replying to Ahmed M. Alaa et al. npj Breast Cancer https://doi.org/10.1038/s41523-023-00514-5 (2023)

Refining prognostication in patients with early breast cancer represents one of the main challenges of modern oncology. In our prior analysis within the phase III ALTTO trial, PREDICT showed to underestimate survival of patients with HER2-positive early breast cancer. Alaa and colleagues tested PREDICT in patients from the UK National Cancer Registry. Consistently with our findings, PREDICT highly underestimated the survival of patients with HER2-positive early breast cancer. In this manuscript, we compare the findings of both studies, and we discuss the future perspective of prognostication in breast cancer.

We read with great interest the Matters Arising manuscript by Alaa and colleagues1, in which the authors present a new prognostic model guided by artificial intelligence to improve the accuracy of survival prediction in patients with early breast cancer.

In their study, Alaa and colleagues analysed the data of 395,862 patients extracted from the UK National Cancer Registration and Analysis Service (NCRAS)1. First, they conducted a validation experiment to test the prognostic performance of PREDICT in this cohort; secondly, they tested a new artificial intelligence-based tool (i.e., Adjutorium, available at https://adjutorium-breastcancer.herokuapp.com/) to evaluate its prognostic performance in the same cohort1.

The first part of their study, namely the validation of PREDICT in the NCRAS cohort, allows an indirect comparison with our previously published findings2, in which we evaluated the prognostic performance of PREDICT in patients with HER2-positive early breast cancer enrolled in the phase III, randomized ALTTO trial3.

Both our results and those from Alaa and colleagues were consistent in showing that PREDICT underestimates the survival of patients with HER2-positive early breast cancer. When observing the magnitude of this underestimation, we found an absolute difference in 5-year overall survival of −6.69% (95% confidence interval −7.55 to –5.83), compared to −16% in the study by Alaa et al. (Table 1). This difference in the magnitude of underestimation could be attributed to the different population analysed. The cohort evaluated by Alaa and colleagues consists in a large real-world population, with diagnoses of breast cancer ranging from 2000 and 2016, and likely heterogeneous in terms of treatments received. On the contrary, our cohort from the ALTTO trials included patients randomized in a phase III trial enrolling between June 2007 and July 2011.

Table 1 Magnitude of survival underestimation by PREDICT in the study by Agostinetto et al. and in the one by Alaa et al.
Full size table

On the other hand, some differences exist between our study and the one from Alaa and colleagues. First, they evaluated the prognostic performance of PREDICT also in patients with HER2-negative breast cancer (analysis not performed in our study, that included only patients with HER2-positive breast cancer). In HER2-negative breast cancer, PREDICT underestimated survival by 1% only. This finding may suggest that PREDICT still has a role in hormone receptor-positive breast cancers, i.e., the subtype for which this tool was originally developed.

Alaa and colleagues also tested PREDICT in patients with estrogen receptor (ER) positive breast cancer from the NCRAS registry data, apparently regardless of HER2-status. In this sub-cohort, PREDICT underestimated survival. Nonetheless, we believe that it would be more informative to add further granularity in this analysis, by evaluating separately ER-positive/HER2-negative vs. ER-positive/HER2-positive vs. ER-negative/HER2-positive tumours. Of note, these subgroups are associated with different clinical-pathological and molecular characteristics4, and merging all together could easily results in biases. Going even further, additional attention to more specific subgroups, like ER-low5 and HER2-low6 breast cancers, is getting increasing interest and could provide interesting insights.

Furthermore, Alaa and colleagues presented a new prognostic model, Adjutorium, based on artificial intelligence, to improve the accuracy of survival prediction in patients with early breast cancer. Adjutorium was associated with better survival estimation in all the analyzed sub-cohorts (i.e., HER2-positive and ER-positive). In a prior publication by the same team, Adjutorium was compared to PREDICT and showed improved discriminatory accuracy7. These data are encouraging, and support further investigation of this tool. Hence, we fully support the interest of Alaa and colleagues in testing this model in patients enrolled in the ALTTO trial.

All these research efforts underline the relevance of prognostication in patients with early breast cancer. Traditionally, this was considered as an important topic especially for patients with hormone receptor-positive/HER2-negative disease, for whom prognostication could affect the choice of adding adjuvant chemotherapy before endocrine therapy initiation as well as for extended duration of endocrine therapy (Table 2)8,9. However, also in HER2-positive disease prognostication has paramount importance. Indeed, although most patients receive (neo)adjuvant chemotherapy as per current standard of care, prognostication can still impact crucial aspects of patient’s life, including for instance family planning in premenopausal women10. Moreover, an increasing number of clinical trials are testing de-escalation and escalation treatment strategies to optimize the therapeutic approach of patients with all subtypes of breast cancer. It is likely that in the future there will be an increasing incorporation of molecular information to provide reliable prognostic estimates11. Nevertheless, online tools like Adjutorium retain the strengths to be free, publicly available, fast, and easy-to-use in daily clinical practice.

Table 2 Prognostic models for patients with early breast cancer.
Full size table

In the rapidly changing treatment landscape for early breast cancer, the challenge of prognostic models is to allow a promptly adaptation to the use of new therapies and to the subsequent changes in patients’ outcomes.

Reporting Summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.