Calculating Breast Cancer Recurrence Risk with Artificial Intelligence
In this article
‘Better’ prediction of recurrence risk in early-stage breast cancer
HR-positive (HR+), HER2-negative early-stage breast cancer accounts for approximately 70%of all breast cancers. The most difficult decision in clinical practice is correctly determining, in the adjuvant setting, whether chemotherapy is really necessary . ‘Undertreatment’ may increase recurrence risk, while ‘overtreatment’ may impose unnecessary toxicity.
Although current genomic tests (e.g., Oncotype DX) fill an important gap, they generally remain within an AUC 0.75–0.80 range and make limited use of the tumor architecture/microenvironment information hidden in pathology slides. The RF5-03 presentation brings forward an approach combining digital pathology + clinical data + genomic scores in a single model using deep learning-based artificial intelligence (AI) , supported by prospective validation data.
Note: This content is not medical advice. Individual prognostic decisions should be made by the oncology team in the patient’s full clinical context.
p<0.001
Genomic test
clearer discrimination
HR 3.1 (2.5–3.8)
‘Where’ and ‘as what’ was this study presented?
The findings were shared at the San Antonio Breast Cancer Symposium (SABCS) 2025 meeting, in Rapid Fire Session 5 under the code RF5-03 . The conference abstract appeared in a supplement to J Clin Oncol (suppl): J Clin Oncol. 2025;43(suppl 17):RF5-03.
Study design
This is not a ‘drug trial’; it is a prediction model study. Researchers took, from patients diagnosed with HR+/HER2− early-stage breast cancer, pathology slides and digitized them, then entered them with clinical information (age, nodal status, etc.) and a genomic risk score into the same model. The aim was to see how well the model could discriminate ‘the probability of recurrence over the next 5 years’.
A total of n=2,512 patients were used: the model first ‘learned’ from n=1,200 patients, then, in n=1,312 patients, it was independently validated. Median follow-up was reported as 4.1 years .
Model performance: AUC comparison
Bar: AUC (0–1) • Thin markers: 95% CI for AIWhat does this difference mean ‘in practice’?
An increase in AUC from 0.78 to 0.88suggests that the model can distinguish ‘high risk from low risk’ more sharply. In the example given in the presentation, 5-year recurrence among patients above a certain threshold was 25% , compared with 5% in the low-risk group. This distinction may affect clinical discussions about who would really benefit from chemotherapy in particular.
The Most Clinically Useful Signals from the Presentation
• In the high-risk group: HR 3.1 (95% CI: 2.5–3.8)
• 5-year recurrence: 25% (high) vs 5% (low)
• Performance in the node-positive (+) subgroup: AUC 0.90
• External validation (Asian cohort, n=300): AUC 0.86
Why Might Pathology ‘Change the Game’?
Pathology slides do more than establish a ‘diagnosis’; they also contain clues to tumor behavior, such as cell density, growth pattern, stromal structure and the distribution of immune cells. AI’s promise lies in standardizing these visual signals and adding them to prognostic assessment.
Calibration and reliability
To assess ‘how closely the model’s predictions matched reality’ (calibration), the presentation reported a Brier score of 0.12 (lower is better). The difference in performance between ethnic groups was noted to be limited (~2%).
‘Adverse Effects’: The Right Framework
This AI model is not a treatment; it does not directly cause adverse effects. However, because it may influence decisions, the key clinical concern is the possibility of risk misclassification (unnecessarily intensive treatment or undertreatment). Model outputs must therefore be considered not in isolation, but alongside a physician’s assessment .
Practical message
RF5-03 offers a strong signal that combining ‘genomics + clinical data + pathology’ may improve recurrence prediction in HR+ early-stage breast cancer. If similar validations emerge from other centers, such AI models could make adjuvant decisions more selective and personalized.
Source (presentation abstract): Lazar AJ, Lehman CD, et al. ‘AI-based recurrence prediction model in HR+/HER2− early breast cancer: Prospective validation.’ J Clin Oncol. 2025;43(suppl 17):RF5-03.
Was this article helpful?
Your feedback helps us improve our content.
Reader comments
Share your comment
Your name and comment are published only after administrator approval. Your email address is never shown to visitors.
Be the first to comment.