arising from Elisa Agostinetto et al. NPJ Breast Cancer https://doi.org/10.1038/s41523-022-00452-8 (2022)
In a recent paper in NPJ Breast Cancer, Agostinetto et al.1 demonstrated the poor concordance between recently improved survival data for HER2-positive early-stage breast cancer with outcomes predicted by PREDICT 2.1. We replicated these findings in large-scale cohorts extracted from the UK and US patient registries and demonstrated that a publicly available machine learning-based prognostic model provides improved predictive accuracy.
We do need improved prognostic tools that issue reliable predictions across all patient subgroups, and so we developed a publicly available predictive model, Adjutorium (https://adjutorium-breastcancer.herokuapp.com/), using an advanced automated machine learning algorithm to predict individual breast cancer mortality, all-cause mortality and benefit from adjuvant therapies2. The model makes predictions based on seven patient-specific variables: age at diagnosis, ER status, HER2 status, tumour grade, number of lymph nodes, tumour size, and whether the tumour was detected via routine screening or symptoms. We trained and internally validated the model using data for 395,862 patients from the UK National Cancer Registration and Analysis Service (NCRAS). Because of differences in drug access and practice, we then externally validated the model using data for 571,635 patients from the US Surveillance, Epidemiology, and End Results (SEER) Program.
Our validation experiments show that PREDICT exhibits patterns of survival underestimation that bear some resemblance to those reported by Agostinetto et al. In particular, PREDICT underestimated 5-year survival in patients with HER2-positive breast cancer by 16% (Fig. 1). On the contrary, Adjutorium underestimated survival in the same cohort by only 0.9%. Unlike the analysis conducted by Agostinetto et al. using the ALTTO data, we found that PREDICT underestimates survival in HER2-positive patients more severely than in HER2-negative patients, in which PREDICT underestimated survival by 1%. These discrepancies might be a result of the differences in the time period over which data was collected in our analysis cohort compared to the ALTTO trial (both the internal and external data sets comprised breast cancer diagnoses spanning the years 2000–2016), as well as the differences between the treatment regimens used in our observational cohort and those used in the trial data. However, the observation that PREDICT underestimates survival in HER2-positive patients remains consistent across both studies.
The prognostic accuracy improvements achieved by Adjutorium were not just in the HER2-positive patients but also across all subtypes of patients stratified by HER2 and ER status (Fig. 2). Most notably, in patients with ER-positive breast cancer, survival estimates of PREDICT were underestimated by 23% as compared to 7% for Adjutorium. In the overall population of patients involved in our large-scale study, PREDICT underestimated survival by 5.47% as opposed to 0.94% for Adjutorium. In both the internal and external validation cohorts, Adjutorium and PREDICT v2.1 disagreed on the treatment decisions. In all comparisons, Adjutorium showed significantly greater ROC curves and better concordance with the population data. Treatment decisions informed by Adjutorium are less likely to over- or under-treat patients2. The improved accuracy of Adjutorium is likely a result of using flexible (nonparametric) machine learning models, more recent patient data and a larger sample size compared to PREDICT.
In order to improve accessibility of our model, we provided an easy-to-use online tool for breast cancer prediction based on the Adjutorium model (https://adjutorium-breastcancer.herokuapp.com/), where patient features can be easily input to a graphic interface that survival time with different treatment options. The interface style will be familiar to users of PREDICT v2.1, but ours is an open-source software that enables other researchers to easily re-fit the model as more data becomes available.
Validation of prognostic models using data from clinical trials is crucial for evaluating the clinical utility of these models in informing therapeutic decisions. Both PREDICT v2.1 and our model were derived from retrospective data with confounding factors that influence both patient outcomes and treatment assignments. Some of these confounding factors might be unobserved (e.g., patient-level socio-economic variables), which might undermine the quality of prognostic predictions issued by these models. This might be a possible explanation for why PREDICT exhibits a significant underestimation of survival outcomes. Clinical trials resolve any biases induced by confounding by randomising treatment assignments. Hence, it would be very useful if Dr. Agostinetto and her colleagues would consider validating Adjutorium in the ALTTO trial data to test whether its prognostic gains still hold in patient data that does not suffer from confounding.
References
Agostinetto, E. et al. PREDICT underestimates survival of patients with HER2-positive early-stage breast cancer. NPJ Breast Cancer 8, 87 (2022).
Alaa, A. M., Gurdasani, D., Harris, A. L., Rashbass, J. & van der Schaar, M. Machine learning to guide the use of adjuvant therapies for breast cancer. Nat. Mach. Intell. 3, 716–726 (2021).
Acknowledgements
A.L.H. is funded by the Breast Cancer Research Foundation and Friends of Kennington Cancer Fund. M.v.D.S. directs the Cambridge Center for AI in Medicine, which is funded by AstraZeneca and GSK.
Author information
Authors and Affiliations
Contributions
A.M.A. conducted the numerical experiments. A.M.A., A.L.H. and M.v.D.S. collated and wrote the paper.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Alaa, A.M., Harris, A.L. & van der Schaar, M. Matters Arising: PREDICT underestimates survival of patients with HER2-positive early-stage breast cancer. npj Breast Cancer 9, 13 (2023). https://doi.org/10.1038/s41523-023-00514-5
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41523-023-00514-5
This article is cited by
-
Reply to: PREDICT underestimates survival of patients with HER2-positive early-stage breast cancer
npj Breast Cancer (2023)