Disentangling the obesity paradox in upper gastrointestinal cancers: weight loss matters more than body mass index

Abstract

Objectives:

The obesity paradox, whereby obesity appears to confer protection against cancer-related mortality, remains controversial. This has not yet been evaluated in upper gastrointestinal cancers.

Designs:

We identified esophageal, cardia, and non-cardia gastric adenocarcinomas in the Veterans Health Administration between 2006–2016. Multivariable Cox proportional hazard models evaluate the impact of BMI at- and prior to- cancer diagnosis on mortality, adjusting for demographics, clinical characteristics, weight loss, and clinical stage (early: T1B/2N0; locally advanced: ≥T2N+).

Results:

We identify 1,308 patients: 99% male, median 66 years. In early disease, relative to BMI 30, BMI 18 and 20 at diagnosis had increased risk of death (HR 1.83, 95%CI: 1.38–2.44 and HR 1.50, 95%CI: 1.20–1.87, respectively, p<0.0001). Patients with BMI>30 did not. In locally advanced disease, at diagnosis BMI 18 (HR 1.58, 95%CI: 1.0001–1.48, p=0.05), BMI 20 (HR 1.46, 95%CI: 1.01–2.09, p=0.04), and BMI 25 (HR 1.20, 95%CI: 1.04–1.38, p=0.01) had increased risk of death, but BMI>30 did not.

In models assessing premorbid BMI and weight loss, increasing amounts of weight loss were associated with mortality independent of BMI in early cancers. For locally advanced cancers, without weight loss, there was no association with death, regardless of BMI.

Conclusion:

The predominant driver of mortality across clinical stages is weight loss. The obesity paradox appears to exist in early stage disease only. Future studies should investigate mechanisms for the obesity paradox, accompanying physiologic changes with weight loss preceding diagnosis, and if patients with low BMI and weight loss benefit from early nutritional support.

INTRODUCTION

Obesity is a known risk factor for numerous types of malignancies.[1] Among upper GI neoplasms, obesity is considered a risk factor for the development esophageal adenocarcinoma (EAC) and potentially cardia gastric adenocarcinomas (CGA), but not non-cardia gastric adenocarcinomas (NCGA).[1–3] Yet for many cancers, obesity appears to be protective against cancer-related mortality. This is known as the “obesity paradox,” and while well recognized in cardiovascular literature, remains controversial in the field of oncology.[4–8] As cachexia and weight loss often reflect cancer severity, the association between weight, particularly body mass index (BMI), and improved cancer outcomes could reflect several methodological flaws, including confounding, reverse causality, and collider bias.[6,7]

Upper gastrointestinal cancers have a unique relationship with weight given direct impacts on enteral nutrition. Limited data relating obesity and outcomes have emerged from studies of upper GI cancers. Previous studies have largely been limited to surgical candidates, with associated selection biases,[9–11] and focused on weight at diagnosis rather than evaluating premorbid BMI and weight change.[8,9] To further understand the interaction of pre-morbid obesity, weight loss to time of cancer diagnosis, and outcomes, in upper GI cancers, we accessed a large granular longitudinal database that could address these associations for patients with UGI cancers across multiple stages.

METHODS

Study Cohort

This retrospective cohort study was conducted using the Veterans Health Administration (VHA) Corporate Data Warehouse (CDW), which includes data from the unified electronic medical records of all VHA facilities (hospitals and outpatient) since October 1, 1999. To identify cancers, we utilized the Veterans Affairs Clinical Cancer Registry (CCR).[12] The CCR is a comprehensive, national database of cancers diagnosed and treated in the VHA since 1995. We queried to identify primary site gastric and esophageal adenocarcinomas. We excluded cancers of unknown primary or unknown histology and any non-adenocarcinomas. We only included those adenocarcinomas which were classified by the CCR with a known location of 1) EAC, 2) CGA, or 3) NCGA. Tumors of unspecified gastric location and overlapping cardia/non-cardia tumors were excluded. We included early (T1B/2N0) and locally advanced (≥T2N+) cancers, as determined by CCR documentation of clinical staging. We excluded patients with metastatic disease, given vastly differently survival as compared to early and locally advanced disease.

Study Exposures

Among all patients we identified all available weights and heights in the CDW. For both height and weight, we excluded implausible values (below 1^st and above 99^th percentiles). We used median height per patient to calculate BMIs. We evaluated BMI at cancer diagnosis BMI (within 1 month of cancer diagnosis date) and premorbid BMI (Supplemental Methods). When evaluating premorbid BMI, we also included weight loss, defined as the differences of weights from the premorbid state to cancer diagnosis, as a percent of the premorbid weight.

Other covariates included age, race, ethnicity, gender, smoking, diabetes, frailty, and poverty level. Zip code-level poverty is based on 2010 census data, categorized based on percentage of people within a zip code below the federal poverty line. Frailty was identified using the validated Hospital Frailty Risk Score[13], and included frailty codes within the 2 years preceding cancer diagnosis.[14] Treatments received were listed in the CCR. Standard of care was considered to be surgery +/− adjuvant chemotherapy (early) and neoadjuvant systemic chemoradiotherapy then surgery (locally advanced) for EAC, and surgery then systemic therapy (early) and neoadjuvant therapy then surgery (locally advanced) for CGA/NCGA.[15,16]

Statistical analysis

We conducted multivariable Cox proportional hazard analyses to evaluate the impact on survival by 1) BMI at cancer diagnosis and 2) premorbid BMI and percent weight loss from premorbid state to cancer diagnosis. Given different follow up times among cancer stage, we conducted analyses per clinical stage: early (T1B/2N0) and locally advanced (≥T2N+). The outcome was overall survival, with follow-up time beginning at date of cancer diagnosis and ending at date of death or right censoring as of May 1, 2020. The aforementioned covariates were evaluated. Death was obtained from the VA Vital Status File.[17]

Stata/IC 15.1 (College Station, TX) was used to perform backward selection to detect candidate models, with inclusion of all clinically significant hazard ratios (HRs), P<0.10. Type of upper GI cancer (esophageal, cardia gastric, non- cardia gastric) was forced into the model, even in cases where it did not meet statistical significance, given obvious clinical importance (Supplemental Methods). We excluded persons <18 years, those with prevalent upper GI cancers with unknown diagnosis date or incomplete data (no clinical stage, histologic confirmation, etc.). To identify weight loss and premorbid BMI, we modeled the change in weight over time, and using segmental and changepoint analysis, identified the point at which BMI began to decrease that could be attributable to cancer (Supplemental Methods). There was a decline in weight at 290 days prior to cancer diagnosis. The most proximate available weight prior to this range was identified per patient. Prior to multivariable regression, we evaluated for proportionality of hazard, and linearity assumption of all variables. The univariable association between BMI and death using locally weighted polynomial regression (LWPR) was noted to be non-linear. This was not unexpected, as BMI is known to have a “J-shaped” relationship with many outcomes, and for this reason is often transformed using restricted cubic splines.[18] To identify an appropriate spline model, we began with a three-knot construction using Harrell’s recommended percentiles,[19] and assessed visual agreement with the LWPR curve. We then refined the model by making incremental changes in knots until adequate agreement was achieved for each analysis. Although restricted cubic splines allow for precise modeling of a non-linear variables, a corollary is that associated HRs cannot be interpreted in models as linear. Accordingly, we display figures across all BMI ranges and then predict and present relative hazard ratios among BMI 18, 20, 25, 30, 35, and 40 for each analysis.

Sensitivity analyses evaluated whether receipt of standard of care treatment type varied by BMI. This was performed using multivariable logistic regression (receipt of standard of care treatment as outcome) with the same co-variates as above.

The Institutional Review Board of the Corporal Michael J. Crescenz VA Medical Center approved this study.

RESULTS

We identified 11,825 patients diagnosed with esophageal, gastric cardia, or gastric non-cardia adenocarcinoma between 2006–2016 (Figure 2). Of these, 1,308 (11.1%) were early or locally advanced clinical stage, with a median age at diagnosis of 66.0 years, of whom 99% were male. We identified 805 (61.5%) EACs, 300 (22.9%) CGAs, and 203 (15.5%) NCGAs. Table 1 compares the patient characteristics among the 3 cancer subtypes. There was no significant difference in age, sex, or smoking, but those with NCGA were more likely to be Hispanic or Latino (11.8% vs 2.7% and 1.7%, for EAC and CGA, respectively, P<0.001), less likely to be White (48.8% vs 86.1% and 87.7%, P<0.001), and more likely to reside in areas of higher poverty (4.4% versus 0.6 and 0.7%, P<0.001). Those with NCGA were also more likely to be diabetic. BMI at cancer diagnosis, premorbid BMI, and percent change in weight was similar across all groups.

Figure 2:

Flow diagram of patients with early and locally advanced esophageal, cardia, and non-cardia gastric adenocarcinoma in the Veterans Health Administration

Table 1:

Patient characteristics

	Esophageal adenocarcinoma (n=805)	Gastric cardia adenocarcinoma (n=300)	Non-cardia gastric adenocarcinoma (n=203)	p-value
Age at cancer (years), median (IQR)	65.5 (61.0, 71.9)	66.7 (61.3, 73.2)	66.7 (61.4, 75.3,	0.13
Male gender	800 (99.4%)	298 (99.3%)	201 (99.0%)	0.85
Ethnicity				<0.001
Hispanic or Latino	14 (1.7%)	8 (2.7%)	24 (11.8%)
Not Hispanic or Latino	749 (93.0%)	281 (93.7%)	173 (85.2%)
Unknown	42 (5.2%)	11 (3.7%)	6 (3.0%)
Race				<0.001
White	693 (86.1%)	263 (87.7%)	99 (48.8%)
Black / African American	27 (3.4%)	19 (6.3%)	80 (39.4%)
American Indian / Alaska Native	0 (0.0%)	2 (0.7%)	4 (2.0%)
Asian	1 (0.1%)	2 (0.7%)	2 (1.0%)
Native Hawaiian / Pacific Islander	7 (0.9%)	2 (0.7%)	4 (2.0%)
Unknown	77 (9.6%)	12 (4.0%)	14 (6.9%)
Percent residing below poverty level in zip code where patient resided at cancer diagnosis				<0.001
<10%	204 (25.3%)	85 (28.3%)	38 (18.7%)
10–25%	466 (57.9%)	159 (53.0%)	90 (44.3%)
25–50%	104 (12.9%)	40 (13.3%)	59 (29.1%)
>50%+	5 (0.6%)	2 (0.7%)	9 (4.4%)
Unknown	26 (3.2%)	14 (4.7%)	7 (3.4%)
Ever smoker	92 (11.4%)	47 (15.7%)	25 (12.3%)	0.17
Diabetes	262 (32.5%)	110 (36.7%)	88 (43.3%)	0.01
Frailty Risk Category				0.04
Low (<5)	430 (53.4%)	134 (44.7%)	95 (46.8%)
Intermediate (5–15)	261 (32.4%)	107 (35.7%)	70 (34.5%)
High (>15)	42 (5.2%)	20 (6.7%)	23 (11.3%)
Unknown	72 (8.9%)	39 (13.0%)	15 (7.4%)
Hospital Frailty Risk Score, median (IQR)	3.9 (1.6, 7.5)	4.8 (2.1, 8.5)	4.9 (2.7, 9.2)	0.002
BMI at cancer diagnosis				0.15
18.5–25	191 (23.7%)	79 (26.3%)	52 (25.6%)
<18.5	19 (2.4%)	9 (3.0%)	9 (4.4%)
25–30	241 (29.9%)	83 (27.7%)	76 (37.4%)
30–35	148 (18.4%)	63 (21.0%)	30 (14.8%)
35+	74 (9.2%)	24 (8.0%)	14 (6.9%)
Unknown	132 (16.4%)	42 (14.0%)	22 (10.8%)
Premorbid BMI				0.05
18.5–25	105 (13.0%)	45 (15.0%)	46 (22.7%)
<18.5	4 (0.5%)	3 (1.0%)	1 (0.5%)
25–30	231 (28.7%)	88 (29.3%)	61 (30.0%)
30–35	163 (20.2%)	52 (17.3%)	41 (20.2%)
35+	101 (12.5%)	40 (13.3%)	20 (9.9%)
Unknown	201 (25.0%)	72 (24.0%)	34 (16.7%)
Percent change in weight from premorbid weight to cancer diagnosis, median (IQR)	4.9 (0.3, 9.6)	6.0 (1.2, 10.6)	4.5 (0.5, 8.8)	0.27
Clinical stage				0.001
Locally advanced	601 (74.7%)	198 (66.0%)	131 (64.5%)
Regional	204 (25.3%)	102 (34.0%)	72 (35.5%)
Follow up time, median (IQR)	1.9 (0.9, 4.6)	2.0 (1.0, 4.7)	2.7 (1.3, 6.0)	0.01
Deceased	666 (82.7%)	246 (82.0%)	146 (71.9%)	0.002

At-diagnosis BMI

We conducted multivariable Cox proportional hazard analyses evaluating the association between death and BMI at cancer diagnosis, for early and locally advanced upper GI cancers (Supplemental Tables 1–2). Figure 1A and 1B show adjusted hazard ratios for death by at-diagnosis BMI. In early disease, BMI 18 at diagnosis had increased risk of death compared to BMI 30 (HR 1.83, 95% CI: 1.38–2.44), as did BMI 20 (HR 1.50, 95% CI: 1.20–1.87), p<0.0001 for both. Relative to BMI 30, a BMI 25, 35, or even 40 did not have a significantly different HR for death (Supplemental Table 3). Among locally advanced cancers, relative to a BMI 30, an at-diagnosis BMI 18 (HR 1.58, 95% CI: 1.0001–1.48, p=0.05), BMI 20 (HR 1.46, 95% CI: 1.01–2.09, p=0.04), and BMI 25 (HR 1.20, 95% CI: 1.04–1.38, p=0.01) had increased risk of death. A BMI at diagnosis of 35 or 40 did not have a significantly different HR for death (Supplemental Table 4) with a trend towards lower mortality.

Figure 1:

Adjusted hazard ratios for death by at-diagnosis BMI for early (1A) and locally advanced disease (1B), and by premorbid BMI and weight loss for early (1C) and locally advanced disease (1D).

Premorbid BMI and weight loss

After modeling at-diagnosis BMI, we then modeled the relationship between premorbid BMI and weight loss with death using multivariable Cox proportional hazard models (Supplemental Tables 5–6). Figures 1C (early disease) and 1D (locally advanced disease) demonstrate the adjusted hazard ratios for death, by premorbid BMI and percent weight loss. There was no statistical interaction between premorbid BMI and percent weight loss in any models suggesting independent associations with mortality. Among early cancers, increasing amounts of weight loss were associated with increased mortality (Supplemental Table 7) independent of BMI at diagnosis. Across all weight loss strata, a premorbid BMI 18 and 20 was consistently associated with higher mortality than a premorbid BMI 30, with significant but much smaller increases in mortality associated with higher premorbid BMIs such as 35 and 40. For locally advanced cancers (Supplemental Table 8), when a patient had no weight loss, there was no significant association of premorbid BMI and mortality. However, increasing percentages of weight loss were strongly associated with mortality irrespective of premorbid BMI.

Receipt of standard of care

We then sought to determine if patients received differential treatment by BMI at diagnosis. Table 2 displays predicted odds of receiving standard of care therapy, by at-diagnosis BMI. Among patients with early cancer, at-diagnosis BMI was not significantly associated with receipt of standard of care therapy (p=0.31). Among patients with locally advanced cancers, patients with BMI 35 were more likely to receive standard of care (according to their cancer) compared to BMI 30 (OR 1.67, 95% CI 1.19–2.34, p=0.003).

Table 2:

Predicted odds of receiving standard of care therapy, by at diagnosis BMI

BMI	Odds ratio (95% confidence interval)	p-value
Early cancers
18	0.69 (0.36 – 1.30)	0.25
20	0.89 (0.56 – 1.40)	0.60
25	1.14 (0.82 – 1.60)	0.44
30	REFERENCE
35	1.14 (0.90 –1.45)	0.29
Locally advanced cancers
18	0.79 (0.30 – 2.02)	0.61
20	0.79 (0.37 – 1.69)	0.54
25	0.82 (0.61 – 1.10)	0.19
30	REFERENCE
35	1.67 (1.19 – 2.34)	0.003

p-value compares odds versus BMI of 30

We repeated this analysis evaluating only receipt of chemotherapy. In multivariable logistic regression models, BMI at diagnosis was not associated with receipt of chemotherapy for early-stage (p=0.95) or locally advanced (p=0.64) disease. There was no evidence that BMI < 30 at-diagnosis significantly altered access to standard therapy or chemotherapy as a mediator of survival impact.

DISCUSSION

Weight loss, a symptom of most cancers, involves the loss of both adipose tissue and muscle mass. Weight loss is caused by reduced caloric intake, often driven by anorexia and dysgeusia, as well as increased resting energy expenditures related to inflammation and upregulated thermogenesis.[20] Weight loss in cancers of the upper digestive organs may be exacerbated by symptoms of the cancer itself, including dysphagia, odynophagia, early satiety, and post-prandial pain.[21] Obesity has been found to be a risk factor for certain UGI cancers. Moreover, weight at diagnosis, particularly when excessively high or excessively low, theoretically could impact patient selection for certain therapies after a cancer diagnosis. Thus, evaluating the impact of weight on cancer outcomes from UGI cancers can be highly confounded.

This is the largest analysis to date, to our knowledge, of obesity paradox in upper GI cancers, using BMI at cancer diagnosis, premorbid BMI, and percent change in weight preceding diagnosis in the largest sample evaluated to date. Previous studies evaluating the obesity paradox have suffered from numerous methodological flaws, including confounding, reverse causality, and collider bias. To overcome these methodological shortcomings, we adjusted for multiple relevant factors that could drive mortality (frailty, diabetes, smoking, age), and analyzed our cohort using both at-diagnosis and premorbid BMI (accounting for weight loss potentially related to the cancer). We included all three cancer subtypes to address heterogeneity within cancer subtypes.[6]

Two distinct patterns of the association of premorbid weight, at-diagnosis weight, weight loss, and mortality emerge in different cancer stages. In early-stage upper GI cancers, lower BMI at-diagnosis and pre-diagnosis are both associated with increased mortality, compared to overweight/obese individuals. Importantly, worsening weight loss was associated with increased mortality, independent of BMI at diagnosis. By contrast, in locally advanced cancer, while at-diagnosis BMI is associated with mortality, after adjusting for weight loss, BMI is no longer significant. This key finding highlights that for early stage cancers the obesity paradox holds, but for locally advanced cancers, we find only weight loss (not BMI) impacts survival. This discrepancy emphasizes the need to evaluate both pre-morbid and at-diagnosis BMI in obesity paradox studies.

Why weight loss has a more dominant effect on BMI with locally advanced cancers versus early cancers is unclear. Cancer cachexia and associated weight loss are evolving concepts, but are clearly associated with mortality.[22] Worsening weight loss may reflect more micro-metastatic disease, the burden of which could be differential between the two stages, and which correlates with overall survival.[23,24] It is possible that in the early stage cancers, where survival is better than locally advanced disease, weight loss may not be as strongly accompanied by other impairments that are associated with death.[25] There may be a difference in the tumor biology itself by differing stage, or some other undefined patient or tumor characteristic.[26] Future studies should attempt to elucidate this. It is important to note that unintentional weight loss is associated with increased mortality, and we are unable to distinguish intent in our cohort study.[25] However, large amounts of weight loss are unusual from traditional weight loss methods, and are likely to be unintentional.[25,27] Weight loss itself (independent of cancer site) is a poor prognostic sign, and our findings reinforce this. Future studies should investigate the potential impacts of quantifying weight loss once cancer has been diagnosed, as a potential intervention point. [28–31] Future studies should also look at how these patients with weight loss fare in other metrics, such as sarcopenia and nutritional status and begin collecting large-scale data with these measurements (since BMI is widely considered an imperfect marker of health and weight).[4,32]

We herein confirm previous findings that increasing age, diabetes, and smoking are risk factors for death, among all UGI cancer subtypes and stages, consistent with previous literature.[33–35] Interestingly, NCGA appeared to have a better prognosis in early but not locally advanced disease but otherwise survival outcomes were similar for the 3 cancer types.

Our study reliably demonstrates that for early upper GI cancers, a premorbid or at-diagnosis BMI 18 or 20 has a demonstrably worse prognosis than a normal or overweight BMI. We attempted to further elucidate the reason for this finding by evaluating treatment receipt, but underweight and normal weight persons were not less likely to receive standard of care therapy or chemotherapy in our cohort, suggesting differential administration of treatment was not the cause of the mortality differences. Our findings corroborate previous studies, which have also demonstrated that underweight persons with cancer had an increased risk of death. These include a cohort study of colorectal cancer patients in Germany, and patients with liver metastases in colorectal cancer in Europe.[36,37] These and our findings are also consistent with potential clinical explanations of the obesity paradox, including less aggressive tumor biology in obese patients, differential pharmacokinetics of cancer treatment in obesity, and the possibility that obesity provides nutritional reserve against cancer cachexia.[6,38–40] As such, there may be some validity to the obesity paradox, and future studies should continue to investigate this, particularly considering the role of nutritional supplementation at cancer diagnosis for underweight and low-normal weight individuals.

There are several limitations to this study. As a retrospective study, causality cannot be determined, and our findings do not demonstrate causality between BMI and mortality. While the use of the VHA limits generalizability, it is considered one of the most complete US based databases available, and a marker of quality for and reflection of health care in the US.[41–45] However, as in all databases, there is some missingness of data (missing at random). There is also some concern for misclassification of data, both in terms of the data for covariates as well as cancer diagnosis, though this misclassification of data should not be significantly more pronounced than in other large database studies. Misclassification of cancers is also possible but our demographic findings (that NCGA are associated with racial and ethnic minorities, as opposed to EAC/CGA) and overall survival times suggest external validity. BMI in particular is prone to misclassification, though we removed biologically implausible values.[46] BMI is not necessarily indicative of health, and including other metabolic factors will account for this to some degree, but BMI remains an imperfect measure of obesity.[47–49]. The association of BMI and survival may still be due to some other, unmeasured factor, such as sarcopenia.[47,50] In a large retrospective database study we are unable to measure this, or other unmeasured confounders. Other comorbidities and underlying disease states could prove to be unmeasured confounders, but previously conducted studies on this question have not suffered greatly from the lack of this data.[51] While these factors (cardiovascular disease, low physical activity, etc.) may represent unmeasured confounders, they have mechanisms in common with EAC and CGA related to BMI, so not adjusting for them may also provide a manner in which to avoid over-adjustment of the model.[48] Cancer specific survival was not available to us, leading us to use overall survival. To address possible reverse causality, we identify and stratify by stage of cancer diagnosis. That in a separate model we are analyzing premorbid BMI and rate of BMI change in pre-diagnosis period addresses concerns regarding the restriction of analyses to diagnosed cancer patients. We did not include distant cancer stages as these patients have poor survival, limiting extrapolation of our results to that group.

The strengths of our study include that it is the first study to evaluate the obesity paradox in upper GI cancers in a granular fashion on such a large scale. Most importantly, we overcome methodological flaws by prior studies evaluating the obesity paradox, and demonstrate the importance of appropriate methodology in this area.

CONCLUSION

We demonstrate that for early cancers, the obesity paradox exists, but weight loss pre-cancer diagnosis is an additional, strong driver of mortality. In locally advanced cancers, weight loss is the driver of mortality, across all BMIs. Age, diabetes, and smoking contribute to mortality risk. The amount of weight loss should be an important consideration in clinical practice, as it has prognostic implications. Future studies should expand on our findings by evaluating the association between weight patterns (pre- and post-cancer diagnosis) and mortality, investigating other physiologic impacts of weight loss preceding cancer diagnosis, considering if nutritional supplementation is warranted at cancer diagnosis (even for normal weight persons), evaluating other markers of health, apart from BMI, and investigating biological mechanisms that could explain the obesity paradox.

Highlights.

• Upper gastrointestinal cancers have a unique relationship with weight given impacts on nutrition

• Limited data relating obesity and outcomes have emerged from studies of upper GI cancers

• We utilize a large cohort of individuals with esophageal adenocarcinoma, cardia gastric adenocarcinomas, and non-cardia gastric adenocarcinomas to further understand the interaction of pre-morbid obesity, weight loss to time of cancer diagnosis, and outcomes

• For early cancers, the obesity paradox exists, but weight loss pre-cancer diagnosis is an additional, strong driver of mortality

• In locally advanced cancers, weight loss is the main driver of mortality, across all BMIs

Abbreviations:

BMI

body mass index

CCR

Clinical Cancer Registry

CDW

Corporate Data Warehouse

CGA

cardia gastric adenocarcinomas

EAC

esophageal adenocarcinoma

NCGA

non-cardia gastric adenocarcinomas

VHA

Veterans Health Administration