Metaplastic breast cancer (MBC) is a rare but aggressive subtype of breast cancer that accounts for less than 1% of all breast malignancies. It tends to respond poorly to standard chemotherapy and carries a worse prognosis than common breast cancer subtypes, making accurate survival prediction especially important.
This study used machine learning to build survival prediction models for MBC patients, drawing on data from the SEER (Surveillance, Epidemiology, and End Results) database - a large national cancer registry covering millions of patients across the United States.
The best-performing model, CatBoost, was validated on an independent external dataset and showed strong ability to predict whether a patient would survive 1, 3, or 5 years after diagnosis.
Metaplastic breast cancer is unusual because its cells transform (metaplasize) into non-breast cell types such as squamous cells, spindle cells, or bone/cartilage-like tissue. This makes it biologically distinct from typical breast cancers and harder to treat with standard protocols.
Most MBCs are triple-negative - meaning they lack estrogen, progesterone, and HER2 receptors. This rules out targeted hormone therapies and HER2-directed treatments that benefit many breast cancer patients. Chemotherapy response rates are low, and outcomes are worse than for other triple-negative breast cancers.
Because MBC is rare, individual treatment centers see few cases, making it difficult to accumulate enough data for reliable statistical analyses. Large registry databases like SEER are essential for studying such rare cancers.
The researchers identified 1,604 MBC patients from SEER diagnosed between 2010 and 2018. They collected features including age, tumor size, grade, stage, lymph node involvement, surgery type, radiation, and chemotherapy. The dataset was split into training and internal test sets.
Six machine learning models were trained and compared: CatBoost, Random Forest, Logistic Regression, Support Vector Machine, K-Nearest Neighbors, and Decision Tree. Each model predicted 1-year, 3-year, and 5-year survival outcomes.
CatBoost - a gradient boosting algorithm that handles categorical variables particularly well - emerged as the top performer. Its predictions were then validated on a separate external dataset not used during model development, providing a more rigorous test of real-world generalizability.
On internal testing, CatBoost achieved AUC scores of 0.833 for 1-year survival, 0.806 for 3-year survival, and 0.810 for 5-year survival. AUC (Area Under the Curve) measures how well a model separates survivors from non-survivors - values above 0.8 are considered good.
External validation was even more impressive for short-term prediction: the 1-year AUC reached 0.937. This suggests the model captures genuine prognostic signals rather than patterns specific to the SEER training data.
Feature importance analysis revealed that tumor grade, stage at diagnosis, lymph node involvement, and treatment type were the strongest predictors of survival - findings consistent with established oncological knowledge.
Beyond survival prediction, the study used the model to explore which patient subgroups benefit from radiotherapy. The results revealed important nuances that could directly inform treatment planning.
Radiotherapy was found to be beneficial for patients undergoing breast-conserving surgery (lumpectomy) with M0 (no distant metastasis) disease, and for patients with advanced local disease (T3-T4 tumors or N2-N3 lymph node involvement) undergoing mastectomy.
Critically, radiotherapy did NOT show a significant survival benefit for patients with early-stage disease (T1-T2 tumors, N0-N1 lymph nodes) who underwent mastectomy. This suggests that routine post-mastectomy radiotherapy in early-stage MBC may be unnecessary and could be omitted to spare patients side effects.
One of the most valuable applications of this kind of model is in personalizing treatment for rare cancers where randomized controlled trials are difficult or impossible to conduct due to small patient numbers.
The CatBoost model could be used at the point of diagnosis to estimate prognosis and help oncologists decide on treatment intensity. A patient predicted to have poor 1-year survival might be considered for more aggressive intervention or clinical trial enrollment.
The external validation step is particularly important - it means this model shows promise for use in hospitals and regions outside the SEER catchment area, increasing its potential applicability.
This study demonstrates that machine learning - particularly gradient boosting methods like CatBoost - can generate clinically useful survival predictions for metaplastic breast cancer despite its rarity.
The finding that radiotherapy benefit varies by surgical approach and disease stage has immediate clinical relevance. It supports a more individualized approach to treatment planning rather than blanket application of standard protocols.
Future work should prospectively validate these models in multi-center clinical settings and explore integration with molecular biomarker data to further refine predictions.