Jin Feng · Yu Xinmiao · Wang Menghan Department of Breast Surgery, The First Affiliated Hospital of China Medical University
At the 2025 San Antonio Breast Cancer Symposium (SABCS 2025), a reanalysis of the landmark NSABP B-20 trial (Abstract No. RF3-03) was presented, exploring the prognostic value and chemotherapy benefit–predictive potential of a multimodal artificial intelligence (MMAI) model based on digital pathology and clinical data in hormone receptor–positive (HR+), node-negative early breast cancer.The study suggests that MMAI may serve as a low-cost, tissue-sparing risk stratification tool, offering new evidence to inform adjuvant chemotherapy decision-making. Oncology Frontier invited Professor Jin Feng, Professor Yu Xinmiao, and Dr. Wang Menghan from the First Affiliated Hospital of China Medical University to provide an in-depth interpretation of the study, with the aim of offering meaningful insights for clinical practice.
Study Overview
Study Title
Chinese: 基于数字病理的多模态人工智能模型在HR+且淋巴结阴性乳腺癌中的验证：用于预后评估和化疗获益预测—NSABP B-20试验分析
English: Evaluation of a Digital Pathology–Based Multimodal Artificial Intelligence Model for Prognosis and Prediction of Chemotherapy Benefit in Node-Negative, Hormone Receptor–Positive Breast Cancer Patients: Analysis of the NSABP B-20 Trial
Background
In patients with HR-positive early breast cancer (EBC), accurate stratification of distant metastasis (DM) risk is central to adjuvant treatment decisions, particularly regarding the addition of chemotherapy (CT).
A multimodal artificial intelligence (MMAI) approach—integrating histopathologic image features with clinical variables—has previously demonstrated prognostic value in the ABCSG-08 trial, offering a potential low-cost complement to multigene assays.
This study leveraged the landmark NSABP B-20 randomized phase III trial, which enrolled HR-positive, lymph node–negative early breast cancer patients and compared:
- Tamoxifen (TAM) alone
- TAM + methotrexate + fluorouracil (MFT)
- TAM + cyclophosphamide + methotrexate + fluorouracil (CMFT)
The aim was to externally validate the MMAI algorithm and explore its potential role in guiding adjuvant chemotherapy decisions in HR+/LN- early breast cancer.
Methods
Patients from the NSABP B-20 trial were included if they had:
- Digitized H&E-stained surgical slides
- Complete clinical data
- Long-term follow-up information
Using a pre-specified, locked MMAI algorithm, patients were categorized into:
- Low-risk
- Intermediate-risk
- High-risk groups
The primary endpoint was distant metastasis (DM).
1. Prognostic Performance
Only patients in the TAM-alone arm were analyzed. A Fine–Gray competing-risk regression model was used to evaluate the association between MMAI score and DM, reporting subdistribution hazard ratios (sHRs) with 95% confidence intervals (CIs).
2. Prediction of Chemotherapy Benefit
In the full cohort including TAM, MFT, and CMFT arms, MMAI scores were incorporated as either:
- A continuous variable, or
- A binary variable (low risk vs intermediate/high risk)
Interaction between treatment arm and MMAI score was tested using Fine–Gray models.
Interaction analyses were performed in:
- The overall cohort, and
- Age-stratified subgroups (≥50 years vs <50 years), consistent with the original NSABP B-20 follow-up design.
Results
A total of 1,763 patients received MMAI scores and were included in the analysis, representing approximately 75% of the original NSABP B-20 population. The median follow-up was 14.6 years.
Risk Group Distribution
- Low risk: 1,189 patients (67%)
- Intermediate risk: 180 patients (10%)
- High risk: 394 patients (22%)
1. Prognostic Value
In the TAM-alone arm, MMAI score was strongly associated with DM risk.
- Continuous MMAI score: Increasing score was significantly associated with higher DM risk sHR = 1.94 (95% CI: 1.57–2.41, p < 0.001)
- Categorical risk groups: High vs low risk: sHR = 3.97 (95% CI: 2.57–6.16, p < 0.001) Intermediate vs low risk: sHR = 2.78 (95% CI: 1.47–5.23, p = 0.002)
These findings consistently demonstrate the robust prognostic discrimination of MMAI, whether treated as a continuous or categorical variable.
2. Prediction of Chemotherapy Benefit
- Overall cohort: No statistically significant interaction was observed between treatment arm and MMAI score, regardless of whether MMAI was modeled as a continuous or binary variable.
- Age-stratified analysis: A significant interaction emerged only in patients aged ≥50 years (p = 0.01). In this subgroup, 32% were classified as intermediate/high risk by MMAI. Among these patients, adding chemotherapy reduced the 10-year DM rate from 21% to 10% CMFT: 10% vs TAM: 21% (relative reduction of 52%) In contrast, low-risk patients showed minimal difference in 10-year DM rates regardless of chemotherapy use CMFT: 7% vs TAM: 5%
- In patients <50 years, MMAI did not show predictive interaction with chemotherapy benefit, although chemotherapy appeared beneficial across both low- and higher-risk groups.
Conclusions
In patients with HR-positive, node-negative early breast cancer from the NSABP B-20 trial, the MMAI model demonstrated excellent and independent prognostic performance. However, it did not predict chemotherapy benefit in the overall population.
In exploratory age-stratified analyses, MMAI risk classification (intermediate/high vs low) predicted benefit from CMF chemotherapy in patients aged ≥50 years. These findings suggest that MMAI may serve as a low-cost, tissue-sparing alternative to genomic assays, particularly in older HR+/LN- patients, to support adjuvant chemotherapy decision-making.
Expert Commentary
In recent years, multimodal artificial intelligence (MMAI) has advanced rapidly in breast cancer research. Its core concept is to integrate routine digital H&E pathology slides with basic clinical information (age, tumor size, T/N stage, etc.).
Platforms such as Artera have demonstrated that MMAI scores independently predict distant metastasis risk in large phase III trials (e.g., WSG PlanB and ADAPT) involving over 5,000 HR+/HER2- early breast cancer patients. Similarly, the RlapsRisk BC computational pathology model has significantly improved 5-year metastasis-free survival (MFS) prognostication in ER+/HER2- disease.
Against this backdrop, the current study provides an important external validation of MMAI in the well-characterized NSABP B-20 trial, which features long-term follow-up and randomized chemotherapy comparisons. In the TAM-alone arm, MMAI demonstrated strong discrimination of 10-year DM risk, with nearly a four-fold difference in sHR between high- and low-risk groups—underscoring its solid prognostic capability.
A key clinical question is whether MMAI can guide adjuvant chemotherapy decisions. In the overall population, the lack of significant interaction suggests that MMAI is not yet sufficient to guide chemotherapy use across all HR+/LN- patients. However, the exploratory benefit observed in patients aged ≥50 years, where chemotherapy reduced 10-year DM from 21% to 10% in intermediate/high-risk patients, suggests that MMAI may help identify a subset of truly chemotherapy-sensitive individuals in this age group.
This apparent age-related difference aligns with observations from prior studies such as NSABP B-20, TAILORx, and RxPONDER, which have shown varying absolute chemotherapy benefits across age groups. Nevertheless, the evidence here is weaker than that from prospective trials and is based on legacy CMF chemotherapy, necessitating further validation in the context of modern systemic therapy.
Overall, the MMAI model—derived from routine digital pathology and simple clinical data—demonstrates reliable long-term prognostic stratification in HR+/LN- early breast cancer and may offer valuable guidance in resource-limited settings. Future work should focus on validating MMAI alongside established genomic assays within modern treatment frameworks, to determine whether digital pathology–driven multimodal AI can truly help clinicians optimize chemotherapy use—maximizing benefit for high-risk patients while avoiding overtreatment in low-risk individuals.
Professor Jin Feng
Professor Yu Xinmiao
Dr. Wang Menghan
