A hierarchical approach to combine histological grade and immunohistochemical factors to identify high-risk luminal breast cancers

Background The luminal subtype accounts for ~70% of newly diagnosed breast cancer patients. Although it has a better prognosis, approximately 30% of them develop a late relapse. Identifying those patients is of interest to improve treatment decisions. Methods A retrospective observational, single-centre study based on data from medical records of 572 non-metastatic (I–III) invasive ductal breast carcinoma patients, 448 with luminal tumours and 124 with triple-negative tumours. Kaplan–Meier, Cox regression and time-dependent Cox regression were carried out to obtain the prognosis value of risk factors. Results During a median observation of 5.5 years, 105 distant metastasis events and 105 all-cause deaths were observed. In addition to known clinicopathological factors (i.e., age, tumour size and lymph node metastasis), the high semi-quantitative expression of both hormone receptors was associated with distant metastasis-free survival (DMFS) (adjusted hazard ratio (HaR): 0.524 (0.316–0.867), p = 0.012) and overall survival (OS) (adjusted HaR: 0.486 (0.286–0.827), p = 0.008). The stratified analysis made it possible to identify risk modification factors. Subsequent stratification by histological grade, Ki-67 and semi-quantitative PR expression or, mainly, the composite semi-quantitative expression of hormone receptors (cHR) enabled the identification of luminal breast cancer patients of adjuvant schema at higher risk for metastasis and death. However, initial analyses including patients of neoadjuvant therapy pointed to a path of subsequent stratification by cHR and histological grade, also enabling grouping of luminal breast cancer patients with similar prognosis for DMFS (cHR ≤ 4+ G2 or G3 versus triple-negative, adjusted HaR: 0.703 (0.415–1.189), p = 0.189) and OS (cHR ≤4+ G2 or G3 versus triple-negative, adjusted HaR: 0.662 (0.403–1.088), p = 0.104). Conclusion The semi-quantitative expression of both cHR, Ki-67 proliferation index and histological grade can identify luminal breast cancer patients at greater risk of developing metastasis and death when combined in a hierarchical fashion, and could be useful for a better prognosis stratification in services from low- and middle-income countries.

Tumours were classified according hormonal receptor positivity as 1+ whenever ≥1% and ≤10% of tumours cells expressed the receptor with low intensity; 3+ whenever 60%-90% of tumours cells expressed the receptor with moderate/strong intensity; and 4+ whenever >90% of the tumour cells expressed the receptor with strong intensity. Tumours with an expression greater than 1+ and less than 3+ are considered undefined as it is a very broad spectrum [46,47]. Therefore, these patients were not included in the initial analyses.
Posteriorly, patients were also reclassified by the sum of both HR expression patterns in a system from 1 to 8 points. As no consensus has been established for 2+ classification, patients in this situation were included only if the other HR had an expression pattern of 4+, being classified as 5, or 1+, being classified as 3, or 0, being classified as 2. The 2+ classification was only considered if it was accompanied by the percentage and labelling of both receptors to enable the sum of their semi-quantitative expressions. The expression of both hormone receptors, and their composite (sum), was evaluated as continuous variables.
From a total of 2,186 records, 1,567 were excluded based on the aforementioned criteria. After exclusion criteria were applied, 619 patients with non-metastatic, invasive ductal carcinoma of no special type histology of luminal or triple-negative subtype with complete clinicopathological reports were included in the study.

Statistical analysis
Distributions were analysed using the Kolmogorov-Smirnov test. Continuous variables with normal distribution were described as mean (± standard deviation) and non-parametric variables as median (minimum-maximum); categorical variables were described as frequencies.
The association between interdependent categorical factors was evaluated using Pearson's χ 2 test. The association was considered positive (direct) when the adjusted standardised residuals had a value >(+2.0) and considered negative (indirect/inverse) when the value was <(-2.0).
To assess the degree of agreement in the classification of patients between two different approaches, Cohen's Kappa test was used.
To assess the bivariate correlation in the presence of an ordinal variable, Spearman's correlation test was used.
To determine the independent prognosis value of the variables, the Cox proportional hazards regression model was used. The proportionality of categorical variables was assessed using the Kaplan-Meier (KM) estimator curve associated with the Log-Rank test.
The proportionality, and linearity, of the risk of continuous variables, was tested by categorising them into defined percentiles. Risk proportionality was assumed as long as and when there were constant risks, evaluated by the KM estimating curve associated with the Log-Rank test. Associated with this, the correlation between the partial residuals generated by the univariable Cox regression with the observation time for each outcome was evaluated; time dependence was assumed in the presence of correlation between these variables and visually analysed by a scatterplot.
In violating the prerequisite of proportionality of risks, time-dependent Cox regression models were used. For categorical variables, observation times were categorised as the time of inflection (crossing) between the curves, and the T_COV variable with interaction (*) with this categorisation was used. Univariable analysis of prognosis factors in a time-dependent context was carried out in the presence of the T_COV* variable.
The Cox regression model with the maximum prognosis value for the analysed cohort was obtained using a two-block analysis. In the first block, candidate factors were included and the Stepwise Forward Wald method was used with an entry p-value equal to 0.10 and output p-value equal to 0.05 to reduce covariate collapsibility [48] and overfitted models; in the second block, the forced entry method of known prognosis factors, like T, N and age, and treatments was used, when pertinent. A second model was built inserting all variables with a p-value < 0.05 in univariable analysis, or with a known theoretical value, to correct any possible overfitting. In risk stratification analyses, the models were always adjusted by T and N for distant metastasis-free survival (DMFS), and T, N and age for overall survival (OS). The simple contrast was used to establish the reference level for categorical variables; the repeated contrast was used to carry out clustering of levels with similar prognosis within a variable.
The aforementioned statistical analyses were carried out with IBM SPSS v25.0.
Using Jamovi v1.6.5.0 software, continuous variables were categorised according to optimal cut-off points obtained by the system after survival analysis of continuous predictors.
A p-value < 0.05 was considered significant for all the aforementioned analyses.

Analysis of risk factors in patients with luminal breast cancer
Of the 619 patients included in the study, 495 and 124 were identified as having luminal and triple-negative breast cancer, respectively. In a median observation period of 64.9 months, 112 events of distant metastasis and 112 deaths were observed. Detailed data are described in Table 1.
To identify the risk factors associated with the development of distant metastases and death in patients with luminal breast cancer, analyses only in this subgroup were carried out.
Survival analyses for continuous variables were carried out for Ki-67 and the quantitative expression of the hormone receptors (HRs) and the optimal cut-off points were identified. For Ki-67, the survival analysis showed no association with DMFS (hazard ratio (HaR): 1.01 (1.00-1.02), p = 0.116). But for OS, an increment of 1% of tumour cells expressing Ki-67 increased by 1% the odds of death (HaR: 1.01 (1.00-1.02), p = 0.009). The optimal cut-offs retrieved by the analysis were >11 and >12 for DMFS and OS, respectively.
Posteriorly, it was tested whether there is a prognosis value of the semi-quantitative expression of HRs and what the optimal cut-off points are. However, 41 patients were classified in category 2+ of the ordinal (cross) system, which has a very broad expression (10%-60% of expressing cells) and generates inconsistencies for semi-quantitative reclassification. Furthermore, nine patients had no quantitative expression either in percentage or in the ordinal system. For this reason, these 50 patients were not included in further analyses. Thus, only 445 patients with luminal tumours were included in subsequent analyses.
For DMFS, the PR semi-quantitative expression (HaR: 0.83 (0.70-0.99), p = 0.035) was significant, meaning a 17% risk decrease by an increment of 1+. The composite semi-quantitative expression of hormone receptors (cHR) showed only a trend for OS (HaR: 0.90 (0.80-1.01), p = 0.064), meaning a 10% risk decrease by an increment of 1 point. The optimal cut-offs retrieved by the analysis were >1 and >4 for PR and cHR, respectively.
Although with a moderately strong correlation (Spearman's rho: 0.398, p < 0.005) and association (Pearson's χ 2 : 176.768, p < 0.005) between ER and PR, the semi-quantitative expression of ER was not significant for any outcome. As the results were obtained through the analysis of variables that can violate the assumption of risk proportionality, they were categorised according to the cut-off points obtained in the previous analysis, and proportionality was tested using a KM curve and Log-Rank test. Cut-off points >12, >1 and >4 were selected for Ki-67, PR and cHR, respectively, as a function of significance in the results of previous analyses.
The cHR cut-off enabled the inclusion of three additional patients due to the single expression of ER on their tumours, totalising 448 luminal patients included in further analysis. The clinical data of the 572 patients, 448 with luminal tumours and 124 with triple-negative tumours, for further analysis from this point on are presented in Table 2.
Overall, the categorisation of Ki-67, the semi-quantitative expression of PR and the composite of both hormone receptors showed significance, or were close, for both outcomes (Figures 1-3).
The qualitative expression of hormone receptors and the histological grade also have prognosis values. Therefore, they were evaluated by the KM method. Unlike the semi-quantitative expression, the qualitative expression of hormone receptors was only significant for OS ( Figure 4). The histological grade also showed significance, or was close, for both outcomes ( Figure 5).
Differently from pathological factors (T and N -data not displayed), the risks remain proportional during the first 60 months (5 years), with changes at later times and intersections being observed close to 120 months (10 years) for some factors (Figures 1-5). Therefore, timedependent analysis was carried out using an interaction term between the categorised times of observation and the variables of interest.
For DMFS, only the semi-quantitative expressions of PR or cHR resulted in being independent after correction by other factors by stepwise and insert methods, respectively (Table 3).
For OS, the stepwise model included both Ki-67 and qualitative expression of HR as significant (Table 4). Due to strong to very strong correlations between semi-quantitative PR expression and cHR (Spearman's rho: 0.802, p < 0.005); semi-quantitative PR expression and qualitative HR expression (Spearman's rho:0.553, p < 0.005); and between cHR and qualitative HR expression (Spearman's rho: 0.667, p < 0.005), the second model included the factor with the lowest p-value (cHR), which resulted in being significant after covariation (

Stratified analysis allowing the identification of risk modification factors
Previous analyses resulted in the inclusion of different factors according to outcomes. However, histological grade and Ki-67 proliferative index are known risk factors in patients with luminal breast cancer. Aiming at a greater segregation of risks, stratified (hierarchical) analyses were carried out to study the effect of modification by other variables. As the semi-quantitative expressions of PR and cHR were independent variables for both outcomes, the hierarchy started with them.  Starting with the semi-quantitative PR expression level, both Ki-67 and histological grade showed some stratification only in patients with tumours classified as ≤1+. In preliminary analyses, G2 and G3 showed higher overlap for both outcomes. Therefore, these categories were grouped. The Ki-67 level showed significance for both DMFS (Log-Rank; p = 0.003) and OS (Log-Rank; p = 0.007), and histological grade for both DMFS (Log-Rank; p = 0.022) and OS (Log-Rank; p = 0.012) as well. As no interaction was observed (p > 0.05), conventional Cox regression was carried out with the stepwise method. The histological grade (G2/G3 versus G1) resulted in being independent for OS (adjusted HaR   Starting with cHR, both Ki-67 and histological grade showed some stratification only in patients with tumours classified as ≤4+. In preliminary analyses, G2 and G3 showed higher overlap for both outcomes. Therefore, these categories have been grouped. The Ki-67 level showed significance for both DMFS (Log-Rank; p = 0.020) and OS (Log-Rank; p = 0.020), and histological grade for both DMFS (Log-Rank; p = 0.047) and OS (Log-Rank; p = 0.017) as well. As no interaction was observed (p > 0.05), conventional Cox regression was carried out with the stepwise method. The histological grade (G2/G3 versus G1) resulted in being independent for both DMFS (adjusted HaR    None of the classifications by stratification shows 100% agreement with the other. For this reason, all were analysed. As the subsequent classification by cHR and histological grade followed the order of factors found in the initial analysis, its value in multivariable analysis has always been calculated in further analysis. Triple-negative cancer patients were added as the third level to each category to identify the best risk stratification. As a comparison, the classification of tumours as luminal A and B was according to the expression of PR and Ki-67 (luminal A: PR≥20% and Ki-67<14%; luminal B: PR<20% and/or Ki-67±14%). As there was 100% agreement between the cut-off points obtained (>12%) and that found in the literature (>14%) (Cohen's Kappa: 1.000, p < 0.005), the former was maintained. As most medical records did not report the percentage of PR-expressing tumour cells, this cut-off point was replaced by PR>1+. Alternatively, the same classification was carried out but with a modified Ki-67 cut-off to 20% as per the modern guidelines.
The stepwise model retained the hierarchical risk stratification by Ki-67 and PR (Tables 5 and 6). Although the classifications similar to those currently accepted are significant, the classifications previously obtained proved to be superior and capable of identifying patients with luminal breast cancer whose prognosis is similar to that of patients with triple-negative breast cancer (Tables 5 and 6).   The KM curves of the three fittest approaches are shown in Figures 6-8.    Subsequently, multivariable analyses were carried out to test the possibility of performing clusters. Both risk classification strategies proved to be significant for DMFS and OS. Through repeated contrast, it was possible to observe a clear distinction between the second and third groups for both strategies in the two outcomes tested (data not shown). Therefore, the first two and last two groups were grouped. These classifications have a high, but incomplete, degree of agreement (Cohen's Kappa: 0.876, p < 0.0005) with each other. Figures 9 and 10 show the survival curves including triple-negative patients also on an adjuvant regimen only.

Adjustments for treatments do not change the prognosis value of risk ratings
Finally, it was analysed whether treatments could be a confounding factor. But first, it was tested in which context chemotherapy results results in a better prognostic for patients with luminal cancer in order to improve the classification of systemic treatment.  Patients with luminal breast cancer from an adjuvant regimen were segregated according to biomarkers and it was evaluated whether chemotherapy implies a better prognosis, since analyses with all patients did not demonstrate a potential benefit of chemotherapy. As it may be an effect modification, the * operation was carried out between pN (N0, N1 and N2/N3) and chemotherapy (No and Yes).
As chemotherapy was not significant in the models to validate its interaction with lymph node metastasis, subgroup analyses were carried out. In fact, chemotherapy was observed to be the only factor associated with OS (HaR: 0.263 (0.092-0.752) p = 0.013) in patients with pN2/N3 and Ki-67>12% disease, with a trend for DMFS in association to radiation therapy for DMFS (adjusted HaR: 0.346 (0.115 -1.045) p = 0.060).
Therefore, systemic treatment was re-categorised based on the previous results, assigning appropriate treatment classifications to patients whenever they received hormone therapy, and they received chemotherapy if and only if they had pN2/3 disease and Ki-67>12%, which resulted superior to the initial classification, according to the St Gallen guidelines.
Finally, models were adjusted to include adjuvant regimen-only treatments. For this, time from the end of the adjuvant therapies (chemotherapy or radiotherapy) until the event or censorship was considered; patients who developed events during adjuvant therapy (chemotherapy or radiotherapy) were excluded. A total of 419 patients were included in the analyses. Two models using the two-block approach were built. The first model included all risk ratings in the first block; the second model was designed excluding the first block and the risk classification obtained in the first model.
The stratifications with Ki-67 in the first hierarchical level continued to be statistically superior to the others, even compared to those described in the literature, with change from PR, at the second hierarchical level referring to DMFS, to cHR, at the second hierarchical level, and to OS (Tables 7 and 8). Nonetheless, the second model for both DMFS and OS agreed that the sequential risk stratification by histological grade, Ki-67 and cHR as independent factors is superior to the others (Tables 7 and 8).

Discussion
The identification of luminal breast cancer cases with a high risk of metastasis development is critical for a therapeutic decision [10,17]. However, the only prognostic and predictive test currently available with grade 1A evidence for both therapeutic decision and staging is the Oncotype Dx genomic test [49][50][51], which is expensive.
Studies have shown consecutive risk stratification in luminal cancer cases by Ki-67 [9,10], with subsequent prognostic gain by segregation according to the semi-quantitative expression of PR [52], but the unification with the histological grade remains a difficulty [10,17,38]. Ehinger et al [53] observed that, in patients with luminal ER + tumours, the semi-quantitative expression of PR and Ki-67 discriminates prognosis only in moderately differentiated (G2) tumours. Furthermore, although other studies show the Ki-67 index as a risk segregation criterion in G2 tumours [12,54], Liang et al [55] observed similar prognosis between patients with G3 tumours and high Ki-67 expression in G1 and G2 tumours regarding recurrences. We were able to incorporate these three factors but only in patients of adjuvant regimen, with consecutive hierarchical levels by histological grade, Ki-67 and, finally, the semi-quantitative expression of both hormone receptors, or just of the progesterone receptor. Even so, the stratification path that the initial analysis pointed to was another, promoting consecutive stratification by cHR and later by histological grade, showing similar prognosis between G2 and G3 diseases with low expression of both hormone receptors. Notwithstanding, this stratification did not result in a good risk segregation for DMFS in patients of adjuvant schema as we observed for consecutive stratification by histological grade, Ki-67 proliferation index and, finally, cHR.   Interestingly, all classifications obtained in the analysis turned out to be superior to the currently accepted luminal breast cancer classification (luminal A and B) [10,17]. However, due to the limitation in reporting the expression of hormone receptors, only one approximation was made. In fact, the used PR >1+ category represents tumours with more than 10% expressing cells [46,47], which is biased while making proper comparisons.
Even though the classification incorporating PR levels proved to be interesting, the classifications involving cHR respected some theoretical foundations. For example, luminal tumours expressing only the PR are more aggressive, being a factor of poorer prognosis for relapses and death, even when compared to ER + /PRcases [56][57][58][59][60]. Additionally, a recent meta-analysis demonstrated that low ER-expressing (1%-10%) tumours have an impaired endocrine response [61], a typical characteristic of more aggressive luminal tumours [62]. Thus, classification considering cHR naturally classified tumours with low ER expression and expression of only one hormone receptor as similar, although this approach is not largely validated.
Another important theoretical factor is how the interaction between the receptors takes place. In oestrogen-responsive tumours, PR is expressed as a response to the latter via the ER [63], with subsequent modification of the ER interaction network and expression of different target genes in a feedforward fashion [64]. In line with this, we observed that low ER-expressing tumours are likely to express low PR levels, whereas high ER-expressing tumours express high levels of PR, suggesting that this effect probably occurs in a dose-dependent manner. Thus, the semi-quantitative expression of both HRs could respond for part of the highly intricate biology of breast cancers, as we observed for both studied outcomes, and it respects the fact that the qualitative expression of HR is a more important prognosis factor for OS.
Genetic studies with different panels show segregation of the histological grade into two genetic grades (low and high), which have a higher prognosis value than the histological grade [65]. Interestingly, there is a high degree of agreement between the histological grade G1 and the low genetic grade and the histologic grade G3 with the high genetic grade, but high segregation in the histological grade G2 between these two genetic grades [66,67]. In fact, the histological grade G2 is the classification with the greatest disagreement among the three [68][69][70], and some studies suggest that the proliferative index can promote a prognosis distinction precisely in G2 [12,54].
A study by Sotiriou et al [71] observed that low and high genetic grades can segregate patients with PR + tumours according to endocrine responsiveness and cell proliferation, whereas the low genetic grade has lower proliferation and greater endocrine responsiveness. We observed that the semi-quantitative expression made efficient segregation of the risk factor, but only in G1/G2 patients with a high Ki-67 proliferative index, which finally defined the luminal patients in two groups, which is in harmony with the study by Arima et al [72], in which they observed that in luminal patients (HER2 -) the expression of PR does not add a prognosis factor in patients with low proliferative index by Ki-67, but the high expression of RP (≥20%) partially counterbalances the negative effect of the high proliferative index by Ki-67, as we observed in this study.
The low sample number of patients with G1 and G3 tumours prevented us from evaluating whether these other two factors would cause any segregation of risks. However, Ehinger et al [53] observed that patients with luminal breast cancer (ER + /HER2 -) of histological grade G3 have a similar prognosis factor, regardless of whether they are classified as luminal A or B based on the Ki-67 proliferative index and qualitative PR expression. In fact, there is a consensus that high histological grade (G3) is a factor of worse prognosis in luminal A patients [10].
Furthermore, an expression of both hormone receptors, not just PR, correlates with the Oncotype Recurrence Score [35,37,73], and has a prognosis value for distant metastases by the IHC4 predictive model [74], which has a similar prognosis value to Oncotype Dx for early metastases (up to 5 years) [74][75][76]. Thus, there is more than enough scientific support to analyse the semi-quantitative expression of both receptors as a prognosis factor, and it is versatile in allowing both risk segregation for distant metastasis and potential staging for overall survival, unlike the qualitative expression which is a known staging factor for overall survival [25], as we also observed in the initial analyses.
In this study, the histological grade and the Ki-67 proliferative index proved to be mutually exclusive in several analyses. One possibility is the correlation that exists between them [77], because the mitotic index is one of the classification criteria of the Nottingham system [11,45]. Another possibility is the changes that neoadjuvant chemotherapy causes in the tumour microenvironment and in the selection of subclones with different phenotypes, causing changes in the expression of hormone receptors, mitotic index and Ki-67 [78][79][80][81][82][83], which may explain why it was possible to incorporate histological grade and Ki-67 only in patients on the adjuvant regimen, but not in the group including those on the neoadjuvant regimen.
Although showing the superiority of prognosis segregation in the incorporation of the three analysed parameters (histological grade, Ki-67 and semi-quantitative expression of hormone receptors), having presented other stratification models is interesting for some situations. For example, the treatment guidelines approved for the public health system in Brazil make the assessment of histological grade and identification of the three important tumour receptors mandatory, but not Ki-67 [84].
An interesting factor related to the study of these biomarkers is the potential benefit of chemotherapy. Although the focus was not the discovery of a predictive classification system for the benefit of chemotherapy, it was possible to observe that Ki-67 is the best predictive factor for this response, but only in patients with pN2/N3 disease. This is in agreement with data in the literature, which show that Ki-67 is the best predictor of response to chemotherapy in patients with luminal breast cancer [85,86]. Interestingly, we could observe that patients with G1/G2 Ki-67>12%/cHR≤4+ or G3 disease show a modification of effect by chemotherapy when there is the involvement of four or more lymph nodes (pN2/N3) for OS. In fact, there is evidence that patients with tumours of good biological risk (low grade, low proliferation and increased expression of hormone receptors) have little or no benefit from chemotherapy, even in the presence of metastatic lymph nodes, but the contrary is true [62]. Thus, this hierarchical classification can even help in clinical decision-making for treatment.
This study has several limitations, as the long elapsed period, which implies the incorporation of new therapies. Other limitations are the possibility of IHC analysis artefacts [4], which we did not have access to, and the reproducibility and feasibility of different antibody clones [4,[87][88][89][90], which were used by the Institution, that could dampen our results. Also, the low number of patients representing certain histopathological and immunohistochemical features, such as G1 and G3 tumours, and tumours with low expressions of hormone receptors and/or only one hormone receptor led to results with very wide confidence intervals due to the low number of events observed in these categories.
Regarding Ki-67, we analysed only one dichotomisation, according to the statistical methodology used. However, it is known that there are multilevel categorisations of Ki-67 that better relate to the development of relapses and response to chemotherapy [17,38,62,85,86].
In addition, we could not observe the median time in most analyses, being observed practically only for the second category of classification of luminal patients of all reclassifications, but not for the first category or patients with triple-negative tumours, which makes the results on prognosis factors dependent on the survival curves of this cohort.
But the main limitation is the classification of the semi-quantitative expression of hormone receptors. Due to the lack of standardisation of the semi-quantitative expression of these, which has recently been incorporated into the guidelines of the public health system in Brazil [84], we needed to standardise the reports to the ordinal system of crosses by Sannino and Shousha [46,47]. This leads to the loss of information that more refined systems provide, such as Allred or H-score systems [91,92]. Despite these limitations, our results do not greatly contrast with those in the literature.
A study with a large cohort and improved immunohistochemical data (percentage and intensity) could show the validity of the presented approach and whether it is possible to integrate histological grade and Ki-67, along with semi-quantitative expression of hormone receptors, to identify patients at high risk of developing distant metastases and death.

Conclusion
The different risk stratifications prove to be advantageous when there are limitations of information obtained by IHC techniques according to clinical practices adopted in regions with few resources, such as in low-and middle-income countries. The fact that risk stratification with better identification of high-risk patients, similar to those with triple-negative cancer, was generated by the incorporation of all the variables included attests to the need to include a detailed analysis of these factors in the routines and guidelines of countries such as Brazil.