Numerous epidemiological studies have examined the association between physical activity and pancreatic cancer; however, findings from individual cohorts have largely not corroborated a protective effect. Among other plausible mechanisms, physical activity may reduce abdominal fat depots inducing metabolic improvements in glucose tolerance and insulin sensitivity, thereby potentially attenuating pancreatic cancer risk. We performed a systematic review to examine associations between physical activity and pancreatic cancer. Six electronic databases were searched from their inception through July 2009, including MEDLINE and EMBASE, seeking observational studies examining any physical activity measure with pancreatic cancer incidence/mortality as an outcome. A random effects model was used to pool individual effect estimates evaluating highest vs. lowest categories of activity. Twenty-eight studies were included. Pooled estimates indicated a reduction in pancreatic cancer risk with higher levels of total (five prospective studies, RR: 0.72, 95% CI: 0.52–0.99) and occupational activity (four prospective studies, RR: 0.75, 95% CI: 0.59–0.96). Nonsignificant inverse associations were seen between risks and recreational and transport physical activity. When examining exercise intensity, moderate activity appeared more protective (RR: 0.79, 95% CI: 0.52–1.20) than vigorous activity (RR: 0.97, 95% CI: 0.85–1.11), but results were not statistically significant and the former activity variable incorporated marked heterogeneity. Despite indications of an inverse relationship with higher levels of work and total activity, there was little evidence of such associations with recreational and other activity exposures.
Pancreas cancer is the thirteenth most common malignancy in Europe but is the fifth most frequent cause of cancer death.1 Although relatively rare it grows aggressively, it metastasises early and is often detected late; consequently, 5 year survival is formidably low at only 5% in Europe and ∼75% of patients die within a year of their initial diagnosis.2 Current pharmacological and radiological modalities confer little benefit to this neoplasm; a focus on methods for both its prevention and earlier detection are therefore pivotal. Despite this dismal prognosis, the etiology of pancreatic cancer remains poorly understood. Cigarette smoking is the most strongly associated risk factor for this disease; a meta-analysis of 82 articles (42 retrospective and 40 prospective studies) has shown that smoking may increase the risk of this neoplasm by as much as 75% (RR: 1.74, 95% CI: 1.61–1.87).3 In addition, long-standing type II diabetes4, 5 and obesity6–8 have also been associated with this multifaceted disease. Despite the potential for physical activity to ameliorate the effect of these latter risk factors, findings from observational epidemiological studies examining physical activity and pancreatic cancer risk remain ambiguous.9, 10
Physical inactivity and obesity, particularly abdominal adiposity, increase the risk of developing insulin resistance.11–13 Regular exercise has shown prolonged positive effects on insulin sensitivity,14 augmenting glucose uptake through enhanced fat oxidation independently of insulin pathways.15–17 Higher levels of C-peptide and circulating insulin (markers of hyperinsulinemia) have been associated with a twofold risk of pancreatic cancer.18 In a state of hyperinsulinemia, circulating levels of insulin-like growth factor binding proteins (IGFBP-1 and 2) may be lowered increasing IGF-1 bioavailability favoring aberrant cell proliferation,19 a molecule also thought to play an important role in the transformation of pancreatic ductal cells.19 Physical activity and weight control have obvious potential to moderate serum levels of IGFBP-1 and 2, particularly as low plasma levels of IGFBP-1 have been linked with obesity and sedentary activity.20 We, therefore, performed a systematic review to examine associations between physical activity and the risk of pancreatic cancer.
Material and Methods
Six electronic databases were searched from their inception through July 2009 (PubMed, MEDLINE, EMBASE, Cochrane Library, CINAHL plus [Cumulative Index of Nursing and Allied Health Library] and SCI [Science Citation Index]). The keywords pancreas/pancreatic cancer and the Medical Subject Heading (MeSH) “pancreatic neoplasms” were combined with exploded index terms and the truncated synonyms carcinoma, neoplasm, tumour and malignancy using the OR operator. For physical activity, we used exploded index terms for the MeSH heading “motor activity” and combinations of keywords including total/occupational/recreational/transport physical activity, leisure activities, travel, work, walking, exercise, exertion and movement. These two concepts were then combined using the AND operator. We additionally used an epidemiological study search filter.21 A secondary citation search of the reference lists of all included studies was also performed. The search was limited to humans, and no language limitations were imposed.
Each electronic database was searched by the principal reviewer (M.O'R.); three reviewers then independently scanned the titles and abstracts of all identified articles after duplicate removal (L.M., M.C. and M.O'R.). Where the type of study and its relevance couldn't be ascertained from the title alone, the abstract was also read. Those clearly not relevant were excluded. If no abstract was available or where the articles' significance was ambiguous, it was included for further scrutiny. One reviewer (M.O'R.) read all the articles under consideration and assessed their eligibility for inclusion according to predefined eligibility criteria (Fig. 1). We included only observational studies of a case-control or cohort study design. In studies wherein populations had been reported on sequentially, only the most recent publication was included. Articles that included subjects with a previous history of pancreatic malignancy were excluded. Given the rapid fatality of this rare disease studies, which used mortality as a proxy for incidence, were also included.
Data extraction was undertaken by the principal reviewer (M.O'R.); two reviewers then independently checked this extraction for accuracy (H.M., M.C.), discrepancies were resolved via discussion. The closing date for the identification of articles was July 31, 2009. We designed separate data extraction forms for case-control and cohort studies recording detailed information pertaining to individual study characteristics, participant demographics, physical activity exposure assessment and covariates for adjustment. The Newcastle-Ottawa quality assessment Scale (NOS) (http://www.ohri.ca/programs/clinical_epidemiology/oxford.htm) was used to assess the methodological quality of studies based on the selection, comparability and exposure methods of cases and controls; we based study comparability on adjustment for age and cigarette smoking (the two most prominent risk factors for pancreatic cancer), additionally, amongst prospective studies our criterion for detection bias was a minimum of 5 years follow-up. From the results section of each article, we extracted the reported relative risks (RR), hazard ratios (HR) or odds ratios (OR) and 95% confidence limits for pancreatic cancer in relation to four exposure variables (total, recreational, occupational/household and transport physical activity) and measures of physical activity intensity. Whenever stated age or multivariable adjusted and both combined- and sex-specific effect measures were transcribed, therein systematically correcting for confounders in the analysis. In studies wherein more than two levels of physical activity were reported, we extracted estimates for the highest vs. the lowest category of activity. In articles where authors reported on physical activity exposures at more than one time period, the time-point with the most cases was selected.
We used a random effects model as described by DerSimonian and Laird22 to account for both within-study sampling error (variance) and also between-study variation. Wherever possible, we used reported (multivariate adjusted) or derived (unadjusted) combined relative risks, otherwise sex-specific effect measures were incorporated. Our principal quantitative synthesis involved a highest vs. lowest activity category meta-analysis of studies examining total and occupational physical activity as well as recreational activity and its intensity (light, moderate and vigorous) plus walking/cycling. In five studies which measured recreational physical activity homogenously (i.e., metabolic equivalents MET hr/wk), we estimated the linear increase in pancreatic cancer risk per unit increase in MET hr/wk by averaging the reported upper and lower MET hr/wk frequencies in each activity category and regressing this against the adjusted ORs using the methods described by Greenland and Longnecker23 and conducted meta-analysis on these estimates. Authors were contacted to obtain frequencies not stated in the original article.
The degree of statistical heterogeneity was assessed via the Cochran Q statistic and the I-squared statistic (I2), was used to quantify the variation in the effect estimate attributable to this heterogeneity.24 The degree of heterogeneity was assessed in post hoc sensitivity analysis by eliminating studies in turn and monitoring for changes in the overall estimate and the heterogeneity. We performed subgroup analyses based on study location, gender, follow-up duration and study adjustment for three key confounders (BMI, smoking status and diabetes). In sensitivity analysis, we removed studies with a NOS score of ≤6, studies in which mortality was ascertained primarily via death certificates or in which case ascertainment was not acquired from cancer registries or record linkage. In addition, we excluded studies wherein the method of physical activity exposure was not validated. The retrospective nature of case-control studies renders their findings considerably less convincing; we stratified our analyses on study design and limited sensitivity and subgroup explorations to prospective studies, consequently, the forest plots of our main analyses presented in Figures 2–4 are confined to prospective studies. Evidence of publication bias was examined using the Begg and Mazumdar's25 and Egger et al.'s26 tests. We used Intercooled STATA (SE version 9.2; StataCorp 2005, College Station, TX) for all analyses.
Thirty-two potentially eligible studies were identified (25 prospective, 7 retrospective) from 555 publications. Michaud et al.27 included data on two separate cohorts, the Health Professional's Follow-up Study and the Nurse's Health Study. Six studies reported on different activity exposures within the United Kingdom Whitehall Study,28, 29 the Japan Public Health Centre based cohort30, 31 and the American Association of Retired Persons diet and health cohort,32, 33 respectively. As we analyzed each activity exposure separately, included cases from either cohorts were independent, and therefore, all six studies were incorporated. Four studies34–37 were excluded as they were succeeded by latter publications. In all, 22 prospective and 6 retrospective studies met the criteria for inclusion (Table 1).
Table 1. Characteristics of studies included in the review
Total physical activity
Total physical activity was examined in five prospective cohorts. All studies had a NOS >6, and only one used a validated activity assessment.30 We obtained a pooled RR of 0.72 (95% CI: 0.52–0.995; p = 0.047) when combining pancreatic cancer risk estimates comparing the most active to the least active individuals (Fig. 2). However, there was some evidence of mild heterogeneity I2 = 35.2% (95% CI: 0–76). In sensitivity analysis removing the study by Nothlings et al.38 obtained an attenuated RR of 0.63 (95CI: 0.45–0.88; p = 0.007) with pheterogeneity of 0.35 and an I2 variation of 8.7%. Removal of any further study did not significantly alter heterogeneity or p value; moreover, there was no evidence of a difference between this exposure and pancreatic cancer risk between subgroups (data not shown). One retrospective study39 also examined this exposure reporting a RR of 0.62 (95% CI: 0.35–1.09; p = 0.09).
Recreational physical activity
A total of 16 prospective studies looked at physical activity in leisure time. The pooled RR was 0.94 (95% CI: 0.88–1.01; p = 0.10) comparing the highest and lowest category of recreational activity (Fig. 3). Amongst the five prospective studies describing leisure activity in MET hr/wk, a pooled RR of 0.99 (95% CI: 0.97–1.01) was obtained per 10MET hrs/wk increment. Combining estimates from three retrospective studies40–42 generated a statistically significant pooled RR of 0.74 (95% CI 0.59–0.94; p = 0.012). Heterogeneity was small amongst prospective studies, hence, subsequent study removal or further sensitivity, and subgroup explorations did not significantly alter the results (data not shown).
Occupational physical activity
Work activity was assessed in five prospective studies with a pooled RR of 1.0 (95% CI: 0.57–1.76; p =1.0) comparing the highest vs. lowest category of work activity, but there was significant heterogeneity between studies I2 = 83% (95% CI: 62–93). After removal of one study,43 the heterogeneity was markedly reduced, I2 = 0% (95% CI: 0–85); given this study's contribution to the variation in the pooled effect estimate and its crude assessment of this exposure, it was subsequently excluded. The final model of the remaining four prospective studies attained a pooled RR of 0.75 (95% CI: 0.59–0.96; p = 0.023; Fig. 2). Exclusion of any further study failed to appreciably alter the association, and there was no evidence of an effect difference between subgroups (data not shown). Two retrospective studies44, 45 also examined work activity. One study44 based on only four pancreatic cancer cases, reported a threefold increase in risk with a job that required sedentary work for more than 70% of working hours. Brownson et al. (1991) reported an odds ratio of 1.3 (95% CI: 1.0–1.8) comparing moderate vs. low levels of occupational activity.
Transport physical activity
Five prospective studies reported on walking or both walking and cycling29 as a form of commuter activity. All five prospective studies had a NOS ≥6 (Table 1). We obtained a pooled RR of 0.77 (95% CI: 0.55–1.1; p = 0.15; Fig. 3) for pancreas cancer risk comparing individuals partaking in the highest vs. the lowest category of transport activity. There was moderate heterogeneity between the studies I2 = 55.2% (95% CI: 0–84), but this was not significant. Further subgroup and sensitivity analyses did not significantly alter the observed association (data not shown).
Low intensity physical activity
Two prospective studies33, 46 examined low intensity activity in relation to pancreatic cancer risk. The pooled RR was 1.01 (95% CI: 0.77–1.34; p = 0.93) pheterogeneity = 0.53, and the percentage variation attributable to heterogeneity was 0.0% (Fig. 4).
Moderate intensity physical activity
Six studies prospectively investigated activity of moderate intensity and pancreas cancer risk with a pooled RR of 0.79 (95% CI: 0.52–1.20; pvalue = 0.28) for the highest vs. lowest category of activity (Fig. 4). However, this analysis incorporated significant heterogeneity, which could not be explained by differences in study participants, design, follow-up length or statistical adjustments for confounders. All six studies were conducted in North America, and all scored highly on the NOS scale. Two studies27 had a validated method of physical activity assessment, and combining these study estimates gave a pooled RR of 0.45 (95% CI: 0.29–0.69, pvalue = 0.000; pheterogeneity = 0.6). Other sensitivity analyses did not significantly vary the estimate; there was no evidence of a difference in effect between subgroups, (data not shown). One further retrospective study41 also examined moderate activity with respect to pancreas cancer risk reporting a nonsignificant RR of 1.07 (95% CI: 0.64–1.79).
Vigorous intensity physical activity
Nine prospective studies examined vigorous intensity activity and pancreas cancer risk with a pooled RR of 0.97 (95% CI: 0.88–1.07; p = 0.50; Fig. 4). The pheterogeneity was 0.87 (95% CI: 1.0–1.7) with an I2 variation of 0.0%. All nine studies had a NOS score ≥6, and exclusion of any single study and further sensitivity and subgroup analysis did not substantially modify the results (data not shown).
There was evidence of funnel plot asymmetry for the association between total activity and moderate intensity exercise suggesting publication or other biases. Formal tests revealed evidence of publication bias for moderate intensity exercise (Begg's test p = 0.26; Egger's test p = 0.04) but not total activity (Begg's test p = 0.46; Egger's test p = 0.11). Funnel plots for other activity variables were inspected but revealed no evidence of publication bias (not shown).
Our pooled estimates demonstrated reductions in pancreatic cancer risk with higher levels of total and occupational activity (post removal of one study with outlying results). We observed nonsignificant reduced risks for pancreatic cancer with increasing transport activity in individuals reporting some form of regular walking or cycling; however, we showed little risk reduction in those engaging in the highest category of recreational activity. When examining studies of pancreas cancer and exercise intensity, neither low nor vigorous intensity exercise showed significant associations. An inverse association was observed with moderate physical activity, though this did not reach statistical significance and incorporated significant heterogeneity.
A systematic review of 18 studies9 on which the World Cancer Research Report was based, included a meta-analysis of three comparable cohorts but did not find a statistically significant association in relation to recreational activity (RR: 0.98, 95% CI: 0.91–1.05 per 10 MET hours/week). Our findings generally accord those of a more recent systematic review and meta-analysis.47 This review searched two electronic databases through to April 2008; we conducted a more extensive electronic literature search of six databases through to July 2009 and additionally assessed the methodological quality of our included papers. Bao and Michaud excluded studies that relied on job titles as surrogate measures of physical activity and also articles that did not provide age-adjusted RRs. We include data from 10 studies that were either not identified or excluded a priori by the latter authors; furthermore our analyses contains extended follow-up reports from three cohorts and two additional prospective studies.
It is likely that any potential association between physical activity and pancreas cancer risk may be strongly confounded by BMI. In our review, 11 prospective studies reported adjusting for BMI, and when we conducted subgroup analysis based on these studies, the observed associations did not markedly change. Unfortunately, only two studies38, 48 stratified their physical activity analysis within BMI categories detailing RRs and 95% CIs; we were, therefore, unable to assess the degree to which BMI may modify the effect of physical activity, although neither study detected a statistically significant difference between physical activity and BMI category on pancreatic cancer risk. Interestingly Nothlings et al.38 reported an increased risk with higher BMI amongst men only; one plausible explanation to explain this gender disparity may relate to differences in fat distribution between men and women,49 but further studies are warranted. Similarly cigarette smoking may have a significant confounding effect when investigating the association between physical activity and pancreatic cancer risk, however, out of 18 prospective studies that reported adjusting for this variable only 431, 38, 48, 50 provided risk estimates for physical activity categories stratified by smoking status; none of which detected a statistically significant difference between physical activity and current/ever smokers vs. never smokers on pancreatic cancer risk. Only two of these studies examined the same physical activity exposure: Stevens et al.48 reported a RR of 2.31 (95% CI: 1.80–2.82) in current vs. never smokers engaging in vigorous exercise at least once per week, whereas Yun et al.50 reported a more modest RR of 1.02 (95% CI: 0.78–1.34) in current vs. never/ever smokers engaging in moderate/high levels of vigorous leisure-time activity.
Our finding that occupational physical activity was inversely associated with pancreas cancer risk should be interpreted with caution. Two studies examining this exposure43, 51 used crude or surrogate measures of assessment. For instance, in the San Francisco Longshoremen study,51 the level of work activity was allocated in accordance to an individuals' job assignment. This was a relatively highly physically active workforce however, and although the author's report adjusting for age, variations in work style and rest periods, there is likely to be residual confounding, particularly given that the most physically active men were reported to smoke less whilst working. In the remaining three studies,52–54 activity at work was self-reported and expressed as either physical or sedentary. In occupational epidemiology, the Healthy Worker Effect (HWE) has long been considered a source of selection bias. This argument states that for an individual to be employed they must exhibit lower morbidity and mortality rates than the general population; factors that have helped qualify them for their work.55 However, given the late age at onset of pancreatic cancer and that such a diagnosis is unlikely to have affected long-term employability, the influence of the HWE in these studies may be relatively moderate, particularly as the effects of the HWE decline with age and the occupational and population cohorts become more akin.
The apparent disparity between findings of prospective and retrospective studies assessing the relationship between recreational physical activity and pancreatic cancer risk in this review may pertain to the effects of selection bias. All three retrospective articles in question40–42 were population-based case-control studies from North America. Controls selected from the general population are likely to be representative of those at risk of becoming cases; however, this does not preclude the potential for differential recall bias amongst the cases,56 which may partially explain the protective results seen. Two studies,40, 41 chose not to assess physical activity in the immediate years before diagnosis to control for reverse causality (i.e., lower physical activity amongst cases as a result of subclinical disease); In addition, neither study used proxy respondents, since this can lead to non-differential misclassification bias when assessing physical activity.57 Zhang et al.42 in contrast used a substantial proportion of proxy respondents (90% of cases and 10% of controls); however, bias can be minimized by obtaining a high response rate,58 this averaged 70% from both cases and controls in these articles. All 15 prospective studies assessed recreational physical activity predominantly via self-administered questionnaires. Therefore, the likelihood of some degree of measurement error (misclassification bias because of under/over reporting of exposures leading to a loss of precision) cannot be discounted. Pancreatic cancer tends to be advanced at presentation with consequent rapid fatality, therefore, selective recruitment of survivors is possible, as only those cases well enough to take part would have been recruited, contributing to an artificially reduced risk of pancreatic cancer with higher physical activity. However, such bias is likely to have been small given the prognosis for this disease.
In the assessment of physical activity exposures, there is potential for residual confounding from engagement in other protective activities, the effects of which may be most pronounced amongst those in the highest categories of activity as they may be more inclined to exhibit other health behaviors. A recent report measuring a combined healthy lifestyle index of five factors (smoking and alcohol use, dietary quality, BMI and physical activity) illustrated a substantially reduced risk of developing pancreatic cancer in those individuals engaging in more health behaviors (score of 5) RR of 0.42 (95% CI: 0.26–0.66) p trend <0.001 vs. a score of 0 (no health behaviors). Interestingly, even those scoring just one on this index showed a reduced risk.32
We conducted sensitivity analyses on studies that incorporated validated activity assessment tools and removed those studies in which pancreatic cancer incidence was not ascertained via record linkage and similarly those that identified pancreatic deaths via death certificates (as studies of cancer deaths may be more representative of individuals with severe pancreatic cancer). Neither of the latter nor other subgroup analyses based on study characteristics such as gender, adjustment for key covariates (cigarette smoking, body mass index and diabetes), study region (Europe, North America and Asia), follow-up duration (<14 yrs/>14 yrs) or study quality score were significantly different from the estimates in the original models. Heterogeneity was evident, however, in moderate physical activity for all four subgroup analyses. Although given the small number of studies involved, it becomes difficult to adequately assess sources of heterogeneity or publication bias in this exposure.
The Million Women Study48 provided both incidence and mortality data within the same population, to avoid duplicate counts of cases, we choose to analyze only pancreatic cancer fatalities as a greater proportion of deaths accrued and follow-up was longer for mortality data in this cohort. We did, however, analyze the incident data separately; this obtained a slightly less protective pooled RR of 0.95 (95% CI: 0.87–1.04) for recreational activity but did not change the effect estimate for vigorous intensity exercise.
The stronger association that we observed with occupational physical activity may be partly attributed to its more stable and readily recalled nature (tends to be habitual activity and may therefore exert effects over a longer duration), possibly lending to cumulative improvements in insulin sensitivity. With regards to total physical activity, wherein occupational activity is an integral part, it is possible that inclusion of the latter is what is responsible for the pancreatic cancer risk reduction observed, especially as our analysis of recreational activity illustrated nonsignificant effect estimates. Although not statistically significant, we found moderate intensity activity to be more protective for pancreatic cancer than vigorous exercise. One plausible explanation for this may pertain to exhaustive exercise and immunity/inflammation. Intense exercise may induce cellular and circulating oxidative stress and free-radical production, which can cause DNA damage.59 Moreover, intense exercise has also been associated with an increase in lymphocyte apoptosis, dampening immune response.60 However, the level of intensity needed to incur this immune deficit is higher than the ranges reviewed in this analysis, and it is the opinion of the authors that our findings are more likely because of misclassification and measurement error, particularly as vigorous exercise has been seen to be protective at other cancer sites, i.e., breast and colon.10 In addition, a year-long randomized controlled trial of the effect of exercise on body weight and fat mass found moderate to vigorous intensity exercise effective in evoking statistically significant reductions in body weight, BMI and waist and hip circumferences in both men and women,61 conferring an increased intraabdominal fat loss with either a longer duration or greater gain in fitness.
In summary, the findings of our review indicate tentative inverse associations with pancreatic cancer amongst individuals partaking in the highest categories of occupational and total physical activity despite physiologically plausible mechanisms. Moderate intensity activity was seen to be slightly more protective than vigorous or exhaustive exercise. Physical activity is a notoriously difficult exposure to capture accurately in observational studies; therefore, in accounting for the loss of precision in these estimates because of potential measurement error, misclassification and the small number of estimates involved, our results should be prudently interpreted.
We thank Ms. Alex McIlroy for her assistance in the development of the search strategy used in this review. The principal author of this article (M.O'R.) is in receipt of a Department of Employment and Learning (DEL) PhD funded scholarship.