We used a previously established cohort of 683 individuals who underwent radiotherapy and/or metastasectomy for BrM from 2008-2018 at Sunnybrook Health Sciences Centre in Toronto, Canada. Details regarding the cohort and baseline characteristics of included patients have been previously outlined (9). All factors found to be significantly associated with bsPFS or OS in the Sperduto et al. (8) or Gao et al. ( 9) studies were included in this analysis. These included KPS, age, tumour receptor status, presence of leptomeningeal disease, presence of neurological symptoms, number of BrM, location of extracranial metastases, and initial treatment modality for BrM. As KPS was not originally studied in our cohort, methods for imputing KPS are outlined in Supplementary Table 1 .

Given the importance of the BreastGPA in clinical prognostication of OS in BCBrM, we sought to include all prognostic variables that are included in the BreastGPA and treat them as similarly as possible to the way they were treated in Sperduto et al. ( 8 ) That paper cohorted KPS into cohorts of 90-100, 70-80, and ≤60. Our dataset was felt not to be large enough and available estimates of KPS not precise enough to be able to do useful analysis on a three-cohort KPS division as done in Sperduto et al. Accordingly, we mimicked their KPS ≤60 cohort and combined their KPS 90-100 and KPS 70-80 cohorts into a single KPS >60 cohort for our analysis.

For the purposes of our study, we defined bsPFS as the length of time from the date of diagnosis of brain metastases to the date of either disease progression in the brain (as determined based on a combination of radiologic and clinical assessment) or death due to any cause, whichever was sooner. Overall survival (OS) was defined as the length of time from the date of diagnosis of brain metastases to the date of death due to any cause.

The Kaplan-Meier method (20) was used to estimate both bsPFS and OS, with comparison of the low (≤60) versus high (>60) KPS groups using the log-rank test. Univariable and multivariable analyses of bsPFS and OS were performed with Cox proportional hazards models (21). All variables which were prognostic for bsPFS and/or OS in a univariable model (p-value <0.05) were aggregated into a multivariable Cox proportional hazards model (21) for that outcome. Patients for whom data was missing for any of the statistically significant variables from the univariable analysis were omitted from the multivariable analysis. Variables with p-values <0.05 in the multivariable model of an outcome were considered significant predictors of that outcome. All statistics were computed using RStudio Version 1.4.1106.

This study received approval from Sunnybrook Research Institute’s Research Ethics Board.

Data regarding 13 explanatory variables of interest were summarized ( Table 1 ). Information about patient age and initial locoregional treatment modality for BrM was available for all patients. Values for most other explanatory variables were available for 95% or more of the patients in the dataset; the exceptions were tumor receptor status, presence of multiple BrM, and KPS score, with 14%, 28%, and 52% missing data respectively. The great majority of patients (91%) had extracranial metastatic disease, with bone metastases being the most common individual site (68%). 25% of patients had leptomeningeal disease. Patients both above and below age 60 were well-represented (40% and 60% respectively). In our cohort, 36% (n=247) of patients had HR+HER2- breast cancer, 28% (n=189) had HER2+ and 22% (n=153) had triple-negative breast cancer. Most patients (73%) underwent whole-brain radiotherapy (“WBRT”), and a minority received SRS or required neurosurgery (19% and 10% respectively). Notably, 38 patients (6%) received more than one initial therapy for BrM. Table 1 .

Some patients’ charts did not contain sufficient information to determine whether their KPS was ≤60 or >60 at the time of diagnosis of brain metastases (see ‘Supplementary Methods’ section below). Characteristics were compared between the “KPS cohortable” patients to whom a KPS cohort could be assigned and the “KPS not cohortable” patients to whom a KPS cohort could not be assigned (see Table 1 above). There were some statistically significant differences between the two groups; the KPS cohortable patients were more likely to have bone (73% vs 63%), liver (58% vs 48%), and lung (64% vs 49%) metastases. They were also more likely to undergo treatment with SRS (23% vs 15%), and to be free of neurological symptoms at the time of diagnosis of BrM (19% vs 16%). Finally, two groups had slightly different distributions of tumour subtypes, with KPS cohortable patients slightly more likely to have HR+HER2- (39% vs 34%) and HER2+ (30% vs 26%) disease compared to their KPS non-cohortable counterparts.

Within the KPS cohortable group, there were important statistically significant differences between the patient populations in the high-KPS and low-KPS cohorts. Patients with high KPS were more likely to be under the age of 60 (64% vs 48%) and more likely to have lung metastases (69% vs 52%); they were also more likely to have asymptomatic BrM (24% vs 6%) and more likely to receive SRS (31% vs 6%) compared to the low-KPS cohort.

Overall median bsPFS for the population was 9 months (95% CI: 8-10 months) with median follow-up 13 months (95% CI: 10-15 months). A univariable analysis identified presence of leptomeningeal disease, KPS ≤60, presence of neurological symptoms, and not undergoing SRS as significantly associated with shorter bsFPS (see Table 2 for details). Of these, KPS ≤60 was most strongly associated with shorter bsPFS with a hazard ratio of 1.64 (95% CI 1.16-2.33). KM curves for bsPFS by KPS are shown in Figure 1 .

Median OS in the overall population was not reached; median follow-up for OS was 7 months (95% CI: 6-9 months). Presence of lung metastases, liver metastases, leptomeningeal disease, and neurological symptoms at time of diagnosis of BrM were significantly associated with shorter survival, as were KPS ≤60, age ≥60, and not receiving SRS or neurosurgery ( Table 2 ).

The only variable significant for shorter bsPFS in the multivariable analysis was KPS ≤60 (HR 1.86, 95% CI 1.20-2.88), and the only two variables significant for shorter OS in the multivariable analysis were KPS ≤60 (HR 2.95, 95% CI 1.55-5.58) and the presence of liver metastases (HR 1.95, 95% CI 1.06-3.60).

Performance status has been repeatedly identified as an important marker of prognosis among patients with breast cancer (8, 22–24). Our study confirms that the prognostic significance of KPS extends to patients with brain metastases from breast cancer, given that a KPS ≤60 was significantly associated with a shorter bsPFS and OS.

Once KPS was accounted for, none of the other variables examined were significantly associated with bsPFS in our multivariable analysis. This is likely due to an intimate connection between KPS and other variables that are associated with bsPFS (9). Increasing age is associated with higher burdens of comorbidity and greater frailty, which tend to lead to diminishing functional abilities and therefore lower KPS. Similarly, the presence of neurological symptoms at diagnosis often heralds a greater need for support in activities of daily living or even confusion, both of which diminish KPS. This hypothesis is supported by the statistically significant differences in prevalence of age ≥60, neurological symptoms, and lung metastases between high- and low-KPS cohorts in our dataset (see Table 1 above).

This study confirms the findings of Sperduto et al. that low KPS is a negative prognostic factor in patients with breast cancer and BrM (8) in a dataset which included a significant proportion of patients with no neurological symptoms at the time of diagnosis of BrM (17%) as well as of patients with leptomeningeal disease (25%). Neither of these latter populations are well represented in the patient population examined by Sperduto et al. ( 8).

Like in Sperduto et al, triple-negative breast cancer subtype and advanced age were associated with shorter OS and HER2-positive breast cancer was associated with longer OS, albeit only in univariable analyses. Given that these variables were previously shown in the larger Sperduto et al. dataset to be associated with OS in a multivariable analysis that included KPS (8), we suspect that our study was underpowered to confirm this.

One noteworthy difference between the results of our study and that of Sperduto et al. is the impact of extracranial metastasis on OS. In the latter study, the presence of any extracranial metastasis was a negative prognostic marker, albeit a less important one than tumour receptor status or KPS. The presence of any extra-cranial metastatic disease was not significantly associated with OS in our study, possibly because the small proportion (7%, n=49) of patients who presented with brain-only metastatic disease limited our statistical power to detect such an association. In contrast to the BreastGPA, we were able to refine our analysis by metastatic site. The presence of lung and/or liver metastases were individually prognostic for shorter OS in univariable analyses, and the presence of liver metastases was significantly associated with shorter OS in a multivariable model. The impact of location and burden of extracranial metastatic disease on clinical outcomes requires further study in larger datasets.

Unfortunately, there was a relative scarcity of data suitable for KPS imputation in our study. More than half of the patients did not have enough information available in their medical records to confidently assign them to the low- or high-KPS cohort. The omission of these patients created an “excluded middle” of patients whose absence created a dataset weighted towards high and low KPS, emphasizing the effects of KPS on the outcomes studied. Nonetheless, the distribution of population characteristics between the study population as a whole and those for whom a KPS cohort was available was similar ( Table 1 ). Hence, our inability to cohort all patients by KPS is unlikely to have meaningfully altered our results.

Another important caveat is that only a minority of the patients who were successfully cohorted by KPS had explicitly documented performance statuses on any clinically validated scale around the time of diagnosis of BrM. For the remainder, a KPS was calculated indirectly from a PPS. This imputed PPS was often based on only one or two of the criteria evaluated in the PPS metric, with information on the remaining criteria being unavailable. It is likely that some of the patients’ imputed PPS scores would have been different had more complete information been available in the records.

It is also noteworthy that KPS estimation could not be done in a blinded fashion. Bias introduced by the unblinded nature of assigning KPS was mitigated by the adoption of strict criteria delineated in the ‘Supplementary Methods’ section. Finally, the timing of KPS estimation was not perfectly uniform as outlined in Supplementary Table 1 . Other caveats discussed in Gao et al. regarding the single-academic-centre, retrospective nature of this study and the challenges of applying such a study to other centres with different practice patterns and demographics (9) remain relevant in our study as well.