At the 2025 San Antonio Breast Cancer Symposium (SABCS 2025), a study based on data from the NSABP B-42 and TAILORx trials (Abstract No. RF3-07) reported the development and validation of a multimodal, multitask deep learning model (Clarity BCR). By integrating routine digital pathology images with key clinical variables, this model accurately predicts the risk of late distant recurrence (late DR) in patients with hormone receptor–positive (HR+) early breast cancer and effectively identifies those who may benefit from extended endocrine therapy (EET).The model represents a low-cost, scalable precision stratification tool that supports long-term clinical decision-making. Oncology Frontier invited Professor Jin Feng, Professor Yu Xinmiao, and Professor Cao Yu from the First Affiliated Hospital of China Medical University to provide an in-depth interpretation of the study, with the aim of offering insights for optimizing long-term management strategies in HR+ breast cancer.
Study Overview
Study Title
Chinese: 基于NSABP B-42试验数据训练、TAILORx试验数据验证，用于预测激素受体阳性（HR+）早期乳腺癌患者远期转移风险的多模态-多任务深度学习模型
English: A Multimodal-Multitask Deep Learning Model Trained in NSABP B-42 and Validated in TAILORx for Late Distant Recurrence Risk in HR-Positive Early Breast Cancer
Study Design
Model Development and Background
This study developed a multimodal-multitask (M3T) deep learning model, also referred to as Clarity BCR, designed to integrate whole-slide H&E pathology images (WSIs) with key clinical variables to accurately predict late distant recurrence (late DR) in HR-positive early breast cancer. The model also aims to guide decisions regarding extended endocrine therapy.
The model was trained using data from the NSABP B-42 trial, which investigated extended letrozole therapy, and externally validated in the large, prospective phase III TAILORx trial.
NSABP B-42 Training Cohort and Model Architecture
Study population: Postmenopausal patients with HR-positive (ER+ and/or PR+) early breast cancer who had completed 5 years of standard endocrine therapy without recurrence and were randomized to receive either:
- Extended letrozole for 5 years, or
- Placebo
Primary endpoints:
- Prediction of late distant recurrence (DR) risk
- Survival differences between model-defined high- and low-risk groups (hazard ratio, HR)
M3T (Multimodal-Multitask) Model Design
- Multimodal inputs: The model integrates features extracted from H&E pathology images with clinical variables, including nodal status, surgical approach, and age.
- Auxiliary task: A distinctive multitask learning strategy was employed. During training, the model included an auxiliary task predicting bone mineral density (BMD) T-scores, designed to enhance sensitivity to tumor microenvironment characteristics and latent biological features.
- Primary objective: To evaluate the model’s ability to stratify patients by late DR risk and quantify prognostic differences between high- and low-risk groups.
External Validation Cohort (TAILORx)
Validation population: Patients who completed ≥4.5 years of endocrine therapy and remained recurrence-free at 5 years (late DR analysis cohort, n = 4,300).
Primary endpoint:
- Distant recurrence-free interval (DRFI) / late distant recurrence risk
Results
1. Model Performance and Predictive Value in the NSABP B-42 Cohort
- Superior risk stratification: Compared with single-modality models, the M3T model demonstrated the strongest risk stratification performance in the NSABP B-42 cohort.
- High- vs low-risk differentiation: The M3T model stratified patients into high- and low-risk groups with a 10-year distant recurrence hazard ratio of 5.71 (95% CI: 3.50–9.32, P < 0.001).
- Absolute risk difference: The absolute difference in 10-year distant recurrence rates between the high- and low-risk groups was 7.95% (high risk: 9.63% vs low risk: 1.69%).
Guiding Extended Endocrine Therapy (EET)
The M3T model successfully identified patients who derived meaningful benefit from extended letrozole therapy:
- High-risk group (significant benefit): Extended endocrine therapy reduced the 10-year distant recurrence risk by 4.09% (absolute benefit) HR = 0.614, P = 0.015.
- Low-risk group (minimal benefit): The absolute benefit of extended therapy was only 0.49%, with no statistical significance (HR = 0.664, P = 0.378), suggesting that extended therapy may be safely omitted in this group.
Outperforming Traditional Clinicopathologic Factors
The model reclassified risk in a substantial proportion of patients:
- Node-negative (N0) patients: Approximately 29% were reclassified as high risk, indicating that recurrence risk may be underestimated using conventional factors alone.
- Node-positive (N+) patients: Approximately 18.7% were reclassified as low risk. These patients demonstrated excellent outcomes (10-year distant recurrence rates of only 1.23%–2.73%) and derived minimal benefit from extended therapy.
2. External Validation in the TAILORx Cohort
- Independent prognostic validation: In the large external TAILORx validation cohort (N = 4,203; events = 230), the M3T model retained strong independent predictive power for late DR.
- Long-term risk prediction: The 15-year cumulative distant recurrence rate was: 12.94% in the high-risk group 6.59% in the low-risk group
- Statistical significance: HR = 1.893 (95% CI: 1.413–2.534, P < 0.001)
- Multivariable analysis: After adjusting for age, tumor size, grade, Oncotype DX Recurrence Score (RS), surgical type, and chemotherapy use, the M3T risk label remained an independent prognostic factor (HR = 1.61, 95% CI: 1.14–2.27).
Notably, its predictive performance was independent of Oncotype DX (RS HR = 1.12, 95% CI: 0.59–2.14, P > 0.05).
Conclusions
The Clarity BCR (M3T) deep learning model demonstrated strong prognostic stratification in the NSABP B-42 trial and was successfully externally validated for late distant recurrence prediction in the TAILORx trial.
Compared with genomic assays, the M3T model leverages routine H&E pathology slides and standard clinical data, offering a cost-effective and widely implementable alternative. Its clinical significance includes:
- Precision selection: Identifying low-risk patients—including a subset of node-positive patients—who are unlikely to benefit from extended endocrine therapy, thereby avoiding unnecessary toxicity.
- Therapeutic guidance: Confirming that high-risk patients represent the primary beneficiaries of extended endocrine therapy, supporting long-term adjuvant treatment decisions beyond conventional clinicopathologic factors.
Expert Commentary
Late distant recurrence in HR-positive breast cancer is cumulative, with approximately 50% of recurrences occurring beyond 5 years after diagnosis, making long-term risk management a persistent clinical challenge. Traditional clinicopathologic factors—such as nodal status, tumor size, and histologic grade—have limited predictive power for late recurrence, while genomic assays, although informative, face barriers related to cost, platform variability, and accessibility.
The Clarity BCR model addresses these gaps by applying deep learning to integrate digital H&E pathology images with key clinical variables, using multitask learning to automatically extract high-dimensional features from the tumor microenvironment. This enables the model to capture histologic patterns that are invisible to the naked eye yet closely associated with late metastatic risk.
The NSABP B-42 trial, with its rich clinical, pathologic, and long-term follow-up data, provided an ideal training foundation, while the large, globally representative TAILORx trial served as a robust external validation dataset. The model’s strong performance in both cohorts—particularly its stability in the large TAILORx validation—greatly enhances its clinical credibility.
Importantly, the model’s reclassification of risk has meaningful clinical implications. Approximately 18.7% of node-positive patients were reclassified as low risk, suggesting a biologically indolent subgroup with limited benefit from extended therapy. Conversely, over 29% of node-negative patients were identified as high risk, highlighting a group whose recurrence risk may be underestimated in routine practice. This refined stratification can help avoid overtreatment while ensuring that truly high-risk patients receive intensified monitoring and intervention.
Because the M3T model relies on routine pathology slides and readily available clinical data, it offers clear advantages in cost, scalability, and accessibility, particularly in resource-limited settings. As digital pathology infrastructure continues to expand, such AI-driven models are poised to become integral components of long-term, individualized management pathways for HR-positive breast cancer.
As one of the most significant advances in breast cancer research reported in late 2025, continued refinement and prospective validation of this model may usher in a new era of precision-guided prognostication and treatment decision-making for HR-positive early breast cancer.
Professor Jin Feng
Professor Yu Xinmiao
Professor Cao Yu
