To explain the bimodal pattern, a simple metastasis development model was proposed [14,15] consisting of three distinct phases: a single malignant cell, an avascular lesion and a vascularised growth. Tumour dormancy was assumed to be possible at the single cell level and also at the point where, for further growth, angiogenesis is necessary [14]. The model was able to simulate [15] the second peak as a steady stochastic progression from one phase to the next. In order to simulate the first peak, however, it was postulated that for some subsets of patients, transitions from one state to the next were stimulated at or about the time of surgery. The conclusions were that the dominant mode of relapse in early stage breast cancer was the result of events terminating dormancy phases at the time of surgery.

Other explanations for the bimodal hazard rate pattern have also been suggested. Attempts to evaluate the role of local failure in distant failure and survival led to the hypothesis that the second recurrence surge originated by distant metastases started by a prior local failure, involving delicate issues in the formalisation and interpretation of the analyses [5]. Researchers analysing breast cancer mortality have suggested that bimodal patterns could result from the heterogeneous "malignant potential" of breast cancer [12]. According to this latter hypothesis, the two mortality peaks are attributable to different types of failure caused either by early and late tumour relapse or by local tumour recurrence and distant metastasis. Underlying these explanations is the concept that the hazard rate pattern should be explained by a juxtaposing of populations with different, yet uniform, disease development. They suggest that each peak is generated by clustering similar recurrences while different peaks result from distinct categories of recurrence. Instead of that explanation, the proposed tumour-dormancy-based model is essentially saltatory and suggests that the occurrence of different peaks is generated by the intrinsic general process of the metastatic development, although the underlying mechanism still remains unclear.

The first event was local recurrence in 119 patients and distant metastasis in 280 patients, and in 73 cases a contralateral breast cancer was first recorded and further 39 patients developed a second primary as first event.

The hazard rate for any recurrence (including both local and distant disease relapses) displayed a bimodal pattern with a first surge peaking at about 24 months (estimated risk value = 0.016) and a second peak at almost 60 months (estimated risk value = 0.009) (Figure (Figure1a).1a). The same pattern was observed when the whole recurrence risk was split into its components: the risk of local recurrence and the risk of distant metastasis (Figure (Figure1b).1b). The hazard rate curves for both contralateral breast tumours and second primary tumours did not show major peaks but revealed an almost uniform course at a quite constant level of about 0.002 for contralateral breast tumours and about 0.001 for second primary tumours (Figures (Figures1c1c and and1d1d).

The pattern of the hazard function for the analysed events was confirmed by the flexible regression spline models. The selected model, according to the AIC, had 5 df on time including the interaction with the event type indicator, thus supporting the evidence of a different hazard shape behaviour according to the different events. The estimated cause specific hazard curves from the selected interaction models are reported in Figure Figure2.2. The bimodal behaviour of distant metastasis is also evident from such an analysis, as well as the uniform tendency of contralateral breast cancers and second primaries. With regard to local recurrences, the analysis did not yield any evidence of a second peak. However, this fact does not imply the absence of such a pattern because of the parametric nature of the regression modelling approach, focused on major effects rather than local behaviours according to the available sample information.

The bimodal recurrence risk pattern is once again emerging from the clinical data of this new series of patients as a current feature of the recurrence dynamics. The peak position on the time axis is unchanged in comparison to the findings from the previously analysed series of patients [3]. In the present analysis we found lower peaks estimating the three-month recurrence risk level (0.016 versus 0.033 for the first peak and 0.009 versus 0.014 for the second peak) than in the previous study. This finding is well explained by the different characteristics of the two studied populations. Indeed, this analysis included fewer patients with a poor prognosis (10% tumour size of 2 cm or more, 37% N+, no adjuvant chemotherapy for 30% N+ patients) than the previous one (60% tumour size 2 cm or more, 49% N+, no adjuvant chemotherapy for any N+ patient) [3], accounting for the observed differences. When comparing the present series to the T1 subset of the previous series, even peak height differences almost do not exist (data not shown).

As previously observed [2], both contralateral breast cancers and second primaries display a quite constant hazard rate pattern, confirming that the occurrence of contralateral breast cancer should be considered a 'memory-less' stochastic event unrelated to the primary tumour [21,22] or tumours developing in other organs. This concept is further strengthened by the estimated annual risk level, which is greater for contralateral breast tumours than for other second primaries, as it may be expected in these patients who should be considered at higher-than-average risk for breast cancer (Figures 1c, d and and22).

The hazard rate for local recurrence presents a bimodal curve analogous to the curve of distant metastasis. The two curves cross at about eight years, when the risk of local recurrence ceases to decrease, while the hazard rate for distant metastasis goes on to regularly drop (Figures (Figures1b1b and and1c).1c). In previously analysed patients who underwent mastectomy [2], the risk of local recurrence showed a definite decreasing pattern after the second peak and promptly reached an almost null level. It should be taken into account, however, that women undergoing QUART maintain a significant portion of their mammary gland while in patients having mastectomy the breast area is totally cleared of breast tissue. Therefore, patients who undergo QUART may develop a further breast primary (sometimes from ductal carcinoma in situ) in the residual parenchyma while patients who undergo a mastectomy do not. According to this hypothesis, which was devised since the publication of the early reports on breast cancer conservative surgery [16,23] and repeatedly discussed in further reports [24-27], the hazard rate for local recurrence is the superimposition of the curve resulting from the true local recurrences and the straight line paralleling the time axis resulting from the memory-less stochastic appearance of a further ipsilateral breast cancer. The resulting hazard rate curve, therefore, is expected to show a right-sided tail at an almost constant level, as in fact occurs (Figures (Figures1b1b and and2).2). Distinguishing between true recurrence and second ipsilateral primary is clinically relevant and has been widely pursued [26-29], although without firm results until now.

The analysis of the hazard rate for distant metastases to different organs consistently suggests that the recurrence dynamics does not depend on the site of metastasis. This occurrence supports the concept that the recurrence risk pattern results from inherent features of the metastasis development process, which are apparently attributable to tumour cells, although the local micro-environmental host conditions should be permissive for further metastasis growth. This result has been partially anticipated [2] by the bimodal risk pattern for local recurrence, which may be viewed as a type of soft tissue metastasis in patients undergoing mastectomy and lacking residual breast parenchyma. However, no firm conclusion had been reached given that local recurrences may develop even from tumour cell deposits subsequent to incomplete surgical clearing, therefore not being representative of the metastatic process. The results presented here remove any doubt about this issue and suggest that some traits of the metastasis development process are similar in all seeded organs.

The present findings provide new elements for a reassessment of previously proposed explanations of the bimodal hazard rate pattern, all assuming a uniform development of the microscopic disease. The explanation suggested by Fortin and colleagues [5] fails to elucidate both the present and previous [2] findings: their opinion that the "second peak can be explained only by a second event, namely local failure" are not valid for findings obtained from our analyses that are focused on first events. Beyond this methodological drawback, however, we wish to emphasise that the proposed explanation implies that an observed peak should be related to patients dynamically clustered (eg, patients displaying local recurrence). A similar conceptual criticism can be addressed to the work by Yakovlev and colleagues [12], who found a two-component structure of the hazard function in breast cancer survival and suggested possible explanations based on the heterogeneity of "malignant potential" remaining in treated tumours. The suggested explanations, that patients with different types of failure (more vs less rapidly evolving disease) or different site of failure (local vs distant) produce different peaks, assume that each peak is generated by the clustering of cases with similar features, although different peaks result from distinct categories of patients.

The results of the present study argue against these views and support the concept that different peaks are related to the intrinsic general pathway of the metastasis development, not to distinct categories of recurrence. Indeed, our results provide evidence that the recurrence dynamics in different metastatic sites is similar to the recurrence dynamics found in different patient subsets [2,3,8], suggesting that the disease course after surgical removal of a primary tumour apparently follows a basic common pathway with well-defined steps. The proposed tumour-dormancy-based model recognises such steps as metastatic dormant states at the single cell level and avascular micrometastasis level [3,8,14], and relates the hazard rate pattern for recurrence to the non-linear disease development. Within the common rhythm of the recurrence dynamics, the risk levels at a certain time are influenced by tumour and host traits [3,14], suggesting that the pace of the common pathway is governed by a specific mixture of factors.

The bimodal recurrence risk pattern emerged from the clinical data of this new series of patients as a current feature of the recurrence dynamics. The peak position on the time axis was unchanged in comparison to the findings from the previously analysed series of patients, with an early peak at about two years and a second broader peak at about five years with a decay afterwards.

Three subsets of distantly recurring patients (to soft tissue, bone and viscera) were analysed and the analysis did not support any evidence of a different hazard shape behaviour according to the different events. The corresponding hazard rate curves displayed the typical bimodal pattern and, in particular, the position of the first peak on the time axis was at about 24 months, while the later peak emerged at about 60 months for all recurrence sites.

See related editorial by Chambers and Goss, http://breast-cancer-research.com/content/10/6/114