Breast cancer can be classified into four molecular subtypes based on immunohistochemistry which can in turn determine its prognosis and treatment strategy [1]. The luminal A subtype typically presents a favorable prognosis and predominantly responds to endocrine therapy. By contrast, the luminal B subtype is often associated with a relatively poorer prognosis that may require additional chemotherapy or HER2-targeted treatment [2], [3], [4]. Triple-negative breast cancer (TNBC) is a highly aggressive subtype and usually bleak and the main treatment strategy is chemotherapy [5,6]. The HER2-enriched subtype primarily requires HER2-targeted therapy, and it responds well to neoadjuvant therapy [7,8]. Consequently, accurate preoperative determination of molecular subtypes is crucial to the development of personalized treatment plans for patients with breast cancer.

Magnetic resonance imaging (MRI) is one of the foremost methods used for breast cancer diagnosis [9,10]. Multifunctional MRI imaging techniques, such as ultrafast dynamic contrast-enhanced (DCE)-MRI, magnetic resonance spectroscopy (MRS), diffusion kurtosis imaging (DKI) and intravoxel incoherent motion (IVIM) can effectively assess intra-tumor microenvironmental changes and characterize breast cancer tumor heterogeneity [9,[11], [12], [13]]. Numerous studies have confirmed the efficacy of multiparametric MRI and MRI-based semantic features in predicting molecular subtypes of breast cancer [14], [15], [16]. However, there is a lack of studies integrating a range of multiparametric analyses within an identical cohort to predict molecular subtypes. Moreover, research on the comparative effectiveness of multiparametric MRI and semantic features in predicting these subtypes is limited.

Moreover, the efficacy of conventional statistical methodologies frequently dwindles with regard to intricate and multi-dimensional data. Several researchers have developed machine learning-based prediction models for cancer molecular subtypes with promising results [17,18]. Machine learning is a sophisticated computational approach known for its efficiency and predictive accuracy, making it a robust tool for identifying breast cancer molecular subtypes [19,20]. Nevertheless, the decision-making mechanisms within models remain opaque, and this lack of interpretability substantially curtails their widespread implementation.

The SHapley Additive exPlanations (SHAP) presents a robust framework for model interpretability analysis that is characterized by comprehensive applicability and superior visualization [21,22]. Consequently, it facilitates a more intuitive and lucid understanding of machine-learning model interpretations. Recent studies that used the SHAP, for both the prediction of molecular subtypes and risk assessment in terms of sentinel lymph node metastasis, were able to notably enhance the reliabilities and practical applicability of their models [14,23].

The purpose of this study was to develop and validate an interpretable multiparametric model using machine learning and the SHAP method, compared with semantic model.