Diabetes Mellitus and Liver Cancer Risk: An Evaluation Based on a Systematic Review of Epidemiologic Evidence among the Japanese Population

Objective: The potential associations of diabetes mellitus with malignant neoplasms including liver cancer have become a great concern from both clinical and preventive perspectives. Although sufﬁcient evidence for a positive association between diabetes and liver cancer already exists, it would be informative to summarize up-to-date epidemiologic data in Japan. Methods: We systematically reviewed epidemiologic studies on diabetes and liver cancer among Japanese populations. Original data were obtained by searching the MEDLINE (PubMed) and Ichushi databases, complemented with manual searches. The evaluation was performed in terms of the magnitude of association in each study and the strength of evidence (‘convincing’, ‘probable’, ‘possible’ or ‘insufﬁcient’), together with biological plausibility. Results: We identiﬁed 19 cohort studies, one pooled-analysis of seven cohort studies, and seven case–control studies. Of 24 relative risk estimates of liver cancer for diabetes reported in those cohort studies, 17 showed a weak to strong positive association, six revealed no association and one demonstrated a weak inverse association (summary relative risk 2.10, 95% conﬁdence interval 1.60–2.76). Ten relative risk estimates from the case–control studies showed a weak to strong positive association ( n ¼ 9) or no association ( n ¼ 1; summary relative risk 2.32, conﬁdence interval 1.73–3.12). Overall, the summary relative risk became 2.18 (conﬁdence interval 1.78–2.69). Heterogeneity in relative risks was signiﬁcant for the difference in categor-ies of study population ( P ¼ 0.01), but not in study type ( P ¼ 0.39) or sex ( P ¼ 0.33). Conclusions: Diabetes mellitus ‘probably’ increases the risk of liver cancer among the Japanese population.


INTRODUCTION
The prevalence of diabetes mellitus has been increasing in Japan (1), and the potential associations of diabetes with major chronic diseases including malignant neoplasms have become a great concern from both clinical and preventive points of view. For primary liver cancer, most of which (.90%) comprises hepatocellular carcinoma (2), sufficient evidence already exists for a positive association with diabetes mellitus, as illustrated by several meta-analyses showing 2-4-fold increase of summary relative risk (RR) in diabetic vs. non-diabetic individuals (3 -7). Since the publication of these meta-analyses, however, relevant epidemiologic data including those in a large pooled analysis (8) have still been accumulating, particularly in Japan, and summarizing the most recent and previous data would be informative in considering the prevention of liver cancer in this country.
We aimed to review and summarize up-to-date epidemiologic findings on diabetes mellitus and liver cancer among the Japanese, whose dominant risk factors of liver cancer represent hepatitis C and B virus infection (2,9) and alcohol consumption (10). This work was conducted as part of a project of systematic evaluation of the epidemiological evidence regarding lifestyles and cancers in Japan (11).

PATIENTS AND METHODS
The details of the evaluation method have been described elsewhere (11). In brief, original data for this review were identified by searching the MEDLINE (PubMed) and Ichushi (Japana Centra Revuo Medicina) databases, complemented by manual searches of references from relevant articles where necessary. All epidemiologic studies on the association between diabetes mellitus and liver cancer incidence/mortality among the Japanese from 1950 (or 1983 for the Ichushi database) to March 2014, including papers in press if available, were identified using the search terms 'diabetes', 'liver neoplasms', 'hepatocellular', 'cohort', 'follow-up', 'casecontrol', 'Japan' and 'Japanese' as keywords. Papers written in either English or Japanese were reviewed, and only studies on Japanese populations living in Japan were included. The individual results were summarized in the tables separately by study design as cohort or case -control studies.
The evaluation was made based on the magnitude of association and the strength of evidence. First, the former was assessed by classifying the RR in each study into the following four categories, while considering statistical significance (SS) or no statistical significance (NS): (i) 'strong' (symbol or ) when RR , 0.5 (SS) or RR . 2.0 (SS); (ii) 'moderate' (symbol or ) when RR , 0.5 (NS), 0.5 RR , 0.67 (SS), 1.5 , RR 2.0 (SS) or RR . 2.0 (NS); (iii) 'weak' (symbol or ) when 0.5 RR , 0.67 (NS), 0.67 RR 1.5 (SS) or 1.5 , RR 2.0 (NS) and (iv) 'no association' (symbol -) when 0.67 RR 1.5 (NS); the RR used in this paper denotes ratio measures of effect, including risk ratios, rate ratios, hazard ratios and odds ratios. The ratios of observed to expected number of deaths, which were reported in early follow-up studies of only diabetic patients with a general population as a reference group, were also used although their nature was somewhat different from that of RRs. In the case of multiple publications of analyses of the same or overlapping datasets, only data from the largest or most updated results were included. Studies that reported RRs for impaired glucose tolerance only, or did not provide RRs or data necessary for the present authors to calculate relevant RRs were excluded.
After the above process, the strength of evidence was evaluated in a manner similar to that used in the WHO/FAO Expert Consultation Report, in which evidence was classified as 'convincing', 'probable', 'possible' and 'insufficient' (12). Biological plausibility was also taken into account for this evaluation. The final judgment was made based on a consensus of the research group members. When we reach a conclusion that there is 'convincing' or 'probable' evidence of an association, we conduct a meta-analysis to obtain summary estimates for the overall magnitude of association.
In meta-analyses of this paper, we estimated the summary RR of liver cancer for diabetes mellitus by using random effects models according to the method of DerSimonian and Laird because individual RRs across studies were significantly heterogeneous based on the Q statistic (13,14). We also performed random-effects meta-regression analyses with covariates of study type (two categories: cohort or case -control), sex (three categories: men, women or both) and study population (three categories: general population, diabetic patients or patients with chronic liver disease [CLD]) to explore a potential source of the above heterogeneity. The covariate for the difference in event (death or incidence) was not included in these analyses due to the limited number of RRs for liver cancer deaths. All statistical analyses were performed with the STATA statistical package (Stata Corp., College Station, TX, USA). Two-sided P values ,0.05 were considered statistically significant.
A summary of the magnitude of association for the cohort studies and the case -control studies is shown in Tables 3 and  4, respectively. Of 24 RR estimates reported in 20 cohort studies, 10 (8,15,16,18,23,26,27,30,31,33) showed a strong positive association between diabetes and liver cancer, five (8,21,22,32) revealed a moderate positive association and two (17,19) demonstrated a weak positive association, while the remaining seven presented no association (15,17,20,24,28,29) or a weak inverse association (33). Of 10 RR estimates in seven case -control studies, nine (34 -38,40) showed a weak to strong positive association and only one (39) presented no association. Figure 1 illustrates a forest plot of the RRs of liver cancer for diabetes in individual studies and the corresponding summary RR. In this figure, sex-specific estimates are separately plotted. For both the cohort and case -control studies as well as all studies combined, the RRs were turned out to be significantly heterogeneous (P , 0.001, 0.011 and ,0.001, respectively), and so the summary RRs were estimated by a random effects model. The summary RR was estimated as 2.10 (95% CI 1.60 -2.76) and 2.32 (95% CI 1.73 -3.12) for the cohort and case -control studies, respectively. The summary RR for all studies combined became 2.18 (95% CI 1.78 -2.69).
To explore a potential source of the heterogeneity between studies, we carried out random-effects meta-regression analyses with covariates of study type, sex and study population ( Table 5). Table 5 also presents the summary RR of liver cancer for diabetes in each subgroup by a random-effects model. No significant differences in RRs were evident between subgroups by study type (x 2 ¼ 0.75 with 1 degree of freedom [DF], P ¼ 0.39) or sex (x 2 ¼ 2.24 with 2 DF, P ¼ 0.33), but subgroups by study population revealed a significant difference (x 2 ¼ 8.96 with 2 DF, P ¼ 0.01). More specifically, the summary RR in the subgroup of diabetic patients was significantly higher than that in the subgroup of general population (P ¼ 0.004) or CLD patients (P ¼ 0.01). The residual I 2 statistic was 76% without any covariates and 63% with all covariates, and the model with all covariates showed an adjusted R 2 of 28% with an overall model P of 0.03.

994
Diabetes and liver cancer risk in Japanese

DISCUSSION
Overall, 17 of the 24 RR estimates in the cohort studies and 9 of the 10 RR estimates in the case-control studies showed a weak to strong positive association between diabetes and liver cancer risk, indicating that the overall evidence in Japan strongly supports an increased risk of liver cancer among diabetic patients. The summary RR was estimated at 2.2, which is analogous to those previously reported in several meta-analyses, with a range of 1.6-3.6 (3-7). The overall association was almost similar regardless of study type (case-control or cohort studies) or sex (men, women or both) although three RR estimates from two early studies on diabetic patients (15,16) showed a summary RR of 4.6 ( Table 5) that was significantly higher than that in subsequent studies on general populations (summary RR ¼ 2.1) or CLD patients (summary RR ¼ 1.9). Both studies (15,16) differed from the others in that they followed only diabetic patients and compared liver cancer mortality in such patients with that in the general population, adjusting only for age and observation period.
A major concern on the association between diabetes and liver cancer may be that diabetic subjects possibly include patients with hepatogenous diabetes as a complication of an advanced stage of CLD such as cirrhosis (41), thereby showing a higher liver cancer risk in appearance. Hepatogenous diabetes manifests clinically as liver function deteriorates, and it appears difficult to differentiate Type 2 diabetes from hepatogenous diabetes (41). This issue will be particularly problematic for studies on general populations or diabetic patients without clinical information on the status of subjects' liver disease and hepatitis virus markers. However, the majority of recent cohort studies on CLD patients with adjustment for the severity of CLD and hepatitis virus status (19,24,25,27,28,30 -32) also found a positive association between diabetes and liver cancer risk.
As for the diagnosis of diabetes, self-reported histories were used in 6 (8,33 -35,37,38) of the 27 studies evaluated, and the method of ascertaining diabetes was not clearly described in five studies (20,25,26,36,39). Virtually no studies took into account onset age, duration and treatment of diabetes, which appear difficult to verify but likely have influence on the disease course if diabetes truly causes a risk increase of liver cancer. Of note, some anti-diabetic drugs have been suspected to be protective (e.g. metformin (42)) or promotive (e.g. insulin and sulfonylurea (43)) in human carcinogenesis. These issues may have caused some underestimation or overestimation of true associations. Although Type 1 and Type 2 diabetes were not clearly distinguished in most studies, it seems reasonable to assume that most study subjects had Type 2 diabetes because Type 1 diabetes is rare in adults. Besides, diabetic patients may undergo more medical checkups than non-diabetic subjects, leading to increased detection of cancer and thus some overestimation of the positive association.
Additional methodological limitations should be considered. First, selection bias and information bias (e.g. recall bias on self-reported history of diabetes) might have distorted the results, especially in the hospital-based case -control studies (34 -37,39,40). Second, potential confounders were not always considered in the 27 studies evaluated. Hepatitis status, alcohol drinking or obesity (or body mass index) was not controlled in 10 (8,15 -17,21,22,33 -35,37), 15 (15 -18,20 -23,25,27,29,34,36,39,40) or 20 (15 -18,20 -23,25 -27,29 -32,34 -36,39,40) studies, respectively, although whether or not obesity should be controlled may be open to question due to the possible similarity in etiological mechanisms between diabetes and obesity, as discussed below. Moreover, only five studies (8,33,35,37,38) controlled for smoking that is now regarded as a risk factor (44 -46). Finally, publication bias could not be ruled out although statistical tests for the presence of such a bias revealed insignificant results (P ¼ 0.09 and 0.17 by the Begg's and Egger's tests, respectively; data not shown) (13,14). In relation to the biological plausibility for the observed positive association between diabetes and liver cancer, several mechanisms have been proposed. First, Type 2 diabetes is characterized by insulin resistance and resulting hyperinsulinemia. Insulin can exert a potentially mitogenic effect by activating the insulin receptor and then triggering intracellular signaling cascades that have the potential to be both mitogenic and anti-apoptotic (e.g. phosphatidylinositol 3-kinase-AKT pathway) (47) and by interacting with the insulin-like growth factor-1 (IGF-1) receptor playing a pivotal role in cancer cell Figure 1. Forest plot of the relative risks (RRs) and their 95% confidence intervals of liver cancer for diabetes mellitus in cohort and case-control studies evaluated and the corresponding summary RR. For both the cohort studies and case -control studies as well as all studies combined, the individual RRs were turned out to be significantly heterogeneous (P , 0.001, 0.011 and ,0.001, respectively), so a random effects model was used to estimate the summary RR.

996
Diabetes and liver cancer risk in Japanese proliferation (48). Elevated insulin can also increase free IGF-1 (i.e. bio-active form of IGF-1) in blood via reducing the production of IGF-1 binding proteins 1 and 2 in the liver, thereby leading to tumor development (49). This is the most frequently proposed hypothesis, which also represents a possible mechanism underlying the association between obesity and liver cancer (49,50). If this mechanism mainly contributes to hepatocarcinogenesis, adjusting for obesity as a common complication of Type 2 diabetes might be overadjustment. Secondly, hyperglycemia among diabetic patients can increase oxidative stress in the cell due to an overload of glucose oxidation and other mechanisms leading to the production of reactive oxygen species (ROS) such as hydroxyl radical (51). ROS can bind DNA, can cause gene mutations and may induce cancer development. Although it is still unclear whether hyperglycemia is associated with the development of cancer via ROS production, it is noteworthy that long-term iron reduction therapy with phlebotomy and low-iron diet, which is believed to suppress the production of ROS including hydroxyl radical (52), has lowered the incidence of hepatocellular carcinoma in patients with chronic hepatitis C (53). Lastly, patients with Type 2 diabetes often have obesity leading to elevated levels of pro-inflammatory factors such as tumor necrosis factor-alpha and interleukin-6 and decreased levels of adiponectin with anti-inflammatory actions, and resulting chronic inflammation can promote hepatocarcinogenesis (50).

EVALUATION OF EVIDENCE ON DIABETES AND LIVER CANCER RISK AMONG JAPANESE
Based on the results from the epidemiological studies evaluated and the biological plausibility as described above, we conclude that diabetes mellitus probably increases the risk of liver cancer among the Japanese population. Preventing or treating diabetes may be recommended for the prevention of liver cancer, particularly in high risk individuals such as patients with CLD and hepatitis virus carriers. Based on random effects meta-regression including covariates of study type (cohort or case-control), sex (men, women or both) and study population (general population, diabetic patients or CLD patients).