Pretreatment Prediction of Tumor Recurrence in Breast Cancer After Neoadjuvant Systemic Therapy Using Machine Learning With Clinical and CT Radiomics Features.

This study aimed to develop machine learning models for predicting tumor recurrence in breast cancer before neoadjuvant systemic therapy (NST) by integrating clinical and radiomic features derived from pretreatment computed tomography (CT). We retrospectively enrolled 235 patients with 237 breast tumors who underwent contrast-enhanced CT before NST. Datasets were randomly divided into five-fold training and testing sets using semi-random partitioning to ensure similar clinical characteristics between the two subsets. Subsequently, a nested five-fold cross-validation was performed to develop a recurrence prediction model using three machine learning algorithms across clinical, radiomics, and integrated models. The performance of prediction models was compared using the area under the receiver-operating characteristic curve (AUC), and the best clinical and radiomics models were further integrated to develop the final model. Kaplan-Meier analysis with a log-rank test was conducted to compare survival curves between high- and low-risk groups stratified by the prediction models. The comparisons demonstrated that the random survival forest (RSF) clinical model (mean AUC = 0.755) and the Cox-least absolute shrinkage and selection operator (Cox-LASSO) radiomics model (mean AUC = 0.636) outperformed other machine learning algorithms. The integration of the clinical (RSF) and radiomics (Cox-LASSO) models achieved a mean AUC of 0.777 in predicting tumor recurrence. The log-rank analysis revealed significant differences in the survival curves between the high- and low-risk groups stratified by the integration model on the testing sets. In conclusion, the integration of clinical and CT-based radiomics models was helpful for the pretreatment prediction of tumor recurrence in patients with breast cancer after NST.