This exploratory study investigated the utility of combining ultrasound (US) radiomics with machine learning (ML) models for the non-invasive, preoperative prediction of the human epidermal growth factor receptor 2 (HER2) expression status in breast cancer. The research was performed by collecting images of breast masses, extracting a comprehensive set of quantitative radiomic features from them, and using these features to train and validate several classification machine learning models. The findings demonstrated strong predictive capabilities, with the best model, a Support Vector Machine (SVM), yielding an Area Under the Curve (AUC) of 0.865 and an accuracy of 0.814 in the validation cohort. This performance confirms the strong capability of US radiomics features to differentiate between HER2-positive and HER2-negative tumors. The main point from the discussion is that this radiomics-based approach provides a promising, objective, and efficient tool for predicting HER2 status preoperatively. This non-invasive assessment could significantly assist clinicians in making timely treatment decisions and may help reduce the necessity or complexity of traditional diagnostic procedures, such as core needle biopsy, in certain clinical scenarios.