Physical Activity and Pancreatic Cancer Risk: A Systematic Review

Pancreatic cancer is the 4th leading cause of cancer mortality in the United States (1). Because of the late presentation of symptoms and the high metastatic potential, pancreatic cancer is one of the most fatal diseases in humans and >75% of newly diagnosed patients die within a year (2). Unfortunately, little is known about the etiology of this malignancy. Over the past decade, accumulating evidence has suggested that impaired glucose tolerance and insulin resistance play important roles in the development of pancreatic cancer. The hypothesized biological mechanism involves the growth-promoting effects of excess insulin or insulin-like growth factors (3-5). Obesity and type 2 diabetes, which are closely linked to abnormal glucose metabolism and insulin resistance, increase the risk for pancreatic cancer in most epidemiologic studies (6, 7).

Physical activity can improve glucose tolerance and insulin sensitivity by increasing insulin-stimulated glycogen synthesis and enhancing skeletal muscle glucose transport (8, 9). Randomized clinical trials have indicated that physical activity alone or combined with dietary changes can produce weight loss and reduce the risk for type 2 diabetes (10-13). Therefore, physical activity may lower pancreatic cancer risk by regulating glucose metabolism or by modifying other factors such as obesity and diabetes.

Epidemiologic findings are inconsistent about the association between physical activity and pancreatic cancer, perhaps because of the inadequate statistical power resulting from the limited number of pancreatic cancer cases in most studies or a weak relation. In addition, because physical activity encompasses a variety of types (occupational, leisure time, and transport) and characteristics (frequency, intensity, and duration), different aspects of physical activity assessed may also contribute to the divergent results.

Epidemiologic evidence on the relation between physical activity and pancreatic cancer has not been quantitatively summarized. We therefore systematically reviewed and synthesized the available evidence to evaluate whether total physical activity or physical activity of a particular type or intensity is associated with the risk for pancreatic cancer.

We conducted a comprehensive literature search of MEDLINE (1966 to April 2008) and EMBASE (1974 to April 2008) using both exploded index terms (Medical Subject Headings for MEDLINE, Excerpta Medica Tree for EMBASE) and text words of physical activity (the exploded index terms physical activity, exercise, leisure activities, or walking, or the text words physical activity, exercise, leisure activities, walking, or sedentary) in combination with pancreatic cancer (the exploded index terms pancreatic cancer, pancreas cancer, or pancreas tumor, or the text words pancreatic cancer, pancreatic neoplasms, or pancreas cancer). We also searched the reference lists of relevant research reports and review articles. We limited the search to publications in English.

The inclusion criteria were predefined, and the selection process comprised two steps. The first-round selection was based on a review of the identified titles or abstracts. The studies were considered potentially eligible if they were original observational studies, only involved human subjects, and investigated the association between physical activity and pancreatic cancer risk. Studies with mortality due to pancreatic cancer were included because pancreatic cancer is a rapidly fatal disease, with its mortality almost identical to its incidence (1). The second-round selection was based on the full-text review of the retrieved articles. The studies were included if they had individual-based measurement of physical activity, provided age-adjusted relative risks with 95% confidence intervals (95% CI), and did not involve study populations overlapping with other studies. Wherein multiple publications from one study were found, only the most recent report was selected. Studies that failed to provide data to allow the calculation of age-adjusted relative risks were excluded because age is strongly associated with both physical activity and pancreatic cancer risk. Studies that relied on job titles as surrogate measures for physical activity were also excluded because this crude assessment would have led to substantial measurement errors and therefore would have underestimated the association.

The systematic search identified 234 citations: 78 from MEDLINE, 150 from EMBASE, and 6 from hand searching of the reference lists of the relevant research reports and review articles. Of these, 206 citations were excluded after a review of titles or abstracts, and 10 articles were excluded based on full-text review, leaving 18 eligible reports in the systematic review (Fig. 1).

Following the Meta-analysis of Observational Studies in Epidemiology guidelines for reporting meta-analyses of observational studies (14), we recorded information on study design, participant characteristics, assessment of physical activity and pancreatic cancer cases, adjustments for confounding factors, and estimates of associations. We considered studies as prospective if data on physical activity were collected before the diagnosis of pancreatic cancer and as retrospective if the data was collected after the diagnosis.

To comprehensively evaluate the association, we abstracted relative risks with their 95% CIs for pancreatic cancer in relation to total physical activity (occupational and leisure time) and various aspects of physical activity classified by type (occupational, leisure time, and transport) and by intensity (light, moderate, and vigorous). Overall relative risks and sex-specific relative risks were extracted. We preferentially extracted the risk estimates that incorporated frequency, intensity, and duration of physical activity because these components jointly determine the amount of physical activity. For studies that presented multiple levels of physical activity of interest, we used the relative risk for the highest as compared with the lowest level of physical activity to provide the most distinctive contrasts, assuming monotonic relations. For studies that used the highest level of physical activity as the reference group, we derived the relative risk for the highest versus lowest comparison by calculating the reciprocal of the relative risk for the lowest versus highest level of activity. For studies that reported more than one risk estimate for the same comparison, we chose those adjusted for the greatest number of potential confounders. For studies that obtained relative risks for physical activity at different periods, we used the baseline measurements or cumulative averages, if provided, to reduce misclassification of physical activity. Although the Nurses' Health Study and the Health Professionals Follow-up Study were presented in the same article (15), we extracted relative risks separately for these two cohorts because the participants of the Nurses' Health Study and the Health Professionals Follow-up Study were independently recruited. In addition, in the preliminary analysis, we obtained similar results when we used the pooled relative risks of these two cohorts provided in the original paper (15). Although two articles (16, 17) were based on the same cohort study (the UK Whitehall Study), we treated them as two separate studies because different types of physical activity were examined, and the participants in these two analyses were exclusive.

For leisure-time physical activity, it was possible to obtain a continuous measure of association because four prospective studies (15, 18, 19) used the same unit of (Metabolic Equivalent) MET-h/wk for physical activity at leisure time. All of these studies were based on the U.S. population and captured similar activities during leisure time. We therefore additionally extracted relative risks with 95% CIs and the means of MET-h/wk for each category of leisure-time physical activity. Wherein the means of MET-h/wk were not reported, we calculated the midpoint values based on cut points.

For occupational physical activity, we excluded the study by Nilsen and Vatten (20) because they assessed physical activity at work by asking the participants how often they felt physically worn out after work, which is not a good indicator for occupational physical activity.

One author (Y.B.) did the literature search, selection, and data extraction on three occasions. Any differences were resolved through the reexamination of the original reports.

Overall, 18 articles contained 19 studies: 4 studies had information on total physical activity, 3 on occupational physical activity, 16 on leisure-time physical activity, 1 on transport physical activity, 2 on light physical activity, 7 on moderate physical activity, and 8 on vigorous physical activity.

Because the selected studies differed in design, population, and quality of physical activity measurement, we estimated summary relative risks via the DerSimonian and Laird random-effects models to account for variations between the results of individual studies. Throughout the analyses, we used the originally reported overall relative risks, if provided; otherwise, we used sex-specific relative risks because they were independent estimates. Our primary analyses examined pooled relative risks comparing the highest versus the lowest category of the physical activity of interest, including total physical activity, a range of types, and intensities. For leisure-time physical activity, we additionally assessed linear association using GLST (Generalized least squares trend estimation) procedure, which first generated a continuous estimate for each of the studies and then pooled these estimates in the following step.

We assessed heterogeneity across studies by calculating the Cochrane Q statistic and the I² statistic (21, 22). To examine sources of heterogeneity, we conducted stratified and metaregression analyses with study region (North America, Europe, Asia, or multiethnic), sex (men or women), follow-up duration (<10 or ≥10 years), and adjustment for body mass index (BMI), diabetes, or both (yes or no).

To evaluate the influence of study quality on the results, we did separate meta-analyses for prospective and retrospective studies. In addition, we conducted a sensitivity analysis by removing the studies that did not exclude prevalent cancers at baseline, did not adjust for cigarette smoking, or did not adjust for smoking adequately (that is, lacked adjustment for intensity or duration of smoking).

To determine whether or not the pooled risk estimate was affected by a single study, we omitted each study in turn and recalculated the pooled relative risk for the remaining studies. To assess the influence of different categorizations of exposure on the pooled risk estimates, we repeated the analyses for the studies that had at least 4 categories of physical activity with >15 cases in the highest and the lowest categories.

We evaluated publication bias by the Begg (23) and the Egger (24) tests and visual inspection of funnel plots. All statistical analyses were done using STATA version 9.1 (STATACorp). All tests were two-sided; P < 0.05 was considered statistically significant.

The selected studies included 16 prospective cohort studies (15-20, 25-33), 1 nested case-control study (34), and 2 retrospective case-control studies (Table 1; refs. 35, 36). Of the 19 studies, 9 were conducted in the United States (one of them was multiethnic), 1 in Canada, 6 in European countries, and 3 in Japan. Overall, 4 studies involved men only, 3 studies involved women only, and the other 12 studies included men and women. Exclusion of baseline prevalent cancers was not explicitly specified in four prospective studies (16, 17, 26, 34). In all studies, information on usual physical activity was self-reported and collected after 1960. The highest and lowest levels of physical activity in each study according to the type of physical activity are provided in Figs. 2 to 4. Four studies (15, 27-29) tested the validity of the assessment of physical activity, and only one study (27) updated data on physical activity during follow-up. Most studies ascertained pancreatic cancer cases through cancer registries. For prospective studies, the follow-up was nearly complete and the duration ranged from 5 to 30 years. All risk estimates were adjusted for age and all but two studies (20, 34) adjusted for cigarette smoking. Approximately two thirds were adjusted for diabetes and about half of the studies adjusted for BMI. The degree of the adjustment for other potential confounding factors varied across studies.

Four prospective studies (25, 30, 32, 33) examined the association between total physical activity and pancreatic cancer risk. The individual relative risks ranged from 0.42 to 1.24, and the pooled relative risk was 0.76 (95% CI, 0.53-1.09; Fig. 2). Heterogeneity between studies was statistically significant (P = 0.04), explaining 60% of the total variations in results. The study by Stolzenberg-Solomon et al. (32) and the study by Nöthlings et al. (30) contributed substantially to heterogeneity. After exclusion of the study by Stolzenberg-Solomon et al., the P for heterogeneity was 0.14 (percentage of variation explained, 45%), and the pooled relative risk was 0.87 (95% CI, 0.63-1.19). After exclusion of the study by Nöthlings et al, the P for heterogeneity was 0.25 (percentage of variation explained, 28%), and the pooled relative risk was 0.60 (95% CI, 0.39-0.90). Exclusion of any other study did not materially change the P for heterogeneity or the result of no association. Subgroup analyses were not conducted because of the small number of studies.

Occupational physical activity was reported in three prospective studies (25, 26, 32). The individual relative risks ranged from 0.63 to 0.88, and the pooled relative risk was 0.75 (95% CI, 0.58-0.96; P for heterogeneity = 0.52; Fig. 3). Subgroup and sensitivity analyses were not conducted because of the small number of studies.

Leisure-time physical activity was assessed in 16 studies (14 prospective studies and 2 retrospective studies). For the 14 prospective studies (15, 17-20, 25-29, 31, 32, 34), the individual relative risks of pancreatic cancer comparing the highest versus the lowest category of leisure-time physical activity ranged from 0.60 to 1.31, and the pooled relative risk was 0.94 (95% CI, 0.83-1.05; P for heterogeneity = 0.60; Fig. 3). For the 4 prospective studies that used the same unit of MET-h/wk for leisure-time physical activity (15, 18, 19), the pooled relative risk was 0.99 (0.95-1.02) for 10 MET-h/wk increase. For the 2 retrospective studies (35, 36), the individual relative risks comparing the highest versus lowest category ranged from 0.53 to 0.84, and the pooled relative risk was 0.74 (95% CI, 0.58-0.94; P for heterogeneity = 0.58). Because of the discrepancy in the pooled estimates between prospective and retrospective studies, the subsequent analyses were conducted only among prospective studies. In subgroup analyses, no significant difference was found between men and women (P = 0.79) or between the three study regions (P = 0.08 and P = 0.23; Table 2). The association of leisure-time physical activity with pancreatic cancer was also similar for studies with >=10 years of follow-up and those with >=10 years of follow-up (P = 0.71) and for studies that adjusted for BMI, diabetes, or both and those that adjusted for neither of them (P = 0.84). In a sensitivity analysis, exclusion of any single study from the analysis had little impact on the overall finding. The results remained almost the same after restricting to the studies that measured frequency, intensity, and duration of physical activity (15, 18-20, 25, 27), restricting to the studies that validated the measurement of physical activity (15, 27-29), restricting to the studies that had at least 4 categories of exposure with >15 cases in the highest and the lowest categories (15, 18, 19, 25, 27, 28, 32), or removing the studies that included baseline prevalent cancers, did not adjust for cigarette smoking (20, 34), or did not adjust for smoking adequately (data not shown; refs. 17, 26, 31).

Transport physical activity was examined in one prospective study (16). Those who walked or bicycled to work >20 min/d did not have a lower risk for pancreatic cancer compared with those who walked or bicycled to work <10 min/d (relative risk, 1.14; 95% CI, 0.59-2.00).

Light physical activity was reported in two prospective studies (19, 27). The pooled relative risk was 1.01 (95% CI, 0.77-1.34; P for heterogeneity = 0.53; Fig. 4). Subgroup and sensitivity analyses were not conducted because of the small number of studies.

Moderate physical activity was presented in seven studies (six prospective and one retrospective). For the 6 prospective studies (15, 27, 30, 31, 33), the individual relative risks ranged from 0.41 to 1.37 and the pooled relative risk was 0.83 (95% CI, 0.58-1.18; Fig. 4). The case-control study (36) showed no association of moderate physical activity with pancreatic cancer risk among men (relative risk, 1.21; 95% CI, 0.50-2.91) or among women (relative risk, 1.00; 95% CI, 0.53-1.91). Given the inherent biases and weaknesses associated with retrospective designs, the subsequent analyses were conducted only among prospective studies. Heterogeneity across the prospective studies was evident (P < 0.001). To further explore the heterogeneity, we examined whether results differed according to participant and design characteristics. No effect modifications were found by study region (P = 0.62), sex (P = 0.89), follow-up duration (P = 0.62), or adjustment for BMI or diabetes (P = 0.46; Table 2). Furthermore, exclusion of any single study did not markedly influence the P for heterogeneity or the pooled estimate. The results did not change after restricting to the studies that measured frequency, intensity, and the duration of physical activity (15, 27, 30, 33), restricting to the studies that validated the assessment of physical activity (15, 27), or when pooling over the studies that had at least 4 categories of exposure with >15 cases in the highest and the lowest categories (data not shown; refs. 15, 27, 30).

Vigorous physical activity was examined in eight studies (seven prospective and one retrospective). For the 7 prospective studies (15, 19, 25, 27, 31, 33), the individual relative risks ranged from 0.63 to 1.21, and the pooled relative risk was 0.94 (95% CI, 0.80-1.12; P for heterogeneity = 0.72; Fig. 4). The case-control study (36) showed no significant association among men (relative risk, 0.61; 95% CI, 0.28-1.30) or among women (relative risk, 0.68; 95% CI, 0.35-1.32). The subgroup and sensitivity analyses were conducted only among prospective studies. No effect modifications were found by study region (P = 0.15), sex (P = 0.20), follow-up duration (P = 0.23), or adjustment for BMI or diabetes (P = 0.96; Table 2). Exclusion of any single study from the analysis did not change the overall findings. Similar results were yielded after restricting to the studies that measured frequency, intensity, and duration of physical activity (15, 19, 25, 27, 33), restricting to the studies that validated the assessment of physical activity (15, 27), or when pooling over the studies that had at least 4 categories of exposure with >15 cases in the highest and the lowest categories (data not shown; refs. 15, 19, 25, 27).

The funnel plots seemed slightly asymmetrical for total physical activity and moderate physical activity. However, neither the Egger nor the Begg test provided evidence for significant publication bias for the analyses of any exposure.

This meta-analysis did not support a significant association between total physical activity and risk for pancreatic cancer. The risk decreased with high physical activity at work but did not change with physical activity at leisure time. No clear benefit against pancreatic cancer was found with light, moderate, or vigorous physical activity.

Meta-analyses of observational studies are prone to biases that occurred in the original studies (14). We therefore focused on the results from prospective cohort studies and the case-control studies nested within them rather than traditional case-control studies to avoid systematic bias due to differential recall. Prospective design and low loss to follow-up in most included studies also minimized the likelihood of selection bias.

The precise assessment of physical activity presents a great challenge in observational research and the quality of activity measurement varied substantially from study to study. To reduce potential misclassification, we excluded studies that used job titles exclusively as surrogate indicators for physical activity level. We also conducted a sensitivity analysis restricted to the studies that measured activity components, including frequency, duration, and intensity to assess the influence of the measurement quality on the pooled risk estimates. The results were virtually unchanged from the results of all studies included. However, because physical activity is a complex behavior that accumulates many short unstructured activities that occur in varying contexts, self-reported physical activity measurements, even those with estimation of all types and all parameters of activity, can contain a substantial degree of measurement error that may attenuate the magnitude of the associations between physical activity and health outcomes (37). We therefore did a sensitivity analysis, pooling studies with validated assessment of physical activity. Among those studies, the study-specific risk estimates varied greatly, and the pooled relative risks were also near the null. However, the validation instruments used in some studies were self-reported physical activity record (15, 29) or energy expenditure estimated from the same questionnaire (28), which may share the same sources of errors as questionnaire measurements and thus lead to overestimating the validity of the physical activity assessment (37). Therefore, we cannot completely rule out the possibility that the lack of associations in this meta-analysis were due to measurement error in physical activity.

Although physical activity was assessed only once in most studies, changes in physical activity were not likely to substantially attenuate the association given relatively short follow-up duration for many of the cohort studies. In addition, in the study by Lee et al. (27), the only study that updated information on physical activity during follow-up, physical activity was not associated with decreased risk for pancreatic cancer.

Another concern of the present meta-analysis is confounding bias. To reduce confounding by age, we only included the studies that provided age-adjusted estimates. To determine the extent of confounding by comorbidity, we did a sensitivity analysis by removing the studies that included baseline prevalent cancer; these results were similar. Subgroup analysis also showed that the pooled risk estimates were not quite different according to adjustment for smoking, BMI, or diabetes. Other possible confounding factors, including family history of pancreatic cancer, gallbladder disease, social economic status, and diet, were not likely to influence the main findings to a great extent because the risk estimates were similar for studies adjusted for these variables and studies that did not adjust for these variables.

Although the level of physical activity might change due to subclinical pancreatic cancer, exclusion of the first few years of follow-up did not substantially change the results in all studies that did this type of analysis (15-19, 25, 32, 33), which reduced this possibility.

In the primary analyses, we compared the highest versus the lowest category of physical activity, assuming that the relation between physical activity and pancreatic cancer risk is monotonic. This assumption is reasonable based on the results of the included studies. Because studies categorized physical activity differently and therefore the activity level of the highest category in one study might correspond to the activity level of the middle category in another study, this approach of combining studies may underestimate the association. However, when we restricted the analyses to the studies that had at least 4 categories of exposure with >15 cases in the highest and the lowest categories (15, 18, 19, 25, 27, 28, 30, 32), the results were not materially different. In addition, we were able to assess a dose-response relation for leisure-time physical activity and no linear association was found, similar to the null result from the highest versus lowest comparison.

For total physical activity and moderate physical activity, relative risks were highly heterogeneous across studies. We therefore used random-effect models, which aimed to account for variability between studies. For total physical activity, one source of heterogeneity could be the difference in study populations because the two studies that contributed substantially to the variability were conducted among multiethnic cohort (30) or male current smokers (32), whereas the other two studies measured total physical activity mainly comprised Caucasians, with current smokers <25%. However, we were limited in our ability to confirm the source of heterogeneity because too few studies examined total physical activity. For moderate physical activity, tests for interaction indicated that study region, sex, and follow-up duration were not likely to account for a great deal of variations across studies. However, these tests were underpowered and heterogeneity by a variety of characteristics could not be fully investigated because not all studies provided relevant information.

Eleven prospective studies investigated the interaction by BMI for the association between physical activity and pancreatic cancer. The paper by Michaud et al. (15) reported significant effect modification by BMI when pooling the data from the Health Professionals Follow-up Study and the Nurses' Health Study. The health benefit of physical activity against pancreatic cancer was apparent among overweight participants (relative risk, 0.59; 95% CI, 0.37-0.94), whereas no association was found among normal-weight population. The other nine prospective studies (18, 19, 25, 27, 29, 30-33) did not find significant difference between subgroups defined by BMI. These subgroup analyses were based upon very few cases and thus lacked statistical power. Unfortunately, we were unable to estimate the pooled relative risks in obese individuals and in normal-weight individuals because most studies did not provide stratum-specific relative risks. Stratified analysis by the percentage of participants who were obese was not feasible either because the percentage was not uniformly given across studies.

Available data also precluded exploration of critical periods in which physical activity may influence the risk for pancreatic cancer. Publication bias is not likely to be serious in this meta-analysis based on the Egger and the Begg tests.

The inverse association between occupational physical activity and pancreatic cancer should be interpreted with caution because it was based on only three studies. In addition, the observed association could be due to unmeasured confounding. However, the confounding may exist in both directions: on one hand, individuals who have medical conditions such as diabetes are ordinarily excluded from employment as manual laborers, and on the other hand, physically demanding occupations are usually associated with harmful occupational exposures, lower social economic status, and unhealthy lifestyles such as smoking and drinking. Studies that were excluded because of the use of job titles for the exposures had inconsistent findings; a significantly inverse association (38), no significant association (39-41), and a significantly elevated risk (42) were reported with occupations that require more physical activity. However, these conflicting findings were probably resulted from imprecise measurements of occupational physical activity.

Alternatively, the observed benefit of occupational physical activity might be real. Unlike leisure-time physical activity, which is usually unstructured, of shorter duration, and varies over time, occupational physical activity is likely to be continuous physical activity and lasts for many years. Therefore, physical activity at work may have a long-term and more pronounced effect on improving insulin sensitivity, which may protect against pancreatic cancer. In addition, occupational physical activity may be better recalled because of its routine nature and being part of daily life, which decreases misclassification and thus makes the association more likely to be detected.

In the present meta-analysis, high level of total physical activity was associated with a 24% lower risk for pancreatic cancer, but the association was not statistically significant. Although all studies were conducted in developed countries where lifestyle was relatively sedentary, because occupational physical activity likely represents long-term exposure to physical activity and is more likely to be accurately reported, it may contribute a greater proportion to the total estimation of physical activity, which could explain the reduced risk with total physical activity.

Hyperglycemia, hyperinsulinemia, and insulin resistance have been hypothesized to be important in the development of pancreatic cancer (4, 43). Physical activity can enhance insulin-stimulated glycogen synthesis in skeletal muscles by increasing i.m. glucose-6-phosphate concentration (8). Exercise can also increase the expression of GLUT4 glucose transporters and thus activate muscle glucose transport (9). Both of these effects improve insulin responsiveness and lower serum levels of glucose and insulin. Therefore, physical activity may reduce the risk for pancreatic cancer through the regulation of glucose metabolism and insulin sensitivity.

In conclusion, the current data do not support an overall association between physical activity and pancreatic cancer risk. However, because occupational physical activity was inversely associated with the risk in our analysis and it may be critical to the total estimation of physical activity, we cannot exclude the possibility that, with more data, an inverse association between total physical activity and pancreatic cancer will emerge.

In addition to incorporating occupational physical activity for the estimation of the benefits of physical activity on pancreatic cancer risk, future research should make great efforts to improve the accuracy of physical activity measurement. More adequately powered studies are needed to assess the relevant period of physical activity (e.g., physical activity in early life) and whether the association varies by BMI.

No potential conflicts of interest were disclosed.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.