Objective:There were about 30% breast cancer patients accompanied with occult lymph node metastasis (OLNM), with the majority pathological subtype was Luminal A and B. Breast cancer with OLNM is an independent predictive factor of overall survival and disease-free survival. we aim to establish machine learning (ML) models for predicting OLNM in Luminal breast cancer based on clinical pathological characteristics and MRI features.

Methods: Data were retrospectively collected from patients with pathologically confirmed Luminal breast cancer patients who underwent sentinel lymph node biopsy (SLNB). A total of 155 patients with negative diagnosis of lymph node metastasis detected by preoperative clinical and imaging examinations were enrolled, with 31 patients were eventually confirmed to have metastasis (OLNM+) and 124 patients were confirmed to be without metastasis (OLNM-) by pathology. According to an 8:2 ratio, the entire cohort was randomly divided into a training cohort (n = 124) and a test cohort (n= 31). ML algorithms were utilized to construct prediction model, including logistic regression (LR), support vector machine(SVM) and extreme gradient boosting (XGBoost). Model performance was assessed using receiver operating characteristic (ROC) curves, area under the curve (AUC), and validated with the test cohort. Decision curve analysis (DCA) was employed to compare and validate the predictive performance of models.

Results: The ADC value, T2WI signal, peripheral vessel sign, and vascular cancer embolism were showed significant difference between the OLNM+ group and the OLNM- group  (p<0.05), Specifically, the ADC values of OLNM+ group [0.81(0.7, 0.9)]was significantly lower than OLNM- group [ 0.9 (0.7, 1.1)]. The T2WI signals were appeared significantly less in the OLNM+ group compared to the OLNM- group. Moreover, the incidence of peripheral vessel sign and vascular cancer embolism was significantly higher in the OLNM+ group than in the OLNM- group.The AUC of model constructed by SVM, LR, XGBoost were 0.7933, 0.86, 0.8733, respectively. The sensitivity, specificity, and accuracy of XGBoost model and LR model in the test cohorts were consistent at 90.9%, 71.4%, and 80%, respectively. The sensitivity, specificity, and accuracy of SVM model were 91.30%, 50.00%, 80.65%, respectively. Based on DCA results, the XGBoost model demonstrated the highest clinical value compared to SVM and LR models.

Conclusions:

ML model based on clinical pathological characteristics with MRI features has a certain value for predicting OLNM in Luminal breast cancer.