Factors associated with breast cancer recurrences or mortality and dynamic prediction of death using history of cancer recurrences: the French E3N cohort

Breast cancer is the most common type of cancer in women worldwide and the leading cause of cancer death in women [1]. Between 8 and 10% of women diagnosed with breast cancer will present locoregional recurrences [2‐4], and 15 to 30% will develop distant metastases [4].

Tumor characteristics are the main prognostic factors for breast cancer outcome. A tumor size larger than 2 cm, axillary nodal involvement, negative estrogen and progesterone receptors, and high grade have been shown to increase the risk of death after breast cancer diagnosis and the risk of locoregional recurrence and metastases [5]. Young age at diagnosis has also been associated with a worse clinical outcome, partly because of the over-representation of more aggressive subtypes in young cases compared to older cases, such as triple negative or HER2-positive breast cancer, or because of delayed diagnosis and presentation at an advanced stage [6].

Lifestyle factors might also influence the risk of recurrence and/or death after breast cancer diagnosis. Excess weight, which is associated with a higher risk of postmenopausal breast cancer, is also associated with a higher risk of breast cancer recurrence and death [7]. Smoking is weakly associated with breast cancer risk [8] but has been associated with a higher risk of death after breast cancer, with a 60% higher risk in current smokers at breast cancer diagnosis and a 50% risk in former smokers of over 35 pack-years [9, 10]. Similar findings have been reported for the risk of recurrence [10]. Alcohol intake, a risk factor for breast cancer, increases also the risk of recurrence [11, 12]. However studies investigating the interplay between alcohol and other important lifestyle risk factors and certain disease characteristics as well as genetic susceptibility are called for. In contrast, the relation between alcohol consumption and breast cancer survival is less clear [13].

Reproductive factors associated with breast cancer risk could influence breast cancer survival: women with a diagnosis of breast cancer before menopause have a greater risk of recurrence and death [3]. Although hormone replacement therapy (HRT) has been found to increase the risk of breast cancer in healthy women, pre-diagnostic HRT use showed no effect or a decreased mortality in most observational studies [14, 15]. Conversely, the Women Health Initiative (WHI) trial noted that mammary tumors diagnosed in the estrogen plus progestin (E + P) treatment group had worse prognostic characteristics compared to the placebo group [15].

Most studies on prognostic factors for breast cancer recurrence were performed using Cox proportional hazard models or competing-risks regression models [15]. In these studies, the follow-up of the patient ceased after the first relapse, and subsequent recurrences were not considered in the prediction of cancer-related death. The successive relapses can be analyzed using a shared frailty model for the recurrent events to account for the intra-subject correlation. However, in many settings, a joint modeling approach for time to recurrence and for death is required. First, in relation to the recurrences, given the strong relationship between recurrences and breast cancer-related death [3], death should not be considered an independent censoring process. Second, in relation to death, the number of recurrences occurring during follow-up should be considered a time-dependent covariate; in this context, the Cox proportional hazards model is not applicable because it requires that the time-dependent covariates be external (or exogenous), that is, their value at a specific time point t is not affected by the occurrence of an event at a previous time point [16].

To overcome these problems, Mauguen et al. [17] proposed to apply a joint frailty model [18, 19], that accounted both for the dependence between recurrences and the dependence between recurrences and death. They showed that when predicting the risk of death, accounting for relapses led to better prediction performance.

In this paper we applied a joint model to study the effects of tumor characteristics, lifestyle characteristics and reproductive factors on the risk of relapse and death after a first invasive breast cancer diagnosis using incident cases within the population-based E3N cohort. More specifically, we studied the dynamic prediction of the risk of death considering baseline lifestyle characteristics and history of recurrent events.

During follow-up, among the 4926 cases of the study sample, 549 patients (11.2%) were lost to follow-up, 1334 recurrent events occurred including 343 locoregional recurrences, 603 metastases, and 388 s primary breast cancers; 18% of women suffered at least one recurrent event, 12% suffered exactly one event, 4% suffered two events and 2.1% suffered three or more) (see Table 1).

The median time between breast cancer and any recurrent event was 7.2 years (IQR: 3.9–11.2 years). The median time between breast cancer diagnosis and the first recurrence was 3.8 years for the first event, 5.9 years for the second event, and 6.8 years for the third event. Locoregional recurrences generally occurred sooner after breast cancer diagnosis (median = 3.9 years, IQR: 1.2–7.6) than metastases (median = 4.8 years, IQR: 2.4–7.9) or second primary breast cancers (median = 5.4 years, IQR: 2.8–8.5). Most events occurring after the first event were metastases (185 subjects, see Table 2).

During follow-up, 469 women died. The median time between breast cancer and death was 5.6 years (IQR: 3.3–8.9 years). For the women who died, the time between the last non-fatal recurrent event and death decreased with the increasing number of recurrent events. In total, 4.4% of those with no recurrence, 20% of those with one recurrence, 50% of those with two recurrences and more than 70% of those with three or more recurrences died during follow-up.

We observed that 13% of the cases were double hormone receptor negative; 40% were poorly differentiated; 19% had a tumor larger than 2 cm and 28% had nodal involvement (see Table 3).

We observed that before the cancer diagnosis, 11% of the women were current smokers (see Table 4). After the cancer diagnosis, only 6% of the women continued to smoke.

Table 5 presents the results of the multivariate joint frailty model. Negative hormonal receptors, poorly differentiated tumor, large tumor size and nodal involvement were all associated with an increased risk of recurrent events and death. The recent use of MHT was associated with a decreased risk of recurrence (hazard ratio (HR) = 0.75, 95% CI: 0.62, 0.92) but was not associated with the risk of death. Smoking increased the risk of relapse (HR for current smoker vs never smoker = 1.55, 95% CI: 1.16, 2.07). Breast cancers diagnosed after 2000 were significantly less likely to relapse than those diagnosed before. This result is likely linked to the shorter follow-up time for those with a diagnosis after 2000 and thus a lower probability to observe recurrences. The risk of death was higher when breast cancer was diagnosed after age 60 (HR = 4.50, 95% CI: 1.29, 15.78 for 60–70 vs < 50 years and HR = 9.08, 95% CI: 1.89, 43.69 for 70+ vs < 50 years). Menopausal status was not a significant risk predictor.

The risk of death was strongly dependent on the history of recurrence even after adjusting for the covariates (frailty variance θ and power α significantly different from zero).

In this article, we studied jointly the prognostic factors for successive relapses and death on a large cohort of women with a primary breast cancer. We proposed a dynamic individual prediction of the risk of death after breast cancer using a history of cancer recurrences, tumor characteristics, lifestyle characteristics and reproductive factors, showing that at a given time after diagnosis, the risk of death was strongly dependent on the number of relapses.

To date, relatively few studies have attempted to study this association. Our results confirm the relevance of the association between established prognostic factors, MTH and cigarette smoking and the risk of recurrent events or death after breast cancer. The association between lifetime cigarette smoking and risk of breast cancer recurrence has been previously described into the literature [10]. The tumor characteristics (high grade, high tumor size, nodal involvement) were also associated with a higher risk of those two failure times, as previously shown by Mauguen et al. [17]. Conversely, the risk of relapses or death was lower for tumors with positive estrogen or progesterone receptors [23]. We observed as in previous observational studies [24], a lower risk of relapse for postmenopausal women who had used MHT before breast cancer diagnosis. However, women on MHT undergo more frequent mammographic monitoring and thus have an earlier cancer diagnosis [15]. Therefore, the observed better prognosis in terms of relapses for women using MHT in our study is possibly biased. However, we adjusted for the mammographic monitoring as a binary time dependent covariate (with at least one mammographic visit before the time of interest or without mammography). This variable was non-significantly associated with both the risk of recurrence (HR = 1.25, 95% CI: 0.98, 1.59) and the risk of death (HR = 0.62, 95% CI: 0.27, 1.46). Furthermore, this adjustment did not change the effect of the MHT use. Menopausal status (with or without history of MHT use) was not significant in the joint frailty model, even if the same tendencies were observed with the use of MHT.

Smoking exposure after breast cancer diagnosis was associated with an increased risk of relapse but not with the risk of death. This might be explained by the fact that almost half of the smokers stopped smoking after their cancer diagnosis. Obesity, a factor that has been previously associated with the risk of recurrence and death after breast cancer [7], was not significant in our study. However, E3N women are relatively slim (less than 6% were obese at breast cancer diagnosis), which might have resulted in a limited power to detect an association in this subgroup. Our study did not confirm any association of alcohol consumption or education level and the risk of recurrences or death. The year of breast cancer diagnosis was used as a proxy for breast cancer care. We did not obtain a significant association with this year of diagnosis and the risk death however the risk of recurrence was significantly lower after 2000. This might probably be explained by three reasons. Due to the design of the cohort, the breast cancer diagnosed in the 1990’ties will have longer follow-up than breast cancer diagnosed more recently. We also completed these results by a sensitivity analysis on the same sample but more homogeneous in terms of follow-up durations. In this new sample, the follow-up of the subjects was truncated at 10 years and those with a follow-up greater than 10 years were censored at 10 years. The results, on these patients with shorter follow-up times showed really similar findings, with the same association on the risk of recurrence and on death. Furthermore the introduction of Trastuzamab (Herceptin) in 1998 for the adjuvant treatment of Her2+ breast cancer has significantly reduced relapse rates. Finally since 2004 in France, there is a countryside breast cancer screening program for women aged 50–74. This recent screening program is associated with breast cancers diagnosed earlier and at an earlier stage. This can also explain the reduction of recurrence rates with time. However adjusting also for the mammography monitoring as a binary time dependent covariate did not change the associations.

A major limitation of this study is the absence of variables describing the cancer treatments. The year of breast cancer diagnosis, in combination with tumor characteristics, can be viewed as a proxy variable linked to the treatment evolution. The results reflect a lower risk of relapses for those recently diagnosed who are likely using aromatase inhibitors and practicing the technique of sentinel lymph node. However, for patients diagnosed after 1999, the follow-up time is shorter and the probability of observing a relapse is consequently lower.

We have illustrated that some specific profiles of breast cancer patients vary according to tumor characteristics, lifestyle characteristics, reproductive factors and the history of relapses; these factors may predict different probabilities of dying in the future. The proposed approach was able to show that the number of relapses greatly affects the predicted risk of death and that better predictions of death are obtained by considering the history of relapses rather than only considering prognostic factors. For example, the “naïve” probability of death that does not consider the history of relapses was overestimated. Using this individual dynamic prediction tool, we illustrated that death probabilities must be updated each time additional relapses are recorded. In our analysis, we did not consider time-dependent covariates, such as modifications in the patient’s treatment, BMI, smoking habits or MHT use. However, similar to how these time-dependent covariates are allowed in the joint modeling, they are allowed in the prediction calculation. Although all of the analyses were performed using a standard statistical R package, a level of expertise in programming may be required. An easy web application as Adjuvant Online [25] could provide the opportunity for every physician to derive updated predictions for new patients when history of relapses or more outcomes (such as longitudinal biomarkers) are available, as in Krol et al. [26].

In conclusion, survival outcomes after a breast cancer are affected by the occurrence of relapses, and the proposed approach (with joint model) is really appropriate to both study their link and to predict the prognosis of patients suffering from a primary breast cancer and with possible relapses. These dynamic prediction tools would be valuable in everyday medical practice. That is, medical experience coupled with an estimate of a patient’s survival probability that considers all available information for this patient would enable physicians to make better informed decisions regarding their actions and patient care and thus improve clinical output.