Sentinel lymph node biopsy (SLNB) is the standard surgical procedure used to determine whether breast cancer has spread to nearby lymph nodes. Although less invasive than full axillary lymph node dissection, SLNB still carries risks including lymphedema, sensory abnormalities, and shoulder dysfunction - even when the result turns out to be negative.
Neoadjuvant chemotherapy (NACT) - chemotherapy given before surgery - is now standard for many breast cancers. A well-recognized problem is that NACT can cause fibrosis in the primary tumor area that alters lymphatic drainage pathways, making it harder to reliably identify the sentinel node. Studies show that after NACT, the sentinel node identification rate drops and the false-negative rate can reach 14-23% in clinically node-negative patients - far above the accepted 10% threshold.
A growing body of clinical trial evidence (including the SOUND, INSEMA, and BOOG 2013-08 trials) supports omitting SLNB in certain low-risk patients. This trend toward treatment de-escalation - reducing surgical burden while maintaining outcomes - creates a practical need for reliable tools to identify which patients can safely skip SLNB after NACT.
This study is the first to use machine learning to predict sentinel lymph node (SLN) metastasis risk specifically in patients with clinically node-negative (cN0) breast cancer who have completed NACT. The model uses routinely available preoperative clinical and pathological features - no specialized imaging or biomarkers required.
Four machine learning algorithms were compared: XGBoost, Random Forest (RF), Logistic Regression (LR), and Support Vector Machine (SVM). The best-performing model was then interpreted using SHAP (SHapley Additive exPlanations) analysis to identify the most important predictors and explain how the model makes its decisions.
The study included 221 patients from a single center in China treated between January 2017 and January 2025, covering all major breast cancer molecular subtypes. The primary goal was to identify patients at low risk of SLN metastasis who could safely be exempted from SLNB - reducing unnecessary surgical procedures while maintaining patient safety.
All 221 patients had invasive breast cancer confirmed by core needle biopsy, were clinically node-negative (with any suspicious nodes confirmed benign by fine needle aspiration), completed full NACT cycles, and underwent surgery with SLNB. Patients with inflammatory breast cancer, pregnancy-related cancer, prior axillary surgery, and cases where no sentinel node could be identified were excluded.
NACT regimens were assigned based on molecular subtype following NCCN guidelines: anthracycline/taxane combinations for hormone receptor-positive/HER2-negative cancers (24 weeks), anthracyclines/taxanes with or without carboplatin for TNBC (18-24 weeks), and taxanes/pertuzumab/carboplatin/trastuzumab for HER2-positive cancers (18 weeks).
Radiological complete response (rCR) was defined as complete disappearance of the tumor on MRI with no residual detectable on mammography or ultrasound. The post-NACT pathologic tumor category (ypT) and other clinical variables were extracted from electronic medical records. The sentinel node was identified intraoperatively using technetium-99m and/or methylene blue tracing, and pathologically assessed by frozen section and paraffin pathology.
The 221 patients were split 7:3 into a training set (155 patients) and testing set (66 patients) using stratified sampling to maintain the ratio of SLN-positive to SLN-negative cases in both sets. Because only 15.8% of patients had positive sentinel nodes (a class imbalance problem), the Synthetic Minority Over-sampling Technique (SMOTE) was applied to the training set to create artificial examples of the minority class and prevent the model from simply predicting everyone as negative.
All four models were evaluated using five-fold cross-validation on the training set and then assessed on the held-out test set. Performance was measured by AUC, sensitivity, specificity, accuracy, and F1 score. Decision curve analysis (DCA) was used to compare clinical net benefit - a method that accounts for the relative costs of false positives and false negatives at different decision thresholds, giving a more clinically meaningful comparison than AUC alone.
SHAP analysis was applied to the best-performing model to quantify each feature's contribution to individual predictions. SHAP, grounded in cooperative game theory, assigns each feature a value representing its contribution to moving the model's output away from the average prediction, providing both global feature importance rankings and case-level explanations.
On the training set, the more complex tree-based models achieved higher AUCs: XGBoost (0.964) and RF (0.963) outperformed LR (0.924). However, on the independent test set the situation reversed significantly. The LR model achieved the best test AUC of 0.889 (95% CI: 0.886-0.892), while XGBoost dropped to 0.740, RF to 0.824, and SVM to 0.850.
The LR model's test set performance showed specificity of 0.947 (94.7%) and accuracy of 0.864. Sensitivity was lower at 0.448 - reflecting the deliberate prioritization of correctly identifying negative cases (which is clinically appropriate for a tool designed to identify patients who can safely skip SLNB). The F1 score was 0.508.
Decision curve analysis showed the LR model provided the greatest net clinical benefit over the widest range of decision thresholds in the test set (0.10-0.40), consistently outperforming the other models. This means that across a broad range of clinical risk tolerance settings, the LR model would guide the most clinically beneficial decisions about whether to perform SLNB.
SHAP analysis identified three dominant predictors of SLN metastasis: post-NACT radiological complete response (rCR), lymphovascular invasion, and axillary nodes on ultrasonography. These were the features with the highest average absolute SHAP values, meaning they had the largest influence on the model's predictions.
Patients with rCR were strongly predicted to be SLN-negative. The study found that 98% of all patients who achieved rCR had pathologically negative lymph nodes (ypN0), regardless of molecular subtype. Even patients with initially large T3 tumors who achieved rCR had ypN+ rates below 2%. This makes rCR the most powerful single indicator of safe SLNB omission.
Other features significantly associated with SLN positivity included: lymphovascular invasion (tumor cells invading blood or lymphatic vessels around the primary tumor), abnormal axillary nodes on ultrasound, the HR+/HER2- molecular subtype, T3 tumor category, and multifocal or multicentric disease. These features collectively capture both the tumor's intrinsic biology and its response to chemotherapy.
The superior test-set performance of logistic regression over XGBoost and Random Forest reflects a common phenomenon in small medical datasets: simpler models with linear decision boundaries generalize better when sample sizes are limited and positive event rates are low. With only 221 patients and 35 positive cases (15.8%), the complex nonlinear models had too many parameters relative to available data, leading to overfitting.
LR's simple linear structure - combined with L1/L2 regularization to control overfitting - proved well-matched to the approximately monotonic relationships between the identified risk factors and SLN metastasis. Its interpretable odds ratio outputs also align naturally with how clinicians communicate and use risk information in medical decision-making.
A key limitation is the study's small single-center sample, which constrains the model's ability to learn subtle interactions between features and limits external validity. The retrospective design also prevents establishing causal relationships. Future prospective multicenter trials with larger samples, additional biomarkers, and radiomics features are needed before this model could be used in routine clinical practice.
This model aligns directly with the growing clinical movement toward treatment de-escalation in breast cancer - reducing the extent of surgery while maintaining equivalent survival outcomes. Major randomized trials (SOUND, INSEMA, BOOG 2013-08) have demonstrated that omitting SLNB is safe in selected low-risk, early-stage, node-negative patients who undergo breast-conserving surgery without NACT.
This study extends the concept to the post-NACT setting, where the question of SLNB omission has been less studied. The high-specificity LR model could serve as a clinical decision support tool to identify a subset of cN0 breast cancer patients after NACT who are at sufficiently low risk of SLN metastasis that SLNB could reasonably be omitted - sparing them surgery and its associated complications.
Patients who would benefit most from this tool are those with favorable post-NACT characteristics: achievement of rCR, absence of lymphovascular invasion, and normal axillary ultrasound findings - particularly those with HER2-positive or triple-negative breast cancer subtypes where NACT routinely achieves high rates of lymph node sterilization. Prospective multicenter validation is explicitly identified as the necessary next step before clinical adoption.