The Association of Abdominal Adiposity With Mortality in Patients With Stage I–III Colorectal Cancer

Justin C. Brown; Bette J. Caan; Carla M. Prado; Elizabeth M. Cespedes Feliciano; Jingjie Xiao; Candyce H. Kroenke; Jeffrey A. Meyerhardt

Disclosures

J Natl Cancer Inst. 2020;112(4):377-383. 

In This Article

Methods

Study Population and Design

The cohort—Colorectal, Sarcopenia, Cancer and Near-term Survival—was derived from the Kaiser Permanente Northern California (KPNC) cancer registry, with ascertainment of all patients diagnosed with stage I–III invasive colorectal cancer between the years 2006 and 2011, age 18–80 years, who underwent surgical resection for colorectal cancer (n = 4465). We excluded 693 patients without abdominal or pelvic CT images, 411 patients without valid measures of body mass, and 99 patients whose CT images were unreadable because of poor image quality. The final analytic sample included 3262 patients (Supplementary Figure 1, available online). A waiver of written informed consent was obtained by the study investigators. This study was approved by the KPNC and University of Alberta institutional review boards.

Supplementary Figure 1.

CONSORT flow diagram of participant inclusion in the cohort. CRC=colorectal cancer; CT=computed tomography.

Measures of Body Composition

Height (meters) and weight (kilograms) were measured at the time of diagnosis by medical assistants per KPNC clinic standards. BMI was calculated as kilograms per square meter of height (kg/m2). Body composition was measured using CT images originally collected for clinical purposes with sliceOmatic software (V5.0, TomoVision, Montreal, Canada). A single-slice transverse image at the third lumbar vertebra was used, because tissue cross-sectional areas at this lumbar region correlate with whole-body[19] and visceral and subcutaneous tissue volumes both in men and women.[20] Tissues were demarcated with a semiautomated procedure using Hounsfield Unit thresholds of –29 to 150 for muscle tissue, −150 to –50 for visceral adipose tissue, and –190 to –30 for subcutaneous adipose tissue. A random subsample of 50 CT images was analyzed by two investigators blinded to outcome, and the remaining images were analyzed by a single trained investigator blinded to outcome. The coefficients of variation for visceral adipose tissue area and subcutaneous adipose tissue area were 1.1% and 2.7%, respectively.[21]

Covariates

The KPNC electronic medical record was used to obtain baseline information on age, sex, race and ethnicity, and smoking history. The KPNC cancer registry was used to obtain information on the anatomical site of cancer, cancer stage, and the administration of chemotherapy and radiation. Conditions from the Charlson comorbidity index were obtained from the electronic medical record with a 36-month lookback interval from the time of cancer diagnosis.[22] The above-described covariate data were 99.9% complete. The measurement of physical activity was implemented across KPNC beginning in October 2009; consequently, only 447 (13.7%) of our cohort had physical activity measures available from 36 months before colorectal cancer diagnosis and up to 12 months after diagnosis. Sensitivity analyses were conducted including physical activity measures as a covariate. Additional sensitivity excluded patients who had a BMI classified as underweight at colorectal cancer diagnosis (<18.5 kg/m2).

Study Outcomes

The primary study outcome was all-cause mortality, defined as the time from CT image acquisition to death from any cause. The secondary study outcome was colorectal cancer-specific mortality, defined as the time from CT image acquisition to death attributable to colorectal cancer. Deaths were identified from the California state death registry, National Death Index using Social Security Administration data, and KPNC electronic mortality files through December 31, 2016. Deaths were classified as cancer specific if colorectal cancer was documented as an underlying or contributing cause of death on the death certificate through January 31, 2015.

Statistical Analysis

Investigators often model visceral and subcutaneous adipose tissue using categorical variables. However, categorization induces discontinuities in mortality risk between exposure categories that reduce statistical power and are often difficult to justify biologically and interpret clinically. In contrast to prior studies, we modeled visceral and subcutaneous adipose tissue as continuous variables using restricted cubic splines. Cubic splines accommodate nonlinearity and provide statistically efficient and visually intuitive descriptions of prognostic associations.[23] Model parsimony and the Akaike information criterion were used to select the optimal number of knots for visceral and subcutaneous adipose tissue. In the absence of available literature, knot locations could not be biologically prespecified and were based on default locations using data-derived quantities.[23]

Multivariable-adjusted Cox proportional hazards models were used to evaluate associations between baseline visceral and subcutaneous adipose tissue and mortality risk. Models were adjusted for age, race and ethnicity, cancer site, cancer stage, chemotherapy, radiation therapy, smoking history, Charlson comorbidity index, and height. In addition to muscle area, we simultaneously adjusted statistical models for visceral adipose tissue and subcutaneous adipose tissue to evaluate their independent effects. Linearity and nonlinearity were inspected visually using spline plots and examined statistically using likelihood ratio tests.[23] The assumption of proportional hazards was examined by visual inspection of graphical log-log plots and tested statistically in a generalized linear regression of the scaled Schoenfeld residuals on time.[24] Sensitivity analyses that included physical activity as a covariate used the missing-indicator method.[25] The Pearson correlation coefficient was used to quantify the strength of the association between visceral and subcutaneous adipose tissues.

Effect modification was examined by including the multiplicative interaction terms in the regression models and examined using the likelihood ratio test. Based on prior studies that identified sex-by-BMI interactions,[2,26] sex was prespecified as the primary effect modifier of interest. Baseline characteristics were compared between men and women using the χ2 test for categorical variables and the t test for continuous variables. In post hoc analyses, we also examined the following subgroups: age, cancer site, cancer stage, smoking history, and muscle area. Because of known limitations in statistical power, the threshold for statistical significance for interactions was prespecified at P less than .10.[27] To conclude the presence of effect modification, a statistically significant likelihood ratio test P value and clinically meaningful visual differences in effect size estimates from the spline plots were required.[28] All statistical tests were two-sided.

Comments

3090D553-9492-4563-8681-AD288FA52ACE

processing....