The effect of patient characteristics on second primary cancer risk in France

Data from ten French population-based cancer registries participating in the K2-France nationwide study were used to establish a population-based cohort of all patients presenting with a first cancer diagnosed between 1989 and 2004, excluding non-melanoma skin cancers. Included registries cover eight administrative regions of France (Bas-Rhin, Calvados, Doubs, Hérault, Isère, Manche, Somme and Tarn), which comprise six million inhabitants, representing 9.6% of the metropolitan French population. These registries have achieved a high degree of completeness of ascertainment and incidence data are regularly included in the ‘Cancer Incidence in Five Continents’ monograph series [16]. Since data from these public registries are not publicly available, each registry granted individual permission to use its data for the purposes of this study. The vital status of all patients was updated to December 31, 2007. The proportion of patients lost to follow-up (i.e. alive at some date before December 31, 2007 and with no SPC) was 3.36%.

A SPC was defined as the first subsequent primary cancer occurring at least two months (≥61 days) after a first cancer. Extensions, recurrences or metastases were not considered as a SPC according to the IACR rules for multiple primary cancers [17]. Computation of the person-years at risk (PYR) for each individual began after two months of follow-up, and ended at the date of SPC diagnosis, last known vital status, death or December 31, 2007, whichever came first. Patients who developed a synchronous second cancer (<61 days of follow-up) were excluded from this analysis. Data from all available cancer registries were used to determine tumor rank, including data about cancer diagnosed before the 1989–2007 period. Third and subsequent primary cancers were not considered as SPC. Finally, tobacco or alcohol consumption and first cancer treatments were not analyzed because detailed data about these exposures were not available.

The 3rd edition of the International Classification of Diseases for Oncology (ICD-O-3) was used to code invasive tumors [18]. Cancer sites (e.g. head and neck) and subsites (e.g. oral cavity) were defined in accordance with topography and morphology codes used in the EUROCARE project [19]. We focused the analyses on the risk of SPC occurring in a different subsite as the first primary cancer.

The “standard” person-years approach to study SPC incidence was used [4, 20]. The number of observed SPC (O) was compared to the number expected (E) if patients with a first cancer had experienced the same cancer rates as the general population without cancer. Therefore, E was computed by multiplying patient PYRs (allocated by gender, attained age, year of follow-up and first cancer region of diagnosis) with corresponding first cancer incidence rates estimated from the general population. Age-Period-Cohort models, where age and cohort effects were modeled using smoothing splines, were utilized to provide robust first cancer incidence rate estimates, especially for scarce sites of cancer [21]. This enables a calculation of the observed to expected ratio (O/E ratio), also called standardized incidence ratio (SIR), which can be interpreted as the relative risk of SPC among patients with a first cancer compared with the general population without cancer. Another relevant indicator provided by this method is the Excess Absolute Rate, commonly called Excess Absolute Risk (EAR), which is the absolute number of excess cancer cases per PYR among cancer patients, obtained by subtracting the expected incidence rate of cancer from the observed incidence rate of SPC.

The first stage of the analysis was to provide univariate estimations of SIR and EAR values of SPC by gender, age, year of diagnosis, follow-up and site of first primary cancer (stratified by gender). Then, patient characteristics associated with the risk of SPC were explored, using multivariate Poisson regression models described by Breslow and Day, Dickman et al. and used recently in the field of SPC incidence [20, 22‐25].

These models, as opposed to standard Cox regression models, enable account to be taken in the analysis of the number of expected cancers in the general population, and thus a direct modeling of the SIR and the EAR. Ratios of SIRs and ratios of EARs are provided by these models. For example, a ratio of SIR of 1.5 for younger patients compared to older patients means that the SIR is 50% higher among younger patients compared to older ones. In other words, the relative risk of SPC among young patients compared to the general population (SIR) is 50% higher than the relative risk of SPC presented by older patients. SIRs and EARs were modeled separately by gender, adjusting for age, year of first cancer diagnosis, follow-up and first cancer site. Finally, estimates of SIR and EAR values by site of first and second cancer were computed in order to point out which sites deserve special attention during follow-up of cancer survivors.

Byar’s accurate approximation to the exact Poisson distribution was used to compute 95% confidence intervals (95% CI) of SIR and EAR values [20]. With respect to Poisson regression models, assumptions were made that the risk of SPC was constant within each follow-up interval, and that the risk of SPC for any two patient subgroups was proportional over follow-up time. All analyses were performed using SAS version 9.2 (SAS Institute, Cary, North Carolina).

A total of 325,407 patients presented with a first cancer between 1989 and 2004 within the area covered by the cancer registries participating in the K2-France study. During the first two months after first cancer diagnosis, 4,053 (1.2%) patients presented a synchronous SPC, 27,190 (8.4%) died without developing a SPC and 4,197 were lost to follow-up (1.3%). Among the remaining cohort of 289,967 patients, 55.5% were males and the mean age at first cancer diagnosis was 64.2 years (SD 15.0). The median follow-up time was 4.0 years.

The risk of SPC was increased in this cohort of cancer survivors. With 21,226 patients (7.3%) presenting a SPC occurring in a different subsite as the first primary cancer, compared to 15,555.59 expected on the basis of the rates estimated from the overall population without cancer, the SIR was of 1.36 (95% CI, 1.35-1.38) and the EAR was of 39.4 excess cancers per 10,000 PYR (95% CI, 37.4-41.3).

Univariate estimates of SPC risk differed greatly across patient characteristics, as reported in Table 1. It is worth noting that, despite similar SIR values among male and female patients, males presented more than twice the excess risk of SPC compared with females (EAR of 58.6 versus 21.7), due to a higher expected cancer risk among males in the general population. In a similar way, although higher relative risks of SPC were reported among young patients (SIR of 2.14 for ≤44 years-old patients), cancer survivors facing the highest excess risk of SPC were those who developed a first cancer in the age range 45 to 54 years (EAR = 55.8). Finally, the risk of SPC remained elevated during follow-up, with SIR and EAR values of 1.37 and 42.0 for ≥10 years of follow-up, respectively.

Results presented in Table 2 show clearly that the risk of SPC was closely tied to the first cancer site, and that patients with a first tobacco and alcohol-related cancer carried the highest increased risk of SPC. Indeed, male patients carried the highest excess risks of SPC if they were diagnosed with a first cancer of the head and neck (EAR = 380.0), larynx (EAR = 266.3), oesophagus (EAR = 232.1), bladder (EAR = 121.7), lip (EAR = 74.3), kidney (EAR = 66.4), lung (EAR = 62.0) or a chronic lymphatic leukaemia (EAR = 50.2). Patients with a first prostate or colorectal cancer presented a rather small excess risk of SPC, with respective EAR values of 16.5 and 15.7. With respect to female patients, the highest excess risks of SPC were noted for first cancer of the larynx (EAR = 205.9), head and neck (EAR = 194.8), oesophagus (EAR = 126.5), vagina and vulva (EAR = 46.2) or an acute myeloid leukaemia (EAR = 42.7). Finally, patients with a first breast or colorectal cancer presented somewhat lower excess risks of SPC (EAR = 16.8 and 16.4, respectively).

Among male and female patients, multivariate analyses showed that age, year of diagnosis, follow-up and first cancer site were often independently associated with SIRs and EARs (Table 3). Compared with the second year of follow-up, the EAR for more than 10 years of follow-up was not significantly changed among males (EAR ratio 0.87 95% IC, 0.71-1.07) but was increased among females (EAR ratio 1.83 95% IC, 1.28-2.62). Interestingly, this trend was hidden (or even slightly inverted among males) when considering SIR values, due to the increase of expected rates of cancer with attained age during patient follow-up. These results also show that patients with a first cancer diagnosed at 55 to 64 years-old presented the highest excess risk of SPC. The risk of SPC varied widely across first cancer sites (p < .001) but detailed results are not shown, as the clinical relevance of SIR and EAR ratios across patients with different types of first cancer may be questionable. It should be noted that, although the adjustment on cancer registry showed some heterogeneity between registries (p < .001), this adjustment did not upset the results presented in Table 3 (results not shown).

Finally, Tables 4 and 5 detail risk estimates of SPC by site of first and second primary cancer, with significant EAR values superior or equal to 3.0 per 10,000 PYR among males and 2.5 among females.

There were consistent and reciprocal associations between cancer of the head and neck, oesophagus, larynx and lung. Bi-directional risks between lung and bladder cancers were found. Bladder cancers were more frequent after a first head and neck, oesophagus, larynx and kidney cancer.

There were significant associations between cancers of the upper and lower digestive tract (head and neck, oesophagus and colorectum). Among women, there was a bidirectional association between breast and corpus uteri cancers. Among men, patients with a first thyroid cancer presented an increased risk of prostate cancer. There was an association between colorectal and bladder cancers in men and between colorectal and corpus uteri cancer in women. Patients with a first urological cancer (bladder, kidney) presented a dramatically increased risk of prostate SPC. Finally, patients presented an increased risk of lung cancer after Hodgkin’s lymphoma among males, kidney cancer after testicular cancer, colorectal or bladder cancer after prostate cancer or cervical cancer.

Firstly, French cancer survivors face a 36% increased risk of SPC relative to the overall population without cancer (SIR = 1.36). This risk is markedly high compared with published estimates for Denmark (SIR = 0.91), Italy (0.93), Finland (1.12), US (1.14), Vaud Canton, Switzerland (1.2), Osaka Prefecture, Japan (1.21) and Queensland, Australia (1.27) [4‐10]. A strong argument that can be put forward to explain this singular position of France is the particularly high alcohol and tobacco consumption among male adults in France. Indeed, although alcohol consumption decreased in France from 25.0 to 13.7 l of pure alcohol per capita between 1961 and 2004, this consumption remains still high compared with the US consumption level of about 9.4 l per capita over the same period [26]. In addition, estimated age-standardized prevalence of current smoking among male adults was of 36% in 2006 in France, compared with 25% in the US [27]. Finally, in this cohort of French survivors, tobacco and/or alcohol-related cancer sites accounted for more than 50% of the total excess SPC, compared with 35% in the US, as estimated by the Surveillance, Epidemiology, and End Results (SEER) Program [4].

Secondly, multivariate analyses showed that most patient characteristics were independently associated with SIRs and EARs, and that the EAR of SPC remained elevated during patient follow-up. Although previous studies stressed the slightly decreasing trend of SIRs with follow-up [4], the elevated level of the EARs of SPC during follow-up has never been clearly identified, due primarily to the concurrent increase of the expected rates of cancer with attained age. The effect of follow-up on SPC risk may be the result of two distinct reverse effects; on the one hand the extension of carcinogen exposure after first cancer diagnosis strengthens the risk of SPC and, on the other hand, the progressive selection through cancer survival of patients with a more favorable morbidity profile lowers the risk of SPC. Pancreatic cancer is an extreme example of this progressive selection of patients through survival, pancreatic cancer survivors being significantly “protected” against SPC compared with the general population without cancer (SIR = 0.67 in our cohort). This may explain the apparent contradictory results observed among men and women in our study as, considering that survival is often worse among men compared to women in France [28], the higher selection of male patients with favorable profile during follow-up may have outweighed the strengthening effect of the extension of carcinogen exposure on the risk of SPC.

To date, different explanations have been put forward in the literature to explain the specific patterns of first and second primary cancers among cancer survivors. These include high tobacco and/or alcohol consumption, hormonal and nutritional factors, exposure to virus and immunosuppression, genetic predisposition, increased site-specific medical surveillance, late adverse effects of first cancer treatments and interactions among these factors [4, 5, 15].

Several methodological aspects of this analysis require special attention. A first questionable point is the 2-month delay used to define metachronous tumors. Although this limit was consistent with the one used in most previous population-based studies of this type [4‐7, 10], we performed sensitivity analyses using a 6-month delay, which provided very similar results (SIR = 1.35 and EAR = 38.3 versus SIR = 1.36 and EAR = 39.4, as reported in our study).

Secondly, we excluded patients with a synchronous tumor. This is slightly different from the exclusion process of synchronous cancers performed in other studies, where patients with two synchronous cancers could be included, with any subsequent cancer (i.e. third, forth, …) being considered as a new primary [4, 5]. The impact of the inclusion of patients with synchronous SPC may lead, by an anticipation of prevalence cases during the synchronous period, to:

We repeated our analysis by including the 4,053 patients (1.2%) with synchronous SPC and considered any third cancer developed by these patients as a SPC. We found respective SIR and EAR values of 1.37 and 40.0, which shows that including patients with synchronous cancer may have resulted in an only slight overestimation of the risk of SPC.

A third questionable point may be the consideration of the activity period of cancer registries and its impact on cancer rank determination. Indeed, more recently established registries may wrongly consider some SPC as first cancers if the true first cancer diagnosis date is anterior to the beginning of the registration period. In order to appreciate the direction and the magnitude of this bias, we conducted sensitivity analyses in Bas-Rhin, comparing overall SPC risk estimates obtained over fictional registration periods beginning ten years (1979), five years (1984) and zero years (1989) before the study period 1989–2007. We found that, with respect to the ten-year period, SIRs were overestimated by 0.1% and 0.6% for the five-year and zero-year periods, respectively. In other words, considering some SPC as first cancers leads to a small overestimation of SPC risk by keeping patients with higher risk of new malignancies evidenced by the presence of two cancers. In our study, almost all cancer registries were established from 1975 to 1983 (i.e. more than five years before 1989), with the exception of Manche cancer registry, set up in 1994. Considering that cases from the Manche cancer registry only represent 7.4% of our cohort, the overall magnitude of this bias is very small.

Fourthly, new malignancies occurring among patients who migrate from their original cancer registry geographic area may not have been recorded. According to census data, the mean annual migration rate out of any French administrative region (called “département”) was estimated as 3.1% between 1999 and 2004, decreasing with age from 6.9% to 1.2% for 20–29 and ≥ 60 years-old people, respectively [57]. Considering that the mean age at first cancer diagnosis is superior to 60 years in our cohort, it is reasonable to assume that this migration bias, although more significant among young patients, would not upset the overall results presented here. Moreover, a recent study assessing this type of bias among US childhood cancer survivors found no evidence of an outmigration bias in the incidence of SPCs in SEER [58].

Finally, we focused the analyses on the risk of SPC occurring in a different subsite as the first primary cancer, and this for two reasons. First, SPCs arising in the same subsite of non-paired organs are scarce, mainly due to surgical removal of the affected organ [4]. This is a well-known issue while considering prostate cancer survivors, as the high number of expected (E) SPC of the prostate calculated in such an elderly population is largely overestimated compared with the real number of prostate SPC likely to occur among these patients. Second, according to the common set of rules proposed by the International Agency for Research on Cancer (IARC), multiple primaries arising in the same site are only to be recorded when the morphology belongs to different groups [17]. For instance, among breast cancer patients, the number of observed (O) subsequent breast cancers may be underestimated as most of the new malignancies of the breast will not be recorded because of similar morphology. Ultimately, these respective overestimations of E and underestimations of O would both have led to underestimated SIR and EAR values, which rendered it advisable to exclude cancers occurring in the same subsite from SPC risks calculations. It should be noted that such exclusion was also performed in recent studies about SPC incidence performed in the US (an “all excluding same site” estimate of the risk of subsequent cancers is provided) and in the Japanese region of Osaka [4, 9].

Considering the dramatically high risk of tobacco and alcohol-related SPC in France, the results of this study will be of immediate use in a public health perspective for the preparation of the third French Cancer Plan. In particular, strengthening the measures to control tobacco and alcohol epidemics in France and promoting research about smoking cessation and alcohol abstinence interventions adapted to cancer survivors should be considered.

The use of multivariate models to analyze SPC incidence with cancer registry data will provide new research perspectives. Indeed, for specific sites of first cancer which present a particularly high risk of SPC identified in this overall analysis, further analyses are in progress in the framework of the K2-France study to provide detailed results about follow-up strategy and etiological research questions.

Furthermore, as there are no individual-specific risk prediction tools akin to those used for selected primary cancers [15], the use of the multivariate approach described in this paper will allow a determination of which clinical characteristics are associated with the risk of SPC for a given site of first cancer, enabling the construction of a tailored risk prediction tool. Such a tool will facilitate SPC individualized risk assessment by clinicians and contribute to the implementation of optimal prevention and early detection strategies.