Abstract
Background
The impact of the stage of cancer on perioperative mortality remains obscure. The purpose of this study was to investigate whether cancer stage influences 30-day mortality for gastric, pancreatic, and colorectal cancers.
Methods
Data were collected from the National Cancer Database (NCDB) for patients undergoing resections for cancers of the stomach, pancreas, colon or rectum between 2004 and 2015. The main analysis was conducted among patients with cancers stages 1–3. A sensitivity analysis also included cancer stage 4. Descriptive statistics were used to compare the patients’ baseline characteristics. Generalized linear mixed models were used to evaluate the relationship between stage and 30-day mortality, controlling for other disease-, patient- and hospital-level factors. Pseudo-R2 statistics (%Δ pseudo R2) were used to quantify the relative explanatory capacity of the variables to the model for 30-day mortality. All analyses were performed using SAS 9.4 (Cary, NC).
Results
The cohort included 24,468, 28,078, 176,285, and 64,947 patients with stomach, pancreas, colon and rectal cancer, respectively. After adjusting for other variables, 30-day mortality was different by stage for all cancer types examined. The factor most strongly associated with 30-day mortality was age (%Δ pseudo R2 range 14–39%). The prognostic impact of cancer stage (stage 1, 2 or 3) on 30-day mortality was comparable to that of the Charlson comorbidity index.
Conclusions
Cancer stage contributes to explaining differences observed in short-term mortality for GI cancers. Short-term mortality models would benefit by including more granular cancer stage, beyond disseminated status alone.
Keywords: Cancer, Outcomes, Modeling, Stage, Prediction, Survival
Introduction
Gastrointestinal cancers are among the leading causes of cancer-related death in the United States, but can often be cured by surgical resection with good short- and long-term outcomes.1 Short-term outcomes, such as perioperative mortality, morbidity and complications, are influenced by numerous factors. Important patient-related factors include age, sex, comorbidities, and frailty.2 Intraoperative factors known to affect outcomes include surgery duration and blood loss.3 Postoperative issues such as infection, cardiopulmonary complications, and leaks can all influence outcomes.4 In the context of cancer surgery, stage and histologic characteristics are additional potentially prognostic factors.
In general, studies of perioperative outcomes consider only whether a malignancy is disseminated, rather than including stage as a variable. Disseminated disease has been shown to be associated with increased risk of perioperative mortality in cancers of the colon,5 pancreas6 and lung, and is generally understood to entail higher risk for all cancers.7 However, distinctions among earlier stages of cancer have rarely been investigated as drivers of differences in perioperative outcomes.
The purpose of this study was to examine the relationship between stage and perioperative mortality for patients undergoing surgical resection for gastrointestinal (GI) cancers of the stomach, pancreas, colon and rectum cancers, particularly in the setting of non-disseminated disease. The goal was to determine whether stage should be included as a predictive variable in models of peri-operative mortality outcomes. We hypothesized that perioperative mortality is influenced by all cancer stages.
Methods
Data Source and Study Population
This study used the National Cancer Database (NCDB), a clinical oncology database sourced from hospital registry data from more than 1,500 Commission on Cancer (CoC)-accredited facilities. A program of the American College of Surgeons, the commission on Cancer, and the American Cancer Society, this deidentified database contains information on more than 70 percent of newly diagnosed cancer cases nationwide.8,9 The NCDB includes information about patient and cancer attributes as well as management and outcomes.
International Classification of Diseases for Oncology, 3rd edition (ICD-O-3) site codes were used to identify patients with carcinoma who underwent primary surgical resection (Tables 1 and 2 in appendix) for cancers of the stomach, pancreas, colon and rectum from January 1, 2004 to December 31, 2015. These cancers were chosen based on the cancer diagnoses commonly managed by general surgeons. Stomach cancers were limited to non-cardia stomach cancers as tumors of the gastric cardia are different in both their biologic nature and management. Only carcinomas were included in the analysis. Patients with missing survival or stage information were excluded, and the analysis was limited to patients who received at least part of the first course of treatment in the reporting facility (“class of case” 10–22). Demographic data included insurance type, sex, race/ethnicity, age, zip and income quartile. Biological variables such as cancer and surgery characteristics, receipt of chemotherapy, and comorbidities were extracted. Hospital size, region, and proximity to a metro county were extracted based on the site where the operation was performed. Hospital volume was calculated by making a hospital ID-year concatenated variable (each hospital-year is assessed separately) and counting the number of cases by primary organ for each, then calculating approximate quartiles.10 The outcome for comparison was 30-day mortality.
Stage of cancer was determined using the American Joint Committee on Cancer (AJCC) staging system and was evaluated as stage groups 1, 2, 3 and 4. Subjects with stage 4 disease were excluded from the main analysis, but included in sensitivity analyses. The stage used in the analyses reflected the AJCC edition at time of the surgery as this would best reflect operative considerations; standardization by edition was not performed. Clinical (preoperative) stage was used since 30-day mortality is often used in the pre-operative setting to discuss risks of surgery. Age was examined as <55, 56–65, 66–75 and >75. Race/ethnicity was grouped as White non-Hispanic, Black non-Hispanic, Hispanic, Asian/Pacific Islander and other; Hispanic ethnicity was allowed to dominate self-described race such that someone who was both Black and Hispanic would be classified as Hispanic. The NCDB provides a modified Charlson score based on reported comorbidities and International Classification of Disease, 9th Edition (ICD-9-CM) coding. The Charlson score assesses the severity of pre-existing comorbidities (the number of medical conditions weighted according to their relative effects on survival) and was grouped as 0, 1, 2 or 3 or more.11 Receipt of preoperative chemotherapy (with or without radiation) was included as a binary variable. Resections were grouped as local/segmental resections; subtotal or total resections plus hemicolectomies; extended resections; and not specified (see appendix for specific coding scheme). Whether an additional organ had to be resected en bloc was included as a separate binary variable.
Hospital region was examined as South, West, Midwest, Northeast and other. The NCDB has an “urban/rural” variable that includes nine classifications based on the county population and metro county proximity. These classifications were collapsed to reflect only proximity to metro area: metro area, adjacent to metro area, or not adjacent to metro area.
Statistical Analysis
Proportions of all variables by cancer diagnosis were calculated. 30-day and 90-day post-operative mortality rates were also calculated per stage and cancer diagnosis. Generalized Linear Mixed (GLM) modeling was used to evaluate the impact of the identified factors on 30-day mortality. Contributions of each variable were examined in terms of their effect on Nagelkerke / Cragg & Uhler pseudo-R2.12 The main analyses excluded patients with stage 4 disease, to illuminate the significance of stage 1–3 in explaining differences in 30-day mortality. A model was also created to evaluate the significance of stage 4 disease alone, as this is the variable already considered in perioperative risk calculators. A strong effect of hospital volume on mortality was observed in our analysis, and we further investigated the effect of hospital volume on the relationship between stage and mortality. Models were created for cohorts stratified by volume quartile, then odds ratios for stage were compared. A mediation analysis was conducted to directly assess the effect of volume on the relationship between stage and outcomes, and vice versa. Sensitivity analyses for internal validity were performed for (1) inclusion of stage 4 disease, (2) use of 90-day mortality as the outcome, (3) use of random effects to control for clustering by facility and (4) inclusion of a variable to identify patients whose diagnosis was on the same day as their surgery (an indication that the operation was performed emergently). All analyses were performed using SAS 9.4 (Cary, NC). This study was approved by the Montefiore Medical Center/Albert Einstein School of Medicine Institutional Review Board.
Results
Analyses included 293,778 GI cancer patients, stages 1–3, who met inclusion criteria (24,468 stomach, 28,078 pancreas, 176,285 colon, 64,947 rectum). An additional 53,536 patients had stage 4 disease (2,792 stomach, 1,403 pancreas, 42,701 colon, and 6,640 rectum). Table 1 describes the patient population by organ site. Most patients were healthy (Charlson score 0; 69.3%), white (79.4%), had Medicare (53.5%) or private (36.3%) insurance, and were treated in a metropolitan area (84.3%). Figure 1 shows the 30-day and 90-day post-operative mortality rates by organ and stage. Incremental increases in mortality were observed by stage.
Table 1:
Descriptive Statistics
| Primary Site | N | Percent | |||||
|---|---|---|---|---|---|---|---|
| Stomach | Pancreas | Colon | Rectum | ||||
| Totals | N | 24468 | 28078 | 176285 | 64947 | 293778 | |
| % | 8.3% | 9.6% | 60.0% | 22.1% | 100.0% | ||
| Stage | 1 | 42.0% | 43.7% | 47.0% | 27.5% | 123210 | 41.9% |
| 2 | 30.4% | 51.0% | 31.9% | 35.3% | 100943 | 34.4% | |
| 3 | 27.6% | 5.4% | 21.1% | 37.2% | 69625 | 23.7% | |
| Operation Type | Local/Segmental Resection | 48.1% | 78.1% | 34.5% | 69.6% | 160909 | 47.6% |
| Hemi, Subtotal or Total | 15.2% | 12.6% | 62.8% | 24.9% | 162384 | 45.7% | |
| Extended Resection | 32.9% | 6.0% | 0.6% | 2.5% | 13435 | 4.2% | |
| Unknown or Not Specified | 3.8% | 3.3% | 2.1% | 3.1% | 10586 | 2.6% | |
| Additional Organ Resected | No | 90.3% | 23.3% | 93.5% | 100.0% | 304438 | 88.0% |
| Yes | 9.7% | 76.7% | 6.5% | N/A | 42876 | 12.0% | |
| Charlson Comorbidity Score | 0 | 66.8% | 64.1% | 67.8% | 76.5% | 242651 | 69.3% |
| 1 | 24.5% | 28.2% | 23.3% | 18.4% | 77861 | 22.8% | |
| 2 | 6.5% | 5.9% | 6.5% | 3.8% | 19758 | 5.8% | |
| 3 or more | 2.2% | 1.9% | 2.4% | 1.3% | 7044 | 2.1% | |
| Received Preoperative Chemotherapy (with or without Radiation) | No | 66.1% | 85.0% | 99.2% | 40.2% | 286169 | 82.1% |
| Yes | 33.9% | 15.0% | 0.8% | 59.8% | 61145 | 17.9% | |
| Sex | Male | 66.8% | 50.4% | 48.2% | 60.4% | 182584 | 52.7% |
| Female | 33.2% | 49.6% | 51.8% | 39.6% | 164730 | 47.4% | |
| Age | <56 | 20.5% | 17.2% | 17.0% | 32.5% | 75597 | 20.7% |
| 56–65 | 26.1% | 28.5% | 20.7% | 27.8% | 82601 | 23.5% | |
| 66–75 | 29.9% | 33.7% | 26.6% | 23.2% | 91501 | 26.8% | |
| >75 | 23.5% | 20.6% | 35.7% | 16.6% | 97615 | 29.0% | |
| Race/Ethnicity | Non-Hispanic White | 68.0% | 81.5% | 79.9% | 81.5% | 274003 | 79.4% |
| Non-Hispanic Black | 12.7% | 9.8% | 11.6% | 7.8% | 38650 | 10.7% | |
| Hispanic | 9.7% | 4.5% | 4.7% | 5.7% | 18697 | 5.3% | |
| Asian/Pacific Islanders | 7.8% | 2.6% | 2.5% | 3.4% | 10944 | 3.2% | |
| Other | 1.8% | 1.6% | 1.3% | 1.6% | 5020 | 1.4% | |
| Zip Income Quartile | Poorest Quartile | 17.9% | 15.6% | 18.1% | 17.2% | 62152 | 17.6% |
| 2nd Quartile | 22.2% | 22.6% | 23.8% | 24.1% | 82106 | 23.6% | |
| 3rd Quartile | 27.2% | 26.9% | 26.7% | 27.2% | 92786 | 26.9% | |
| Richest Quartile | 32.7% | 34.9% | 31.5% | 31.6% | 110270 | 31.9% | |
| Insurance Type | Private Insurance | 36.8% | 37.7% | 32.1% | 47.2% | 127632 | 36.3% |
| Medicaid | 6.4% | 4.5% | 3.9% | 5.8% | 17312 | 4.6% | |
| Medicare | 50.9% | 52.5% | 58.9% | 40.1% | 181644 | 53.5% | |
| Other Government | 1.0% | 1.0% | 0.7% | 0.9% | 2829 | 0.8% | |
| Uninsured/Unknown | 3.9% | 3.4% | 3.9% | 5.0% | 14137 | 4.1% | |
| Hospital Case Volume Quartile | Bottom | 27.0% | 25.1% | 29.0% | 29.3% | 39170 | 28.5% |
| 2nd Quartile | 27.4% | 24.8% | 23.2% | 26.0% | 52412 | 24.3% | |
| 3rd Quartile | 22.7% | 25.0% | 24.3% | 21.7% | 84646 | 23.7% | |
| Top | 22.9% | 25.0% | 23.6% | 23.0% | 86356 | 23.6% | |
| Metro Context | Metro County | 86.0% | 83.8% | 85.0% | 81.8% | 292700 | 84.3% |
| Adjacent to Metro | 9.8% | 11.3% | 10.4% | 12.2% | 37600 | 10.8% | |
| Not Adjacent | 4.2% | 4.9% | 4.6% | 6.0% | 17014 | 4.9% | |
| Region | Northeast | 24.5% | 21.6% | 20.3% | 18.9% | 70108 | 20.5% |
| Midwest | 23.0% | 25.8% | 25.4% | 27.4% | 89366 | 25.7% | |
| South | 33.7% | 37.6% | 37.4% | 33.7% | 126434 | 36.3% | |
| West | 16.2% | 13.5% | 15.0% | 16.0% | 52347 | 15.2% | |
| Unknown | 2.6% | 1.6% | 1.9% | 4.0% | 9059 | 2.4% | |
Figure 1:
Unadjusted 30-day and 90-Day Mortality by Stage
The GLM regression for each organ site examined the effect of cancer stage on 30-day mortality (Table 2). Increases in odds ratios for mortality were observed for each stage above stage 1 for both 30- and 90-day mortality (Figure 2). For example, compared to stage 1 colon cancer, the ORs for 30-day mortality for stages 2 and 3 were 1.68 (95% CI 1.58–1.79) and 1.75 (95% CI 1.63–1.88), respectively. Likewise, relationships were observed for other biological factors; men, older patients, and patients with more comorbidities had increased mortality rates. Public insurance and uninsured patients were more likely to die than privately insured patients. Hospital volume was also influential; lower-volume hospitals had higher odds of death.
Table 2:
Generalized Linear Mixed Modeling Odds Ratios of 30-day Mortality by Organ
| Stomach | Pancreas | Colon | Rectum | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Comparison | aOR | 95% CI | P value | aOR | 95% CI | P value | aOR | 95% CI | P value | aOR | 95% CI | P value | ||||
| Clinical Stage, vs Stage 1 | ||||||||||||||||
| Stage 2 | 1.35 | 1.14 | 1.59 | 0.0004 | 1.19 | 1.03 | 1.36 | 0.0155 | 1.68 | 1.58 | 1.79 | <.0001 | 1.57 | 1.30 | 1.88 | <.0001 |
| Stage 3 | 1.70 | 1.43 | 2.01 | <.0001 | 1.84 | 1.41 | 2.41 | <.0001 | 1.75 | 1.63 | 1.88 | <.0001 | 1.53 | 1.26 | 1.86 | <.0001 |
| Resection Type, vs Local/Segmental | ||||||||||||||||
| Hemi, subtotal or total | 1.47 | 1.22 | 1.78 | <.0001 | 1.66 | 1.40 | 1.98 | <.0001 | 1.06 | 1.00 | 1.13 | 0.0493 | 0.87 | 0.73 | 1.03 | 0.0969 |
| Extended | 1.22 | 1.03 | 1.44 | 0.0182 | 1.20 | 0.94 | 1.54 | 0.1427 | 1.40 | 0.94 | 2.10 | 0.0998 | 1.19 | 0.75 | 1.88 | 0.4559 |
| Unknown or not specified | 1.29 | 0.89 | 1.87 | 0.1748 | 1.65 | 1.10 | 2.47 | 0.0157 | 1.13 | 0.93 | 1.39 | 0.2189 | 1.10 | 0.74 | 1.65 | 0.6277 |
| Additional Organ Resected | ||||||||||||||||
| Yes vs no | 1.40 | 1.12 | 1.73 | 0.0025 | 2.08 | 1.70 | 2.55 | <.0001 | 1.44 | 1.31 | 1.59 | <.0001 | ||||
| Preop Systemic Therapy (with or without Radiation) | ||||||||||||||||
| Received vs did not | 1.51 | 1.26 | 1.81 | <.0001 | 1.26 | 1.01 | 1.57 | 0.0394 | 0.86 | 0.60 | 1.24 | 0.4215 | 1.75 | 1.49 | 2.06 | <.0001 |
| Charlson Score, vs 0 | ||||||||||||||||
| 1 | 1.11 | 0.95 | 1.30 | 0.1904 | 0.94 | 0.81 | 1.09 | 0.4171 | 1.33 | 1.24 | 1.42 | <.0001 | 1.53 | 1.29 | 1.80 | <.0001 |
| 2 | 1.75 | 1.40 | 2.18 | <.0001 | 1.48 | 1.17 | 1.87 | 0.0010 | 1.85 | 1.69 | 2.02 | <.0001 | 1.98 | 1.53 | 2.55 | <.0001 |
| 3 or more | 2.13 | 1.53 | 2.97 | <.0001 | 1.81 | 1.26 | 2.61 | 0.0014 | 2.63 | 2.33 | 2.97 | <.0001 | 3.26 | 2.31 | 4.61 | <.0001 |
| Sex, vs Female | ||||||||||||||||
| Male | 1.21 | 1.04 | 1.40 | 0.0122 | 1.08 | 0.95 | 1.23 | 0.2442 | 1.34 | 1.27 | 1.42 | <.0001 | 1.41 | 1.21 | 1.63 | <.0001 |
| Age, vs <55 | ||||||||||||||||
| 56–65 | 1.53 | 1.13 | 2.07 | 0.0063 | 1.51 | 1.15 | 1.97 | 0.0029 | 1.55 | 1.31 | 1.85 | <.0001 | 2.08 | 1.52 | 2.84 | <.0001 |
| 66–75 | 2.74 | 2.00 | 3.74 | <.0001 | 2.09 | 1.56 | 2.79 | <.0001 | 2.47 | 2.08 | 2.93 | <.0001 | 2.53 | 1.79 | 3.58 | <.0001 |
| >75 | 4.88 | 3.57 | 6.69 | <.0001 | 3.13 | 2.32 | 4.22 | <.0001 | 6.77 | 5.74 | 7.99 | <.0001 | 6.75 | 4.81 | 9.45 | <.0001 |
| Race, vs White | ||||||||||||||||
| Black | 0.76 | 0.60 | 0.95 | 0.0151 | 0.74 | 0.58 | 0.95 | 0.0186 | 1.01 | 0.91 | 1.11 | 0.8605 | 1.50 | 1.16 | 1.92 | 0.0017 |
| Hispanic | 0.74 | 0.57 | 0.97 | 0.0276 | 0.97 | 0.71 | 1.33 | 0.8532 | 0.66 | 0.55 | 0.78 | <.0001 | 1.01 | 0.72 | 1.43 | 0.9517 |
| API | 0.61 | 0.44 | 0.86 | 0.0041 | 0.98 | 0.64 | 1.51 | 0.9277 | 0.68 | 0.54 | 0.86 | 0.0014 | 0.76 | 0.45 | 1.28 | 0.2937 |
| Other | 0.85 | 0.47 | 1.53 | 0.5912 | 1.04 | 0.60 | 1.79 | 0.8989 | 1.19 | 0.93 | 1.52 | 0.1771 | 0.74 | 0.36 | 1.49 | 0.3980 |
| Zip Income Quartile, vs 25% Lowest Income Zip | ||||||||||||||||
| 2nd quartile | 1.21 | 0.97 | 1.51 | 0.0907 | 1.34 | 1.08 | 1.65 | 0.0074 | 1.24 | 1.13 | 1.35 | <.0001 | 1.14 | 0.90 | 1.44 | 0.2728 |
| 3rd quartile | 1.34 | 1.10 | 1.62 | 0.0036 | 1.17 | 0.97 | 1.41 | 0.1075 | 1.09 | 1.00 | 1.18 | 0.0387 | 1.05 | 0.85 | 1.29 | 0.6805 |
| Top quartile | 1.17 | 0.97 | 1.41 | 0.0947 | 1.05 | 0.88 | 1.26 | 0.5716 | 1.03 | 0.95 | 1.11 | 0.4610 | 1.13 | 0.93 | 1.37 | 0.2324 |
| Insurance Category, vs Private | ||||||||||||||||
| None | 1.49 | 0.93 | 2.41 | 0.1003 | 1.63 | 1.07 | 2.48 | 0.0218 | 1.83 | 1.45 | 2.31 | <.0001 | 1.31 | 0.76 | 2.24 | 0.3313 |
| Medicare | 1.45 | 1.03 | 2.03 | 0.0329 | 1.41 | 0.99 | 2.00 | 0.0568 | 2.00 | 1.67 | 2.39 | <.0001 | 1.78 | 1.21 | 2.64 | 0.0039 |
| Medicaid | 1.07 | 0.87 | 1.31 | 0.5171 | 1.27 | 1.04 | 1.55 | 0.0215 | 1.46 | 1.33 | 1.61 | <.0001 | 1.77 | 1.39 | 2.25 | <.0001 |
| Other Government | 1.00 | 0.46 | 2.18 | 0.9940 | 1.82 | 1.02 | 3.26 | 0.0435 | 0.83 | 0.52 | 1.32 | 0.4300 | 1.71 | 0.74 | 3.92 | 0.2067 |
| Unknown | 0.75 | 0.40 | 1.42 | 0.3820 | 1.34 | 0.79 | 2.26 | 0.2779 | 1.76 | 1.40 | 2.22 | <.0001 | 1.53 | 0.83 | 2.82 | 0.1727 |
| Yearly Hospital Volume, vs 26–50th Percentile | ||||||||||||||||
| 0–10th percentile | 1.13 | 0.93 | 1.38 | 0.2091 | 2.19 | 1.79 | 2.68 | <.0001 | 1.01 | 0.94 | 1.09 | 0.7845 | 1.37 | 1.12 | 1.67 | 0.0021 |
| 11–25th percentile | 1.03 | 0.85 | 1.25 | 0.7463 | 1.36 | 1.11 | 1.67 | 0.0033 | 1.00 | 0.92 | 1.08 | 0.9470 | 1.00 | 0.81 | 1.24 | 0.9949 |
| 51st-75th percentile | 0.81 | 0.64 | 1.03 | 0.0796 | 0.85 | 0.67 | 1.06 | 0.1523 | 0.96 | 0.88 | 1.04 | 0.3175 | 0.74 | 0.57 | 0.95 | 0.0199 |
| Metro County Proximity, vs Metro County | ||||||||||||||||
| Adjacent | 1.10 | 0.89 | 1.37 | 0.3842 | 1.04 | 0.84 | 1.28 | 0.7217 | 1.04 | 0.95 | 1.14 | 0.4257 | 0.97 | 0.77 | 1.22 | 0.8087 |
| Rural | 1.17 | 0.86 | 1.59 | 0.3300 | 0.91 | 0.67 | 1.23 | 0.5240 | 0.90 | 0.79 | 1.03 | 0.1372 | 1.22 | 0.91 | 1.63 | 0.1821 |
| Region, vs Northeast | ||||||||||||||||
| Midwest | 1.06 | 0.86 | 1.31 | 0.5801 | 1.00 | 0.81 | 1.22 | 0.9597 | 0.92 | 0.84 | 1.00 | 0.0435 | 0.76 | 0.62 | 0.95 | 0.0133 |
| South | 1.38 | 1.14 | 1.67 | 0.0010 | 1.21 | 1.00 | 1.46 | 0.0475 | 1.10 | 1.01 | 1.19 | 0.0233 | 0.97 | 0.79 | 1.19 | 0.7768 |
| West | 1.01 | 0.79 | 1.28 | 0.9388 | 0.82 | 0.64 | 1.05 | 0.1181 | 1.10 | 0.99 | 1.21 | 0.0648 | 0.72 | 0.56 | 0.93 | 0.0104 |
| Unclassified | 1.02 | 0.55 | 1.91 | 0.9446 | 0.83 | 0.36 | 1.95 | 0.6732 | 0.65 | 0.47 | 0.89 | 0.0083 | 0.35 | 0.13 | 0.97 | 0.0521 |
Figure 2:
Multivariable Regression: Adjusted Odds Ratios (aOR) for 30-Day and 90-Day Mortality by Cancer Stage, vs Stage 1. Estimates and 95% Confidence Intervals
Table 3 shows the degree of contribution (%Δ pseudo R2) of each variable to the explanatory power of the model. Stage (differences among stage 1, 2 and 3) explained 5.97%, 4.51%, 8.39% and 2.66% of the difference observed for 30-day mortality for stomach, pancreas, colon and rectum cancer, respectively. This is comparable to the contribution of Charlson score (%Δ pseudo R2 4–9%) to the model. Hospital volume was also a significant predictor of mortality risk, particularity for pancreatic cancer (%Δ pseudo R2 31.80%), but also for gastric (5.82%) and rectal cancer (5.61%). Although hospital volume was a predictor of mortality, there were no observed changes in the relationship between stage and mortality among different hospital volume quartiles. Stage did not mediate the relationship between volume and mortality. Volume did not mediate the relationship between stage and mortality except for stage 2 rectal cancer, in which 10% of the increased odds of perioperative mortality was attributable to hospital volume (tables are presented in the appendix).
Table 3:
30-day mortality Contributions of Variables to Pseudo R2
| Effect of Model Contributor on Pseudo R2 | |||||
|---|---|---|---|---|---|
| Contributor | Stomach | Pancreas | Colon | Rectum | |
| Clinical Factors | Clinical Stage (1–3)i | 5.97% | 4.51% | 8.39% | 2.66% |
| Surgery typeii | 4.64% | 17.36% | 1.44% | 0.43% | |
| Preop Therapy | 4.68% | 1.00% | 0.01% | 5.47% | |
| Patient Factors | Age | 27.87% | 13.96% | 39.20% | 24.96% |
| Sex | 1.08% | 0.30% | 2.81% | 2.37% | |
| Race/Ethnicity | 3.38% | 1.37% | 1.07% | 1.28% | |
| Charlson Score | 7.47% | 4.17% | 9.06% | 7.86% | |
| Socioeconomic Factors | Zip Income Quartile | 1.79% | 1.50% | 0.68% | 0.24% |
| Insurance | 1.62% | 2.33% | 2.84% | 3.15% | |
| Hospital Factors | Hospital Volume | 5.82% | 31.80% | 0.20% | 5.61% |
| Metro Context | 0.26% | 0.12% | 0.10% | 0.21% | |
| Region | 4.56% | 3.25% | 1.19% | 1.88% | |
Main model; stage examined as 3 levels (1, 2, 3). All non-stage contributions in table are based on model excluding stage 4.
Type of resection and inclusion of additional organ
Sensitivity analyses showed that, as expected, when stage 4 cancers were included, stage was able to account for a larger part of the variance than stage 1–3 alone, especially for colon cancer (%Δ pseudo R2 38.55% for stage, including 1–4). The contribution of stage at diagnosis to 90-day mortality was higher than that for 30-day mortality (%Δ pseudo R2 10.64%, 6.37%, 9.22% and 2.84% for stomach, pancreas, colon and rectum cancer, respectively). Neither controlling for hospital-level clustering via random effects nor including information about day of surgery changed the odds ratios associated with stage (tables for sensitivity analyses in appendix).
Discussion
Using the NCDB, a significant relationship was identified between cancer stage and perioperative mortality for gastrointestinal malignancies. This relationship persisted despite controlling for other relevant predictors of mortality. While it is generally known that having stage 4 disease at the time of surgery affects short-term outcomes, we found that stratification of risk by cancer stages 1–3 also contributes to explaining differences in 30-day mortality. The magnitude of stage’s contribution to the model was similar to that of comorbidities; distinctions among non-disseminated stages of cancer and Charlson comorbidity scores were similar in their ability to explain differences observed for short-term outcomes.
The causal mechanism connecting stage to perioperative outcomes is not clear. Part of the relationship may be attributable to slight differences in surgeries performed depending on stage. We attempted to adjust for this by including variables for the type of surgery performed and whether resection of an additional organ was needed, however more subtle details of the operation, such as difficulty in dissection or blood loss, were unavailable in the NCDB. The persistence of the effect of stage suggests that the physiologic burden of cancer may also have an effect. For example, pancreatic cancer cachexia has been found to increase with tumor size.13 Thrombotic risk has also been found to increase with stage across many cancers.14,15 Additional examination of nutritional factors or surgical complications could improve our understanding of these potentially contributory factors.
Based on our findings, we suggest that stage of cancer should be considered for inclusion in perioperative mortality risk models. Risk calculators such as the ACS NISQIP calculator,16 POSSUM,17 and the Gupta score18 do not take stage 1–3 into account and are consequently limited in their capacity to predict outcomes for these resections. Existing cancer-specific calculators such as MD Anderson’s pancreatic cancer calculator and Memorial Sloan Kettering’s nomograms for various cancer types include cancer stage, but are designed to evaluate long-term cancer-specific prognosis as opposed to short-term perioperative risk.19,20 Development of a cancer-specific database that could be used for perioperative risk calculation has been proposed by Bilimoria et al (“Onc-NSQIP”)21 but has yet to be developed. Because cancer stage information is not routinely captured in administrative databases, linkage of administrative databases to cancer registries should be encouraged to promote better capture of stage and other cancer-specific prognostic factors.
The only information about stage included in the most common perioperative outcome calculators is a binary variable for the presence of disseminated disease.13 However, the elevation in perioperative risk associated with stage 4 or disseminated disease is probably in part due to emergent and palliative indications for surgery. Hence, the binary variable for disseminated disease likely reflects only the effect of differing surgical indications on perioperative mortality,22 although isolated metastases of colorectal cancers are increasingly being resected with curative intent.23 Including the distinction among stages 1–3 will capture severity of cancer and improve our ability to compare outcomes for oncologic surgeries.
The potential significance of developing improved models of cancer surgery outcomes is multifaceted. In addition to improving accuracy of prognostic information provided to patients in the clinical context, improved models of perioperative mortality could assist with assessing hospital performance. Because hospitals care for populations with distinct baseline risk, improved models of perioperative risk can help hospitals interpret their own outcomes. Since there is significant variability in the outcomes, even among accredited cancer centers, it is important to supply models optimized for cancer surgery.21
If a new database or risk calculator is to be developed, it will be necessary to decide which cancer-specific variables should be included. Stage is significantly associated with mortality across the cancers we studied, but it seems to have the greatest predictive value for colon cancer, regardless of whether disseminated status is included. In contrast, for rectal cancer, the use of neoadjuvant therapy appears to play a bigger role than stage (%Δ pseudo R2 5.47). In addition to preoperative treatment regimens, cancer substages or histological characteristics may also have short-term prognostic significance for certain cancers.24–26 The fact that each cancer may come with unique prognostic associations could create some controversy about which factors merit inclusion in a database meant to apply to all cancer resections. Our findings establish that the AJCC’s four-level staging system provides prognostic value.
For this analysis, we chose to use clinical stage as opposed to pathologic stage. Clinical stage is determined through diagnostic tests and is used to guide the physician in making initial therapy decisions. As demonstrated by the addition of clinical stage prognostic tables to the esophagus and stomach chapters in the recent publication of the AJCC 8th edition staging manual,27 prognosis differs for clinical stage compared to pathologic stage. Since surgical decisions are made using clinical stage information, we felt it more appropriate to use clinical stage exclusive of information obtained from pathologic examination of the surgical specimen.
Aside from patient factors, hospital and provider-level factors are known to be particularly significant predictors of outcomes in cancer surgery.28 The data reported here for pancreatic cancer show that hospital surgical volume had a far greater predictive value than stage or any other factor in the model. Controlling for clustering by hospital factors with random effects modeling instead of discrete hospital variables affected the pseudo R2 but did not alter the contribution of stage (appendix table 2). This suggests that there are additional measures of hospital quality that affect outcomes beyond the factors included in the analysis. The optimal means of incorporating this information into predictive models, and into discussions with patients, remains to be seen.
This study has limitations. Many of the variables in the model interact, and we make no claim that the given models are fully optimized for prognostic purposes. We have analyzed data for four cancers using only one methodology, which limits the assessment of nuances of each cancer’s staging and management. Further study would be necessary to define the best model to be used for clinical prognostication. Stage was collapsed into simple groups without alteration based on changes to the AJCC staging system over time, which can make the cancer- and stage-specific odds ratios difficult to interpret. Use of any pseudo R2 is limiting in that they cannot be interpreted in the same manner as an R2 or coefficient of determination from a linear regression; strictly speaking, the former are better measures of effect size than of goodness of fit.29 Although the percent contributions to pseudo-R2 do not include measures of precision, they nonetheless provide an approximate understanding of the relative magnitude of the variables’ influence on the model.
Conclusion
Cancer stage contributes to explaining the differences observed for short-term mortality for GI cancers. Perioperative risk appraisal and comparative effectiveness studies of cancer surgery outcomes would benefit by including more granular cancer stage, beyond disseminated status alone.
Supplementary Material
Highlights.
Stage 1, 2, and 3 all have different 30-day mortality after resection of GI cancers
After adjustment for other factors, stage contributes to a model of 30 day mortality at a level comparable to comorbidities.
Synopsis:
Stage of cancer is not widely recognized as a predictive factor for perioperative mortality. Our study shows that stage of cancer (beyond disseminated vs not-disseminated) is an independent predictor of perioperative mortality for patients undergoing resection of gastrointestinal malignancies.
Acknowledgements
Funding: Effort by HI was supported by NIH-NCI grant 2K12 CA132783–06. The manuscript content is solely the responsibility of the authors and do not necessarily represent the official views of the NIH.
Footnotes
Meeting Presentation: This project is accepted for presentation at Society for Surgery of the Alimentary Tract’s Annual Meeting in May 2019, San Diego CA
Duplicate publishing: This work has never appeared in print nor has it been submitted to a journal for publication
Disclosure: The authors have no disclosures to report. The authors report no proprietary or commercial interest in any product mentioned or concept discussed in this article.
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
References
- 1.Wayne JD, Talamonti MS. Chapter 5.8 Tumors of the Stomach, Duodenum and Small Bowel. In: Souba WW, Fink MP, Jurkovich GJ, et al. , eds. ACS Surgery: Principles and Practice. 5 ed.: B.C. Decker; 2006. [Google Scholar]
- 2.Sepehri A, Beggs T, Hassan A, et al. The impact of frailty on outcomes after cardiac surgery: a systematic review. J Thorac Cardiovasc Surg. 2014;148(6):3110–3117. [DOI] [PubMed] [Google Scholar]
- 3.Pedroso-Fernandez Y, Aguirre-Jaime A, Ramos MJ, et al. Prediction of surgical site infection after colorectal surgery. American journal of infection control. 2016;44(4):450–454. [DOI] [PubMed] [Google Scholar]
- 4.Cyrus J. Parsa MD, AEL MD, l V. Christou MD, S. FAC, Alden H. Harken MD, S. FAC Chapter 4: RISK STRATIFICATION, PREOPERATIVE TESTING, AND OPERATIVE PLANNING. In: Souba WW, Fink MP, Jurkovich GJ, et al. , eds. ACS Surgery: Principles and Practice. 5 ed.: B.C. Decker; 2006. [Google Scholar]
- 5.Gooiker GA, Dekker JW, Bastiaannet E, et al. Risk factors for excess mortality in the first year after curative surgery for colorectal cancer. Ann Surg Oncol. 2012;19(8):2428–2434. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Swanson RS, Pezzi CM, Mallin K, Loomis AM, Winchester DP. The 90-day mortality after pancreatectomy for cancer is double the 30-day mortality: more than 20,000 resections from the national cancer data base. Ann Surg Oncol. 2014;21(13):4059–4067. [DOI] [PubMed] [Google Scholar]
- 7.Melvan JN, Sancheti MS, Gillespie T, et al. Nonclinical Factors Associated with 30-Day Mortality after Lung Cancer Resection: An Analysis of 215,000 Patients Using the National Cancer Data Base. J Am Coll Surg. 2015;221(2):550–563. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Bilimoria KY, Stewart AK, Winchester DP, Ko CY. The National Cancer Data Base: a powerful initiative to improve cancer care in the United States. Ann Surg Oncol. 2008;15(3):683–690. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.National Cancer Database. 2016; https://www.facs.org/quality-programs/cancer/ncdb. Accessed 3/4/2019.
- 10.NCDB Participant User File Supplemental Documents [press release]. Fellows of the American College of Surgeons2019. [Google Scholar]
- 11.Elixhauser A, Steiner C, Harris DR, Coffey RM. Comorbidity measures for use with administrative data. Med Care. 1998;36(1):8–27. [DOI] [PubMed] [Google Scholar]
- 12.Wolfinger ROCM. Generalized linear mixed models: a pseudo-likelihood approach. J Stat Comput Sim. 1993;4:233–243. [Google Scholar]
- 13.Nemer L, Krishna SG, Shah ZK, et al. Predictors of Pancreatic Cancer-Associated Weight Loss and Nutritional Interventions. Pancreas. 2017;46(9):1152–1157. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Timp JF, Braekkan SK, Versteeg HH, Cannegieter SC. Epidemiology of cancer-associated venous thrombosis. Blood. 2013;122(10):1712–1723. [DOI] [PubMed] [Google Scholar]
- 15.Cronin-Fenton DP, Sondergaard F, Pedersen LA, et al. Hospitalisation for venous thromboembolism in cancer patients and the general population: a population-based cohort study in Denmark, 1997–2006. Br J Cancer. 2010;103(7):947–953. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Liu Y, Cohen ME, Hall BL, Ko CY, Bilimoria KY. Evaluation and Enhancement of Calibration in the American College of Surgeons NSQIP Surgical Risk Calculator. J Am Coll Surg. 2016;223(2):231–239. [DOI] [PubMed] [Google Scholar]
- 17.Copeland GP, Jones D, Walters M. POSSUM: a scoring system for surgical audit. Br J Surg. 1991;78(3):355–360. [DOI] [PubMed] [Google Scholar]
- 18.Gupta PK, Gupta H, Sundaram A, et al. Development and validation of a risk calculator for prediction of cardiac risk after surgery. Circulation. 2011;124(4):381–387. [DOI] [PubMed] [Google Scholar]
- 19.Katz MH, Hu CY, Fleming JB, Pisters PW, Lee JE, Chang GJ. Clinical calculator of conditional survival estimates for resected and unresected survivors of pancreatic cancer. Arch Surg. 2012;147(6):513–519. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Prediction Tools - A Tool for Doctors and Patients. 2019; https://www.mskcc.org/nomograms. Accessed 3/4/19, 2019. [Google Scholar]
- 21.Bilimoria KY, Wang X, Cohen ME, et al. Do cancer centers need to assess quality and outcomes? Introducing a customized ACS-NSQIP for oncology (Onc-NSQIP). Journal of Clinical Oncology. 2011;29(4_suppl):475–475.21220609 [Google Scholar]
- 22.Solsky I, Friedmann P, Muscarella P, In H. Poor Outcomes of Gastric Cancer Surgery After Admission Through the Emergency Department. Annals of Surgical Oncology. 2017;24(5):1180–1187. [DOI] [PubMed] [Google Scholar]
- 23.Headrick JR, Miller DL, Nagorney DM, et al. Surgical treatment of hepatic and pulmonary metastases from colon cancer. Ann Thorac Surg. 2001;71(3):975–979; discussion 979–980. [DOI] [PubMed] [Google Scholar]
- 24.Kobel M, Kalloger SE, Santos JL, Huntsman DG, Gilks CB, Swenerton KD. Tumor type and substage predict survival in stage I and II ovarian carcinoma: insights and implications. Gynecol Oncol. 2010;116(1):50–56. [DOI] [PubMed] [Google Scholar]
- 25.van Rhijn BW, van der Kwast TH, Alkhateeb SS, et al. A new and highly prognostic system to discern T1 bladder cancer substage. European urology. 2012;61(2):378–384. [DOI] [PubMed] [Google Scholar]
- 26.Mobadersany P, Yousefi S, Amgad M, et al. Predicting cancer outcomes from histology and genomics using convolutional networks. Proceedings of the National Academy of Sciences of the United States of America. 2018;115(13):E2970–e2979. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Amin MB, Edge S, Greene F, Byrd DR, Brookland RK, Washington MK, Gershenwald JE, Compton CC, Hess KR, Sullivan DC, Jessup JM, Brierley JD, Gaspar LE, Schilsky RL, Balch CM AJCC Cancer Staging Manual (8th edition). Springer International Publishing: American Joint Commission on Cancer; 2017. [Google Scholar]
- 28.Ihse I. The volume-outcome relationship in cancer surgery: a hard sell. Ann Surg. 2003;238(6):777–781. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Bewick V, Cheek L, Ball J. Statistics review 14: Logistic regression. Crit Care. 2005;9(1):112–118. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.