Abstract
Objective: To establish a model based on clinical and delta-radiomic features within ultrasound images using XGBoost machine learning to predict proliferation-associated nuclear antigen Ki-67 value ≥ 15% in T(2)N(X)M(0) stage primary breast cancer (BC). Method: Data were collected from 228 randomly selected BC patients who received ultrasound screening and postoperative pathologic assessment from April 2015 to September 2018. The patients were classified into the study group (n = 80) and control group (n = 148), and the data were apportioned into the training set and test set at a 7:3 ratio based on time intervals. In the training set, crucial factors were identified from clinical features and grayscale and delta-radiomic features within ultrasound images, by using the chi-square test, t-test, and rank-sum test. The clinical model, imaging model, and combined model were built using multivariate logistic regression, respectively. The model's predictive performance and clinical net benefit were assessed using DeLong's method and decision curve analysis. Meanwhile, an XGBoost algorithm is used to establish a prediction model to verify the above results. Results: The crucial factors affecting Ki-67 value ≥ 15% included BMI, lymph node metastases, BC volume, CA153, pathology type, tumor boundaries, tumor morphology, elastography score, and delta-radscore. The predictive performance of the combined model [AUC 0.857, OR 0.0290, 95% CI 0.793-0.908] was considerably improved on the training set than the clinical model [AUC 0.724, OR 0.0422, 95% CI 0.648-0.792] and the imaging model [AUC 0.798, OR 0.0355, 95% CI 0.727-0.857]. The decision curve analysis also confirmed that the combined model delivered a higher clinical net benefit, and the verification on the test set yielded similar results. The nomogram and the calibration curve plotted based on the combined model achieved satisfactory clinical effects. The SHAP value of the XGBoost algorithm also confirmed that lymph node metastasis, BC volume, elastography score, and delta-radscore are the best independent factors for predicting BC Ki-67 value ≥ 15%. Conclusion: The XGBoost machine learning-based combined model integrating clinical features and delta-radiomic features on ultrasound images was able to predict the Ki-67 value ≥ 15% in an efficient and noninvasive manner, providing important clues for clinical decision-making and follow-up in BC.