Background: This study was conduct to develop a predictive model for differentiating between benign and malignant Breast Imaging Reporting and Data System (BI-RADS) 4A, 4B breast lesions, which featured non-parallel and small on ultrasound. Methods: For this retrospective study, 327 patients were recruited in the First Affiliated Hospital of Nanjing Medical University from June 2020 to May 2024. Patients were divided into training set (N=229) and validation set (N=229) at the ratio of 7：3. Logistic regression analysis was used to identify risk factors and develop a predictive model to differentiate benign and malignant breast lesions. The effectiveness of the model was evaluated by the receiver operating characteristic (ROC) curve and the decision curve analysis (DCA). Results: The proportion of malignant tumors was 36.1% (118/327) in this study. With univariate and multivariate analyses, a predictive model compromised age, margin, elasticity assessment and US-BI-RADS was built and shown as a nomogram. The area under the ROC curve was 0.846 and 0.798 in the training and test cohort, respectively. DCA demonstrated that the model could achieve benefits for patients. Moreover, the study stratified the breast lesions into three risk groups according to the risk scores calculated by the nomogram. Patients were regarded as “high risk” with a risk score more than or equal to 0.7, and as “low risk” with a risk score less than or equal to 0.1. The proportion of malignancy was 8.8% and 82.1% in low- and high-risk group. Conclusion: The study established a predictive model based on clinical and ultrasound features. The model could effectively differentiate benign and malignant non-parallel and small BI-RADS 4A, 4B breast lesions. This model can distinguish different risks, with a lower probability (8.8%) of malignant breast lesions in the low-risk group and a higher probability (82.1%) in the high-risk group.