Postop Complications and Survival in Colorectal Cancer
Postop Complications and Survival in Colorectal Cancer
To examine the association of postoperative complications, we acquired and linked 2 major Veterans Affairs (VA) databases, the VA Surgical Quality Improvement Program (VASQIP) database and the VA Central Cancer Registry (VACCR) database. To our knowledge, this is the first time that 2 large, systemwide databases of this type have been linked to examine our study questions.
The VASQIP database was originally developed in 1991 to measure and improve the quality of surgical care in the VA hospital system. The database is a collection of risk-adjusted immediate postoperative outcome data collected for thousands of surgical procedures from 123 VA hospitals nationwide. VASQIP collects detailed preoperative risk data—including patient demographics, comorbidities, and preoperative laboratory values—and detailed outcome data for the 30-day postoperative period such as mortality, morbidity across 21 complication groups, and readmissions. A dedicated nurse reviewer collects and abstracts data at each facility, assuring accuracy and completeness.
The VACCR collects diagnostic and treatment data for all malignancies treated in the VA health care system using standardized reporting protocols. These data include diagnoses, histology codes, clinical and pathologic stage, operative data (primarily using Facility Oncology Registry Data Standards codes), other treatment data such as receipt of chemotherapy and radiation, and outcome data such as long-term survival.
The study population consisted of all patients (1) 18 years of age and older, (2) with a primary diagnosis of histologically confirmed, invasive, nonmetastatic (stages 1–3) CRC between January 1, 1999, and December 31, 2009, and (3) who underwent radical curative-intent colon or rectal resection. Patients with stage 0 (in situ) disease and those undergoing endoscopic or transanal surgery were excluded.
Patients meeting the study inclusion criteria were identified from the VACCR database using International Classification of Diseases for Oncology (ICD-O-3) histology codes, site codes (for colon and rectum), International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) diagnosis codes, and Current Procedural Terminology (CPT) and Facility Oncology Registry Data Standards (surgery) codes. Using scrambled social security numbers, the VACCR and VASQIP were merged into a master analytic dataset containing demographic, preoperative, and immediate postoperative data collected by VASQIP and cancer-specific data collected by the VACCR.
To include only the primary curative-intent operations as study cases, patients with more than 1 procedure were excluded. In this manner, reoperations were not counted as unique cases but were captured as complications of the primary case. To ensure fidelity, the linkage was validated by including only those cases that demonstrated conformity of the operative date for each surgery from each database to within 24 hours. More than 93% (14,030) of cases met the conformity requirements, and these cases comprised the initial linked dataset. A total of 12,075 met the study inclusion criteria and comprised the study data set.
The primary outcome of interest was overall survival. Survival time was defined as date of surgery to date of death, with patients alive at last follow-up censored on the date of last contact. To minimize the effect of early deaths that resulted from postoperative complications, deaths within 90 days of operation were excluded from survival analysis. The primary variable of interest for the first part of the analysis was the presence of any postoperative complication. Other variables studied included (1) demographic variables (age and sex); (2) additional patient- and disease-specific variables such as American Joint Committee on Cancer cancer stage, site of disease (colon vs rectum), comorbidity [classified according to the American Society of Anesthesiologists (ASA) classification], preoperative nutritional status (preoperative serum albumin), and preoperative functional status (VASQIP-defined categorized variable); and (3) treatment-specific variables including type of surgery, number of lymph nodes examined (a surrogate for quality of cancer operation), intraoperative blood transfusions, chemotherapy (none, preoperative, postoperative), and radiotherapy (none, preoperative, postoperative).
A descriptive analysis of the entire sample population was performed. The study population was then categorized into 2 study groups on the basis of the presence or absence of any postoperative complication. A univariate comparison of patient, disease, and treatment-specific factors was performed by study group. Categorical variables were compared using χ test and continuous variables were compared using the student t test for dichotomous comparison groups or 1-way analysis of variance for multiple comparison groups (using the normality assumption of the central limit theorem).
Long-term survival for the entire group was estimated using the Kaplan-Meier method and compared by study group using the log-rank test. The analysis was stratified by cancer stage and site of disease, and pairwise comparisons for all individual strata were performed. Using univariate and multivariate Cox regression analysis, the association of any complication with long-term outcome was determined, adjusting for the hypothesis-based confounding factors described previously.
Details of the specific type of surgery—such as partial versus total colectomy/proctectomy, laparoscopic versus open surgery, and presence or absence of colostomy—were included in the descriptive analysis but omitted from the multivariate analysis, given that type of surgery is associated with the main risk factor of interest (postoperative complications), but should not be associated with long-term survival independent of other risk factors already included in the multivariate model (eg, site of disease). With respect to the ASA variable, due to the small number of patients in groups 1 and 5, these patients were combined with groups 2 and 4, respectively, for multivariate analysis to simplify clinical interpretation and to increase statistical degrees of freedom. Similarly, for the functional status variable, the relatively small completely dependent group was combined with the partially dependent group during multivariate analysis.
Multivariate modeling was performed and presented in hypothesis-based fashion, first including only a univariate analysis (model 1), followed by inclusion of patient and disease factors (model 2), and finally after including patient, disease, and treatment factors (model 3-maximal model) to determine the effect of these variables on the complication-survival association.
To determine potential effect modification by cancer stage and site of disease, interaction analysis was performed by adding the respective categorical variable product terms (complication × cancer stage and complication × site of disease) individually to the maximal model (model 3).
After this initial analysis, the study population was then stratified into 3 groups (no complication, noninfectious complications, and any infectious complication). Infectious complications were defined as pneumonia, urinary tract infection, or any SSI (superficial, deep, and/or organ space). A univariate comparison of factors by study group (trichotomized) was repeated. Long-term survival by study group was then compared using the log-rank test with pairwise comparisons. Using multivariate Cox regression analysis, the association of complication type with long-term survival was determined, again adjusting for all hypothesis-based confounding factors in the manner described previously. Interaction analysis was performed as mentioned previously by including the categorical variable product terms (complication type × cancer stage and complication type × site of disease).
To determine whether the severity of the infectious complication influenced outcome, a subset analysis was performed within the postoperative complication cohort. The databases used in the study did not contain enough data to reliably calculate validated complication severity scores such as the Clavien-Dindo classification. Instead, complications were subcategorized by severity of infection as follows. Severe surgical infections were defined as deep SSI, organ-space SSI, and/or systemic sepsis. Nonsevere infections included urinary tract infections, wound dehiscence, superficial SSIs, and pneumonia, all without systemic sepsis. Long-term survival was compared by subgroup—noninfectious complication(s) versus nonsevere infectious complication(s) versus any severe infectious complication—using univariate Kaplan-Meier survival and multivariate Cox regression analysis controlling for all relevant confounders.
The results of all comparative and survival analyses are reported with the appropriate summary statistic(s) and measures of statistical significance. P values of less than 0.05 were considered statistically significant. With respect to missing data, 11% of cases had missing values for cancer stage and less than 1% of values were missing for the remainder of the variables. These were assumed to be missing at random, and cases were omitted for individual comparative and survival analyses only when missing values were encountered. The analysis was performed using IBM SPSS Statistics for Windows, Version 22.0 (Armonk, NY: IBM Corp, 2013).
Methods
Databases
To examine the association of postoperative complications, we acquired and linked 2 major Veterans Affairs (VA) databases, the VA Surgical Quality Improvement Program (VASQIP) database and the VA Central Cancer Registry (VACCR) database. To our knowledge, this is the first time that 2 large, systemwide databases of this type have been linked to examine our study questions.
The VASQIP database was originally developed in 1991 to measure and improve the quality of surgical care in the VA hospital system. The database is a collection of risk-adjusted immediate postoperative outcome data collected for thousands of surgical procedures from 123 VA hospitals nationwide. VASQIP collects detailed preoperative risk data—including patient demographics, comorbidities, and preoperative laboratory values—and detailed outcome data for the 30-day postoperative period such as mortality, morbidity across 21 complication groups, and readmissions. A dedicated nurse reviewer collects and abstracts data at each facility, assuring accuracy and completeness.
The VACCR collects diagnostic and treatment data for all malignancies treated in the VA health care system using standardized reporting protocols. These data include diagnoses, histology codes, clinical and pathologic stage, operative data (primarily using Facility Oncology Registry Data Standards codes), other treatment data such as receipt of chemotherapy and radiation, and outcome data such as long-term survival.
Sample Population and Database Linkage
The study population consisted of all patients (1) 18 years of age and older, (2) with a primary diagnosis of histologically confirmed, invasive, nonmetastatic (stages 1–3) CRC between January 1, 1999, and December 31, 2009, and (3) who underwent radical curative-intent colon or rectal resection. Patients with stage 0 (in situ) disease and those undergoing endoscopic or transanal surgery were excluded.
Patients meeting the study inclusion criteria were identified from the VACCR database using International Classification of Diseases for Oncology (ICD-O-3) histology codes, site codes (for colon and rectum), International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) diagnosis codes, and Current Procedural Terminology (CPT) and Facility Oncology Registry Data Standards (surgery) codes. Using scrambled social security numbers, the VACCR and VASQIP were merged into a master analytic dataset containing demographic, preoperative, and immediate postoperative data collected by VASQIP and cancer-specific data collected by the VACCR.
To include only the primary curative-intent operations as study cases, patients with more than 1 procedure were excluded. In this manner, reoperations were not counted as unique cases but were captured as complications of the primary case. To ensure fidelity, the linkage was validated by including only those cases that demonstrated conformity of the operative date for each surgery from each database to within 24 hours. More than 93% (14,030) of cases met the conformity requirements, and these cases comprised the initial linked dataset. A total of 12,075 met the study inclusion criteria and comprised the study data set.
Statistical Analysis
The primary outcome of interest was overall survival. Survival time was defined as date of surgery to date of death, with patients alive at last follow-up censored on the date of last contact. To minimize the effect of early deaths that resulted from postoperative complications, deaths within 90 days of operation were excluded from survival analysis. The primary variable of interest for the first part of the analysis was the presence of any postoperative complication. Other variables studied included (1) demographic variables (age and sex); (2) additional patient- and disease-specific variables such as American Joint Committee on Cancer cancer stage, site of disease (colon vs rectum), comorbidity [classified according to the American Society of Anesthesiologists (ASA) classification], preoperative nutritional status (preoperative serum albumin), and preoperative functional status (VASQIP-defined categorized variable); and (3) treatment-specific variables including type of surgery, number of lymph nodes examined (a surrogate for quality of cancer operation), intraoperative blood transfusions, chemotherapy (none, preoperative, postoperative), and radiotherapy (none, preoperative, postoperative).
A descriptive analysis of the entire sample population was performed. The study population was then categorized into 2 study groups on the basis of the presence or absence of any postoperative complication. A univariate comparison of patient, disease, and treatment-specific factors was performed by study group. Categorical variables were compared using χ test and continuous variables were compared using the student t test for dichotomous comparison groups or 1-way analysis of variance for multiple comparison groups (using the normality assumption of the central limit theorem).
Survival Analysis: Any Versus No Postoperative Complications
Long-term survival for the entire group was estimated using the Kaplan-Meier method and compared by study group using the log-rank test. The analysis was stratified by cancer stage and site of disease, and pairwise comparisons for all individual strata were performed. Using univariate and multivariate Cox regression analysis, the association of any complication with long-term outcome was determined, adjusting for the hypothesis-based confounding factors described previously.
Details of the specific type of surgery—such as partial versus total colectomy/proctectomy, laparoscopic versus open surgery, and presence or absence of colostomy—were included in the descriptive analysis but omitted from the multivariate analysis, given that type of surgery is associated with the main risk factor of interest (postoperative complications), but should not be associated with long-term survival independent of other risk factors already included in the multivariate model (eg, site of disease). With respect to the ASA variable, due to the small number of patients in groups 1 and 5, these patients were combined with groups 2 and 4, respectively, for multivariate analysis to simplify clinical interpretation and to increase statistical degrees of freedom. Similarly, for the functional status variable, the relatively small completely dependent group was combined with the partially dependent group during multivariate analysis.
Multivariate modeling was performed and presented in hypothesis-based fashion, first including only a univariate analysis (model 1), followed by inclusion of patient and disease factors (model 2), and finally after including patient, disease, and treatment factors (model 3-maximal model) to determine the effect of these variables on the complication-survival association.
To determine potential effect modification by cancer stage and site of disease, interaction analysis was performed by adding the respective categorical variable product terms (complication × cancer stage and complication × site of disease) individually to the maximal model (model 3).
Survival Analysis: Any Infectious Versus Noninfectious Versus No Complications
After this initial analysis, the study population was then stratified into 3 groups (no complication, noninfectious complications, and any infectious complication). Infectious complications were defined as pneumonia, urinary tract infection, or any SSI (superficial, deep, and/or organ space). A univariate comparison of factors by study group (trichotomized) was repeated. Long-term survival by study group was then compared using the log-rank test with pairwise comparisons. Using multivariate Cox regression analysis, the association of complication type with long-term survival was determined, again adjusting for all hypothesis-based confounding factors in the manner described previously. Interaction analysis was performed as mentioned previously by including the categorical variable product terms (complication type × cancer stage and complication type × site of disease).
Subset Analysis: Effect of Severity of Infectious Complication
To determine whether the severity of the infectious complication influenced outcome, a subset analysis was performed within the postoperative complication cohort. The databases used in the study did not contain enough data to reliably calculate validated complication severity scores such as the Clavien-Dindo classification. Instead, complications were subcategorized by severity of infection as follows. Severe surgical infections were defined as deep SSI, organ-space SSI, and/or systemic sepsis. Nonsevere infections included urinary tract infections, wound dehiscence, superficial SSIs, and pneumonia, all without systemic sepsis. Long-term survival was compared by subgroup—noninfectious complication(s) versus nonsevere infectious complication(s) versus any severe infectious complication—using univariate Kaplan-Meier survival and multivariate Cox regression analysis controlling for all relevant confounders.
The results of all comparative and survival analyses are reported with the appropriate summary statistic(s) and measures of statistical significance. P values of less than 0.05 were considered statistically significant. With respect to missing data, 11% of cases had missing values for cancer stage and less than 1% of values were missing for the remainder of the variables. These were assumed to be missing at random, and cases were omitted for individual comparative and survival analyses only when missing values were encountered. The analysis was performed using IBM SPSS Statistics for Windows, Version 22.0 (Armonk, NY: IBM Corp, 2013).