A Machine-Learning Model Using Pre-treatment Multimodal Data to Predict Sentinel Lymph Node Status After Neoadjuvant...

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Accurate preoperative prediction of sentinel lymph node (SLN) status in operable early-stage breast cancer (EBC) after neoadjuvant chemotherapy (NAC) is critical for guiding axillary surgery.
• Current diagnostic methods using single imaging modalities face limitations in achieving high accuracy.
• Developing a multimodal approach is essential to improve preoperative SLN status prediction.

Purpose of the Study:

• To develop and validate an interpretable multimodal machine-learning model for improved preoperative prediction of SLN status in operable EBC post-NAC.
• To integrate radiomics features from MRI, mammography, and ultrasound with clinical predictors.
• To assess the model's performance and generalizability across different breast cancer molecular subtypes.

Main Methods:

• A cohort of 362 operable EBC patients who completed NAC and underwent surgery was analyzed.
• Radiomics features from MRI, mammography, and ultrasound were extracted and combined with clinical risk predictors.
• Light Gradient Boosting Machine (LightGBM) and other algorithms were used to build and validate predictive models, with interpretation via SHapley Additive exPlanations.

Main Results:

• The LightGBM model achieved high AUCs (0.95 internal, 0.93 external cohort 1, 0.91 external cohort 2).
• Key predictive features included DCE-MRI-derived sphericity.
• The model demonstrated varying but significant discriminative ability across molecular subtypes (e.g., AUC 0.98 for Luminal B HER2-positive, 0.79 for triple-negative).

Conclusions:

• The developed LightGBM model shows promising preoperative predictive performance for SLN status in operable EBC after NAC.
• Further validation in larger, prospective, multicenter studies is recommended due to the retrospective design and cohort limitations.
• The model's ability to predict SLN status across different molecular subtypes warrants further investigation for clinical utility.