Sentinel lymph node biopsy (SLNB) is the standard procedure for assessing whether breast cancer has spread to axillary lymph nodes in patients who appear node-negative on clinical examination. The sentinel node is the first lymph node to receive drainage from the breast and is therefore the most likely site of early metastasis - if it is clear, the remaining lymph nodes are almost certainly clear too.
Despite its widespread use, SLNB carries real risks. It demands high technical expertise, and its false-negative rate ranges from 4 to 12% depending on technique. More importantly, 60 to 80% of patients who undergo SLNB receive negative results - meaning the majority are subjected to a surgical procedure, with associated risks of prolonged operative time, upper limb lymphedema, and paresthesia (abnormal sensations), without benefit.
Lymph node status is also deeply connected to prognosis: patients with axillary lymph node metastasis have a 5-year overall survival rate up to 40% lower than node-negative patients. This combination of high clinical importance and a procedural burden falling on many patients who did not need the procedure creates a clear need for accurate, non-invasive pre-surgical prediction tools.
This study developed and compared five machine learning models for predicting sentinel lymph node (SLN) metastasis in clinically node-negative breast cancer patients. The models integrated both pathological features (from core needle biopsy and tumor histology) and imaging features from both ultrasound and MRI of the primary tumor and axillary lymph nodes.
The five algorithms tested were logistic regression (LOGIT), LASSO regularized regression, XGBoost, Random Forest, and gradient boosting machine (GBM). The best-performing model was then analyzed using SHAP (SHapley Additive exPlanations) values to identify which features most influenced predictions and to provide case-level explanations that can be understood by clinicians.
The study was conducted at Chongqing Hospital of Traditional Chinese Medicine in China, enrolling 225 breast cancer patients who underwent SLNB between January 2014 and October 2024. A key strength of the design is the integration of 34 clinical features from three data sources - pathology, breast ultrasound, and axillary lymph node ultrasound/MRI - allowing the models to capture both tumor biology and direct signs of nodal involvement.
From 34 collected clinical features, the study used univariate and multivariate logistic regression to screen down to the most predictive variables. Features passing univariate analysis (p less than 0.1) were entered into a multivariate model, and only features remaining significant at p less than 0.05 were retained. This two-stage approach prevents overfitting from including too many variables in a small dataset while ensuring that selected features have independent predictive value.
The 225 patients were split 70/30 into a training set (159 patients) and validation set (66 patients). Because some clinical features had missing values, multiple imputation (MI) was used to fill gaps using the R package 'mice' - a technique that creates multiple complete datasets and combines results, rather than simply filling in the mean or dropping incomplete cases, which would introduce bias.
To address class imbalance in the training set (61.7% positive, 38.2% negative), the SMOTE (Synthetic Minority Over-sampling Technique) mechanism was applied during cross-validation to avoid models that simply predict the majority class. All five models were trained using 10-fold cross-validation with grid search hyperparameter optimization, and performance was assessed across accuracy, sensitivity, specificity, AUC, precision-recall AUC, and Brier score.
Multivariate logistic regression identified five independent predictors of SLN metastasis: Multifocal disease (multiple tumor foci), lymphovascular invasion (LVI) (cancer cells found in lymphatic or blood vessels near the tumor), Maximum Diameter of the primary tumor, Shape US (whether the tumor is regular or irregular on ultrasound), and Maximum Cortical Thickness of axillary lymph nodes on ultrasound.
Multifocal breast cancer carries a lymph node metastasis rate of approximately 52%, compared to 37.5% in unifocal disease. LVI has long been recognized as an independent predictor of sentinel node involvement, representing the biological process by which cancer cells gain access to the lymphatic system. Tumors larger than 2 cm have significantly higher SLN metastasis rates, consistent with prior studies - and the multivariate model confirmed tumor size as an independent risk factor with an odds ratio of 2.17.
The finding that Maximum Cortical Thickness greater than 3 mm on axillary ultrasound predicts SLN metastasis confirms a widely used clinical threshold. Cortical thickening reflects the accumulation of metastatic tumor cells in the lymph node cortex and is measurable non-invasively during routine preoperative ultrasound - making it an immediately actionable finding. This 3 mm cutoff has been validated in multiple prior studies and is consistent with the SHAP analysis results.
On the training set, Random Forest achieved the highest AUC of 0.9177 (95% CI: 0.874-0.954) and precision-recall AUC of 0.891, with specificity of 0.913 and sensitivity of 0.684. XGBoost was a close second at training AUC 0.9111, followed by GBM (0.9016), logistic regression (0.8974), and LASSO (0.899).
On the validation set, the ranking shifted: logistic regression and LASSO tied for the highest accuracy at 83.3% and AUCs of 0.868 and 0.871 respectively. Random Forest's validation AUC was 0.881 - the highest on the validation set by AUC - but with lower sensitivity (0.667) and accuracy (77.3%) than the linear models. All five models maintained AUC above 0.85 and specificity above 0.85 on the validation set, with no evidence of overfitting.
Decision curve analysis showed that all five models provided net clinical benefit above the 'treat all' strategy across most probability thresholds. When the predicted metastasis probability exceeded 0.75, the Random Forest model showed higher net benefit than the other four models - suggesting it is the preferred choice when a clinician needs high confidence before deciding to forgo SLNB in patients with low predicted probability.
SHAP (SHapley Additive exPlanations) values were calculated for all five models to determine which features had the greatest influence on predictions. Across all five algorithms, two features consistently contributed the most to model decisions: Maximum Diameter (primary tumor size) and Maximum Cortical Thickness (axillary lymph node cortical thickness on ultrasound).
SHAP's grounding in cooperative game theory means each feature's value represents its average contribution across all possible subsets of features - a mathematically rigorous approach that avoids the limitations of simpler importance metrics. SHAP provides both global importance rankings (which features matter most overall) and local case-level explanations (why the model made a specific prediction for an individual patient), with red arrows in the waterfall plots indicating features pushing the prediction toward metastasis and blue arrows indicating features pushing it away.
The SHAP results are clinically interpretable: larger tumors and thicker lymph node cortices are both readily measurable from imaging routinely obtained before surgery, and both have well-established biological rationales for predicting nodal spread. The agreement between SHAP-identified predictors and clinically established risk factors validates that the model is learning genuine biological relationships rather than dataset noise.
The practical clinical application of this model is to identify breast cancer patients who can safely avoid SLNB because their predicted probability of sentinel node metastasis is sufficiently low. Given that 60-80% of patients currently undergoing SLNB have negative results, a model that accurately stratifies patients could spare a substantial fraction of patients from surgery and its associated complications.
The five-feature model is particularly well-suited to clinical workflows because all predictor variables are obtainable from standard preoperative assessments: core needle biopsy provides tumor size, histology, and LVI status; breast ultrasound provides morphology and maximum diameter; and axillary lymph node ultrasound provides cortical thickness measurements. No additional specialized tests are required.
The authors envision these models being deployed as decision support tools in clinical oncology. For patients with very low predicted metastasis probability and favorable imaging features - particularly low-risk patients who would typically require SLNB only for staging purposes - the model provides an evidence-based foundation for potentially omitting the procedure. Prospective multicenter validation remains the necessary next step before routine clinical adoption.
This study successfully developed and validated five machine learning models for predicting SLN metastasis in breast cancer, demonstrating that the integration of pathological and imaging features can achieve meaningful predictive accuracy. The Random Forest model showed the best overall performance, particularly at high probability thresholds where clinical decisions are most consequential.
The use of SHAP explainability is a key strength, ensuring that the model's decisions can be audited and understood by clinicians rather than accepted as opaque outputs. The identification of tumor maximum diameter and axillary cortical thickness as the primary drivers is consistent with established clinical knowledge and provides a coherent biological rationale for the model's behavior.
Study limitations include the retrospective single-center design, small sample size, and absence of external validation - all of which constrain generalizability. Future work should incorporate prospective multicenter studies with larger cohorts, additional biomarkers, and external validation datasets. Successful validation would enable these models to reduce unnecessary surgical procedures and support more personalized treatment decisions for breast cancer patients.