Development and external validation of a machine learning-based multimodal radiomics nomogram for predicting...

Triple-negative breast cancer (TNBC) is a highly aggressive subtype characterized by significant heterogeneity, making accurate prognostic assessment essential for clinical management. However, existing predictive models predominantly rely on single-modality data and lack external validation, which may limit their generalizability and clinical applicability. In light of these limitations, this study sought to explore the feasibility of developing a multi‑center prognostic model incorporating clinical and multimodal imaging data, and to preliminarily examine its potential value in supporting risk stratification and treatment decision-making for TNBC patients.

This retrospective, two-center study aimed to develop and externally validate a machine learning-based multimodal radiomics model. This study analyzed 108 patients with pathologically confirmed TNBC between June 2017 and November 2022. Inclusion criteria: pathologically confirmed TNBC, complete clinical/imaging data, no prior anticancer therapy. Exclusion criteria: incomplete follow-up, poor image quality. Pathological confirmation of TNBC [estrogen receptor (ER)/progesterone receptor (PR)/human epidermal growth factor receptor 2 (HER2) negativity by American Society of Clinical Oncology (ASCO)/College of American Pathologists (CAP) guidelines] served as the reference standard. Clinical predictors screened included age, menopausal status, tumor size, nodal status, World Health Organization (WHO) grade, lymphovascular invasion, perilesional edema, and Breast Imaging Reporting and Data System (BI-RADS) features. Clinical, pathological, and imaging [magnetic resonance imaging (MRI) and digital breast tomosynthesis (DBT)] data were collected. Progression-free survival (PFS) risk factors were identified using Cox regression and Kaplan-Meier analysis with log-rank tests. Regions of interest (ROIs), encompassing the primary tumor as well as 5- and 10-mm peritumoral areas, were manually delineated on MRI [T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), dynamic contrast-enhanced (DCE), diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC)] and DBT [craniocaudal (CC) and mediolateral oblique (MLO) views] images, followed by radiomic feature extraction. Feature selection was performed using the two-sample t-test and least absolute shrinkage and selection operator (LASSO) regression. Five machine learning algorithms [AdaBoost (AB), light gradient boosting machine (LGBM), extreme gradient boosting (XGB), logistic regression (LR), and random forest (RF)] were used to construct models based on clinicopathological, conventional imaging, and radiomic features (tumor-only, tumor + peritumoral, and peritumoral). Rad scores derived from multimodal features were integrated with significant clinicopathological and conventional imaging variables through multivariate Cox stepwise backward regression to construct a predictive nomogram. The nomogram's performance was subsequently evaluated using calibration curves and decision curve analysis.

The study included 108 TNBC patients (mean age, 50.42 years; range, 25-79 years) with a median follow-up of 56 months (range, 12-86 months) until August 31, 2024. The cohort was divided into a training set (n=75; 15 progressions) and an external validation set (n=33; 7 progressions). A nomogram incorporating five independent risk factors for PFS-perilesional edema, WHO grade, lymphovascular invasion, peritumoral 10-mm DBT Rad score, and tumor MRI Rad score-demonstrated strong predictive performance, with areas under the curve (AUCs) of 0.858 [training, 95% confidence interval (CI): 0.788-0.928] and 0.736 (validation, 95% CI: 0.676-0.796), and concordance indices (C-indices) of 0.836 (training) and 0.712 (validation). Calibration curves showed good agreement between predicted and observed outcomes, and decision curve analysis confirmed clinical utility across a range of threshold probabilities.

The proposed nomogram exhibits promising predictive performance for PFS in TNBC and may offer a useful reference for future prognostic investigations.