In July 2026, researchers from the University of São Paulo (USP) and the São Paulo Cancer Institute (Icesp) published the study Ensemble machine learning for predicting breast cancer recurrence and mortality using clinical and hemogram data in npj Breast Cancer, a Nature journal.
In the article, the researchers detail the development of BreastOncoPredict, an Artificial Intelligence (AI) software capable of predicting the risk of recurrence or death in patients with breast cancer based on clinical data and a complete blood count available at diagnosis.
C2PO spoke with Dr. Luciana Rodrigues Carvalho Barros, a scientific researcher at Icesp, a C2PO member researcher, and the study’s lead researcher, to learn more about the development of the program.
C2PO: How did the idea of using a complete blood count to develop AI for predicting breast cancer recurrence come about?
Dr. Luciana Barros: The use of blood counts as a source of information for diagnosis was already well established in the literature. However, we decided to use the complete blood count rather than just white blood cell counts.
We started in 2023 by evaluating data from 4,867 patients, with retrospective information from the previous 15 years (since 2008). This was only possible because we had access to data from Icesp, a large cancer treatment center.
We found that the data were consistent over time, meaning that we could rely on them to try to establish correlations.
The software was trained using data from patients at Icesp, and we assessed which biomarkers were the best predictors and which were not relevant to the predictions. Red blood cells turned out to be relevant, which was something we had not expected.
Several different biomarkers were tested until we arrived at 16 features that, when considered together, provided the best prediction accuracy.
One advantage is that all the information is available at the time of diagnosis: clinical stage, patient age, and the complete blood count.
C2PO: How was the software developed?
Dr. Luciana Barros: We trained dozens of supervised predictive models and then selected the best models to make up the BreastOncoPredict AI. We also used techniques to identify which pieces of information were most important in determining the recurrence risk indicated by the AI.
BreastOncoPredict is an explainable AI, unlike a “black box” AI, in which it is not possible to determine which pieces of information are most important. Using this technique, we found that analysis of the red blood cell series in the blood count—not just white blood cells—was important for the AI’s prediction of recurrence.
C2PO: BreastOncoPredict uses complete blood count data without relying on the tumor’s genetic material, as genetic tests do. What are the advantages of this approach?
Dr. Luciana Barros: Genetic tests require a biopsy from each patient, meaning that a tumor sample must be collected, and are therefore very expensive and inaccessible within the Brazilian Unified Health System (SUS). BreastOncoPredict does not require any additional sample collection beyond the blood count that is already performed as part of patient monitoring.
A major strength of this research may be the democratization of access to recurrence prediction for patients with breast cancer within the SUS, an approach that could also be replicated in low- and middle-income countries.
C2PO: What are the expected benefits and impacts of using the software?
Dr. Luciana Barros: The tool could be used by physicians to support treatment decisions and recurrence monitoring, increasing the frequency of screening examinations for patients at high risk of recurrence while sparing low-risk patients from unnecessary examinations.
For public health managers, it could help manage waiting lists and prioritize patients at higher risk of recurrence, improving the efficiency of the healthcare system.
C2PO: The study was retrospective. Will there now be a prospective study for confirmation? What are the next steps to validate and expand its use?
Dr. Luciana Barros: We will conduct a prospective study at Icesp, through which we hope to confirm the accuracy metrics.
In addition, there will be a validation phase at other centers. The idea is to test the software in other SUS hospitals. In fact, an initial collaboration is already being planned with another public hospital in the state of São Paulo. In the long term, however, the goal is to expand to other SUS centers throughout Brazil. Validation in other settings may lead to minor adjustments to the software, making it more accurate.
BreastOncoPredict, in a version still intended exclusively for research purposes, can be accessed through the official website.
