# Development and validation of an interpretable ensemble model for predicting androgen receptor status in triple-negative breast cancer: a multi-center study

## TL;DR

This study creates a noninvasive model to predict androgen receptor status in triple-negative breast cancer using radiomics and MRI data, which could help guide targeted therapies.

## Contribution

The novel contribution is an interpretable ensemble model combining radiomics and multiparametric MRI for predicting AR status in TNBC.

## Key findings

- The integrated model achieved an AUC of 0.891 in training and 0.863 in external validation.
- Radiomic features like skewness and surface-to-volume ratio were key predictors according to SHAP analysis.
- The model showed high sensitivity (78–85%) and specificity (82–87%) across different patient cohorts.

## Abstract

Reliable assessment of androgen receptor (AR) status in triple-negative breast cancer (TNBC) is critical for targeted therapy but remains challenging due to biopsy limitations from intratumoral heterogeneity. This study aimed to develop and validate an interpretable ensemble model integrating radiomics and multiparametric MRI for noninvasive AR status prediction.

A total of 379 TNBC patients from three institutions were included for model training and external validation. All patients underwent preoperative dynamic contrast-enhanced MRI. Radiomic features were extracted from a Segment Anything Model-based segmentation tool and underwent multi-step selection. Multiparametric MRI features were evaluated using standardized criteria. Three predictive models, including a radiomics model, an MRI model, and an integrated ensemble model, were constructed using a stacking framework with Random Forest, XGBoost, and LightGBM. Model performance was assessed by ROC analysis, calibration, and decision curve analysis. SHapley Additive exPlanations (SHAP) were applied for interpretability.

The integrated model achieved the best performance (AUC = 0.891 in the training cohort), outperforming radiomics (AUC = 0.836) and MRI models (AUC = 0.753). External validation confirmed robustness (AUC = 0.863 and 0.818). The integrated model maintained high sensitivity (78–85%) and specificity (82–87%) across cohorts. SHAP analysis revealed radiomic descriptors, especially skewness and surface-to-volume ratio, as the most influential predictors.

An interpretable ensemble model integrating radiomics and multiparametric MRI achieved robust and generalizable performance for AR status prediction in TNBC. This noninvasive approach may assist in patient stratification for AR-targeted therapy and support personalized treatment strategies.