Machine Learning Model Improves Breast Cancer Risk Stratification

Researchers have developed and externally validated a machine learning clinicogenomic model designed to improve prognostic stratification in patients with early-stage estrogen receptor-positive and human epidermal growth factor receptor 2-negative (ER-positive/HER2-negative) breast cancer. Published in Breast Cancer Research, the retrospective integrative analysis demonstrates that combining a research-grade implementation of the 21-gene recurrence score with immune-related and clinicopathologic features can identify larger groups of patients at low risk of recurrence while maintaining or improving survival estimates compared to the standalone genomic signature.

The study addresses a critical challenge in oncology: how to best integrate heterogeneous biological and clinical data into a single, interpretable model for routine clinical use. While gene expression signatures (GESs) are currently recommended to guide adjuvant chemotherapy decisions in ER-positive/HER2-negative early breast cancer, their clinical utility largely depends on their prognostic value. Previous head-to-head comparisons have suggested that tools integrating clinicopathologic information outperform those based solely on gene expression, yet the optimal method for combining these distinct data sources has remained an open question.

To investigate this, the research team conducted a retrospective cohort study using publicly available and institutional microarray and RNA-sequencing data. The final analysis population included 5,132 patients across four independent datasets: the MetaGxBreast training compendium (n = 1,997), a Gene Expression Omnibus (GEO) dataset (n = 321), the Stockholm Breast Cancer Cohorts (SBCC) dataset (n = 322), and the Swedish Cancerome Analysis Network – Breast (SCAN-B) dataset (n = 2,492). All patients had not received neoadjuvant or adjuvant chemotherapy, with some receiving adjuvant endocrine therapy alone. The primary endpoint was distant metastasis-free survival (DMFS), or a closely related recurrence endpoint where DMFS was unavailable.

The researchers developed a random survival forest (RSF) model, a type of machine learning algorithm suited for survival analysis, which incorporated three key features: a cohort-percentile version of the 21-gene recurrence score, a 14-gene immunoglobulin B-cell (IgG) signature, and a clinical composite risk score (CCRS). The CCRS was constructed using four clinicopathologic characteristics: age, tumour size, histological grade, and nodal status. This integrative approach aimed to capture tumour proliferation, estrogen receptor signalling, immune activity, and standard clinical risk factors within a unified framework.

In external validation, the RSF model demonstrated superior performance compared to the standalone 21-gene recurrence score. The model identified significantly larger low-risk groups, with relative increases ranging from 33.8% to 111.8% across the validation datasets. For instance, in the SCAN-B dataset, the RSF model classified 68.3% of patients as low risk, compared to a much smaller proportion identified by the standalone genomic signature. Despite these larger low-risk groups, the model maintained similar or higher horizon-specific survival estimates, suggesting that the expanded low-risk category did not compromise the accuracy of outcome prediction.

Discrimination metrics, which measure the model’s ability to distinguish between patients who experience earlier events and those who remain event-free, also improved with the RSF model. Across the validation datasets, the model showed absolute increases of 0.035 to 0.058 in the concordance index and 0.030 to 0.064 in the integrated area under the time-dependent receiver operating characteristic curve compared to the standalone 21-gene score. Risk separation, measured by the difference in restricted mean survival time between low- and high-risk groups, was consistently greater for the RSF model, while prediction error remained comparable or lower according to Cox-recalibrated integrated Brier score analyses.

Explainability analyses were conducted to ensure the model’s biological plausibility and transparency. Time-dependent feature contribution analyses indicated that the 21-gene recurrence score and clinicopathologic variables were the primary drivers of early prognostic performance. In contrast, the immune component, represented by the IgG signature, contributed a smaller but more stable effect over time. Partial-dependence plots showed that higher values of the clinical composite risk score and the 21-gene recurrence score shifted predicted survival probabilities downward, whereas higher IgG values were associated with a modest upward shift in predicted survival, reflecting the protective prognostic association of this immune signature.

The study also benchmarked the RSF model against an RSClin-analogous Cox comparator in node-negative patients. The RSClin-analogous model, which integrated the 21-gene recurrence score with age, tumour size, and histological grade, improved discrimination over the 21-gene score alone. The RSF model showed higher discrimination than this clinicogenomic comparator in the MetaGxBreast, SBCC, and SCAN-B datasets, with nearly identical performance in the GEO dataset. Adding the IgG signature to the RSClin-analogous model improved model fit in the training subset and increased the 10-year concordance index in several validation datasets, though this incremental value was cohort-dependent.

Subgroup analyses within the SCAN-B dataset, which had the most favourable clinicopathologic composition, confirmed that the RSF model retained prognostic separation across age and tumour size strata. Interaction analyses did not show evidence that the association between RSF risk and outcome differed by age or tumour size. Similarly, treatment-stratified analyses indicated that the model maintained discrimination in both untreated and endocrine-treated subgroups, although the untreated subgroup had limited event numbers and wider uncertainty. These findings suggest that the model’s performance is robust across clinically relevant subgroups.

The authors highlight several limitations that must be considered when interpreting the results. The study was retrospective and relied on research-grade transcriptomic implementations of gene expression signatures rather than clinically generated commercial assay results. Discordance has been reported between research-based implementations and commercial assays, meaning the model’s outputs should not be interpreted as equivalent to clinically validated risk categories used for adjuvant treatment decisions. Furthermore, the analysis included heterogeneous datasets from different institutions, platforms, and historical treatment periods, which may introduce technical and clinical variability.

Despite these limitations, the findings support the potential value of an explainable clinicogenomic modelling framework for early breast cancer risk assessment. The RSF model’s ability to identify larger low-risk groups while preserving outcome separation suggests it may improve risk stratification by expanding the pool of patients who might be candidates for treatment de-escalation in the future. However, the authors emphasize that these findings are exploratory and require prospective validation in clinically appropriate populations using standardised assay outputs to determine their direct clinical applicability.

The integration of immune-related features into prognostic models represents a significant step forward in understanding the biological determinants of breast cancer recurrence. By combining established genomic signatures with immune and clinical data, the RSF model offers a more biologically grounded approach to risk assessment. This study underscores the importance of multimodal machine learning in leveraging heterogeneous data sources to provide individualised and accurate prognostic information for patients with ER-positive/HER2-negative early breast cancer.

Subject of Research: Development and validation of a machine learning clinicogenomic model for prognostic stratification in ER-positive/HER2-negative early breast cancer

Article Title: Development and validation of a machine learning clinicogenomic model to improve prognostic stratification in ER-positive/HER2-negative early breast cancer

Article References: Sarafidis, M., Sifakis, E. G., Bergh, J., Foukakis, T., & Matikas, A. (2026). Development and validation of a machine learning clinicogenomic model to improve prognostic stratification in ER-positive/HER2-negative early breast cancer. Breast Cancer Research, 28(1), Article 156. https://doi.org/10.1186/s13058-026-02388-4

Image Credits: AI Generated

DOI: 10.1186/s13058-026-02388-4

Keywords: breast cancer, machine learning, prognostic stratification, ER-positive, HER2-negative, gene expression signatures, random survival forest, clinicogenomic model, immune signature, clinical oncology, Development, validation