A Machine-Learning Model Using Pre-treatment Multimodal Data to Predict Sentinel Lymph Node Status After Neoadjuvant Chemotherapy in Operable Early-stage Breast Cancer.
Authors
Affiliations (9)
Affiliations (9)
- Department of Radiology, The Affiliated Hospital, Southwest Medical University, Luzhou, Sichuan, China; Department of Radiology, The Second Affiliated Hospital of Chengdu Medical College, Nuclear Industry 416 Hospital, Chengdu, Sichuan, China.
- Institute of System Science and Technology, School of Electrical Engineering, Southwest Jiaotong University, Chengdu, Sichuan, China.
- Department of Radiology, The Affiliated Hospital, Southwest Medical University, Luzhou, Sichuan, China.
- Department of Radiology, The Suining Central Hospital, Suining, Sichuan, China.
- Department of Radiology, The Third People's Hospital of Chengdu, China.
- Department of Ultrasound, The Sixth People's Hospital of Chengdu, China.
- Department of Epidemic and Health Statistics, School of Public Health at Chengdu Medical College, Chengdu, Sichuan, China.
- Department of Breast Surgery, The Affiliated Hospital of Southwest Medical University, Luzhou, Sichuan, China.
- Department of Radiology, The Affiliated Hospital, Southwest Medical University, Luzhou, Sichuan, China. Electronic address: [email protected].
Abstract
Although accurate preoperative prediction of sentinel lymph node (SLN) status in operable early-stage breast cancer (EBC) after neoadjuvant chemotherapy (NAC) is crucial for optimal axillary surgery, it remains challenging with a single imaging modality. We aimed to develop and validate an interpretable multimodal machine-learning model to improve diagnostic accuracy. In total, 362 operable EBC patients who had completed NAC followed by surgery were enrolled, namely 216 patients in the primary cohort and 146 patients in two independent external test cohorts. Radiomics features derived from magnetic resonance imaging (MRI), mammography and grayscale ultrasound were integrated with clinical risk predictors to build predictive models using Light Gradient Boosting Machine (LightGBM), Random Forest, Logistic Regression, and eXtreme Gradient Boosting algorithms. The optimal model was further selected by balancing complexity and performance, interpreted using SHapley Additive exPlanations, and validated in independent external cohorts stratified by breast cancer molecular subtypes to assess its clinical robustness and generalizability. The LightGBM model incorporating 25 features demonstrated robust performance, achieving AUCs of 0.95, 0.93, and 0.91 in the internal validation cohort, external test cohort 1 and external cohort 2, respectively. It was predominantly driven by DCE-MRI-derived sphericity features and maintained different discriminative ability across molecular subtypes, yielding AUCs of 0.81 for HER2-overexpressing, 0.92 for Luminal B HER2-negative, 0.98 for Luminal B HER2-positive, and 0.79 for triple-negative. The LightGBM model demonstrated promising preoperative predictive ability for SLN status after NAC in operable EBC, but its clinical utility requires validation in larger, multicenter prospective studies, considering the retrospective design, small external cohorts, and subtype imbalance.