Development of a prognostic model for preoperative stage I-III breast cancer using machine learning with integrated cone-beam breast computed tomography data.
Authors
Affiliations (4)
Affiliations (4)
- Department of Medical Imaging Center, Guangxi Medical University Cancer Hospital, Guangxi Zhuang Autonomous Region, Nanning, 530021, China.
- Department of Medical Imaging Center, Guangxi Medical University Cancer Hospital, Guangxi Zhuang Autonomous Region, Nanning, 530021, China. [email protected].
- Department of Experimental Research, Guangxi Medical University Cancer Hospital, Guangxi Zhuang Autonomous Region, Nanning, 530021, China. [email protected].
- Guangxi Key Laboratory of Basic and Translational Research for Colorectal Cancer, Nanning, China. [email protected].
Abstract
This study aimed to develop an integrated model based on cone-beam breast computed tomography (CBBCT) and hematological indicators to predict the prognosis of preoperative stage I-III breast cancer. A retrospective analysis was performed on 243 patients with pathologically confirmed stage I-III breast cancer. A novel machine learning framework for feature selection was employed, which integrates 14 distinct algorithms and explores 101 possible combinations, enhancing the ability to identify the most relevant features in high-dimensional medical imaging datasets. After feature selection, a patient risk score was calculated to construct a nomogram model for breast cancer prognosis. The nomogram model was evaluated using receiver operating characteristic (ROC) curve analysis and calibration curves. Univariate and multivariate regression analyses were conducted to validate the screened features and determine independent risk factors. A machine learning computational framework based on 101 combinations selected 12 prognostic indicators for overall survival (OS) and 18 for disease-free survival (DFS) from 37 CBBCT and hematological features. The model incorporating clinical and imaging indicators achieved an average area under the curve (AUC) value of 0.832 in both the training and validation datasets, demonstrating superior overall survival (OS) prediction performance compared to the clinical model without CBBCT indicators (AUC = 0.777). Similarly, the AUC values for DFS prediction in the training and validation sets were 0.996 and 0.732, respectively. Molecular typing, enhancement curve types, and morphology were independent risk factors for OS in the clinical prediction model. Calcification was an independent risk factor associated with DFS. A nomogram model was established combining the above features. Our study successfully screened prognostic-related CBBCT and hematological features. The developed nomogram showed satisfactory preoperative predictive efficacy for stage I-III breast cancer.