterans Affairs National Surgical Quality Improvement Program. Ann Thorac Surg 2007;83:425e431; discussion 432. 23. Sukumar S, Roghmann F, Trinh VQ, et al. National trends in hospital-acquired preventable adverse events after major cancer surgery in the USA. BMJ Open 2013;3:e002843. 24. Downey JR, Hernandez-Boussard T, Banka G, Morton JM. Is patient safety improving? National trends in patient safety indicators: 1998-2007. Health Serv Res 2012;47: 414e430.
Vol. 222, No. 6, June 2016
Jean et al
Appendix 1. CPT and ICD-9 Codes Used in Selection of the Study Cohort CPT codes: open and thoracoscopic anatomical lung resection
32440 32445 32484 32442 32480 32663 ICD-9 codes to identify anatomic resections for oncologic pathology 031.0 170.3 492 114.3 171.4 492.0 114.4 171.9 492.8 114.5 193 494.0 117.3 195.1 494.1 117.7 197 510.0 150.8 197.0 510.9 151.0 209 512.0 158.9 209.21 512.8 162 209.6 513.0 162.2 209.61 515 162.3 212.3 518 162.4 212.4 518.89 162.5 212.6 747.3 162.8 235.7 748.5 162.9 235.8 748.8 163 239.1 759.6 163.9 417.8 786.6 164 446.4 793.1 164.3 448.0 844.9 164.9 485 V10.42
Analyzing Mortality after Lung Resection
1000.e1