Predicting sentinel lymph node metastasis in breast cancer using an interpretable machine learning approach based on multi-domain clinical features

Abstract

Background

Accurate identification of sentinel lymph node (SLN) status is essential for surgical and adjuvant therapy decisions in breast cancer. Although sentinel lymph node biopsy (SLNB) is less invasive than axillary lymph node dissection, it still exposes node-negative patients to unnecessary surgical risks. Most existing models predicting lymph node status in breast cancer rely on a limited set of features or variables that are not routinely accessible, and their interpretability remains limited. Therefore, developing a comprehensive and explainable model to evaluate the feasibility of omitting SLNB has important clinical significance.

Methods

A retrospective cohort of 1,485 patients trained nine machine learning (ML) algorithms using clinical, imaging, and pathological features to predict sentinel lymph node metastasis (SLNM). An independent prospective cohort of 103 patients served for prospective validation. Multiple imputation enhanced data reliability, and multi-dimensional feature selection simplified models while retaining key information. Model performance was evaluated using AUC, precision, recall, and calibration. SHapley Additive exPlanations (SHAP) were applied to interpret model outputs and quantify feature contributions.

Results

Ten-fold cross-validation showed that the support vector machine (SVM) model achieved the best performance among nine ML algorithms. After optimizing to 26 variables, the internal test achieved an AUC of 0.892 (95% CI: 0.854–0.940), outperforming the clinical model (AUC 0.796, 95% CI: 0.791–0.811). In the prospective cohort, the AUC was 0.775 (95% CI: 0.667–0.865), superior to the baseline radionuclide- and imaging-based model (AUC 0.65, 95% CI: 0.564–0.729). When applying a safety-optimized threshold, the model calibrated well and could spare 58.1% of node-negative patients from SLNB. SHAP analysis identified tumor margin smoothness, AR expression, and nuclear pleomorphism as top predictors. Decision curve analysis supported clinical utility. A Shiny-based web tool enables real-time prediction.

Conclusions

This study developed an SLNM prediction model based on multi-dimensional features available from routine diagnostic examinations. Prospective and exploratory external validation confirmed its clinical applicability. The model’s performance in assessing axillary status was close to that of surgical SLNB, providing preliminary evidence to inform potential de-escalation or stratification of surgical approaches in breast cancer management.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13058-026-02329-1.

Introduction

Breast cancer (BC) is the most prevalent malignancy among women worldwide. In 2022, it accounted for 11.6% of all new cancer cases and 6.9% of cancer-related deaths globally. In China, BC represented 15.59% of all newly diagnosed cancers and 7.94% of cancer-related mortality, highlighting a substantial national burden [1, 2]. Lymphatic drainage is a major route for BC metastasis. Approximately 75% of lymphatic flow from the breast drains into the axillary lymph nodes, making their status critical for accurate staging and prognostic assessment of patients [3]. Given that lymphatic spread in BC typically follows an orderly, stepwise progression, SLNB has enabled omission of axillary lymph node dissection (ALND) in selected cases [4, 5]. In patients with T3 BC, the SLN positivity rate typically ranges from 40% to 65%, whereas only approximately 18% to 30% of early-stage BC patients (T1–T2, cN0) exhibit SLN metastasis [6–9]. Even patients undergoing SLNB alone may experience non-negligible postoperative complications, such as lymphedema, shoulder pain, and numbness [10, 11]. Moreover, minimally invasive candidates who are not suitable for SLNB currently lack reliable methods to assess the likelihood of SLNM. Therefore, to further promote surgical de-escalation, it is imperative to enhance our understanding of the risk factors and biological mechanisms underlying SLN metastasis in BC. Accurate prediction of SLN involvement is crucial for informing personalized diagnostic and therapeutic approaches, as well as for optimizing subsequent management strategies.

Although several studies have explored risk factors associated with SLN metastasis and developed predictive models in BC, these studies present limitations to a certain extent. Most existing studies have included a limited number of predefined features and employed traditional, simplistic variable selection methods to identify combinations of risk factors, subsequently constructing nomograms based on regression models. ML techniques have either not been applied or, when used, often lack interpretability and fail to fully capture latent risk factors or complex feature interactions. Moreover, the relatively small sample sizes or lack of external validation in many of these studies limit the generalizability of their findings [12–17]. In addition, some studies have directly developed deep learning or machine learning models based on radiomics imaging data to predict SLN status, achieving ROC-AUCs ranging from 0.761 to 0.906 on test sets. However, the majority of these radiomic features still rely heavily on manual, pixel-wise annotation by radiologists, which significantly limits their clinical applicability [18–24]. Therefore, there remains a need to develop reliable, accurate, and clinically applicable models for predicting SLNM, with comprehensive incorporation of relevant features.

In this study, we aimed to develop and validate an interpretable ML model based on real-world data to provide a clinically applicable, interpretable, and generalizable tool for individualized assessment of SLNM risk in BC patients. The model was further prospectively validated to ensure its robustness. We employed SHAP to illustrate feature importance and interpret model predictions, thereby elucidating the model’s practical utility in predicting SLN status among BC patients.

Methods

Data source and study population

Our study included two patient cohorts: a retrospective training cohort and an internal prospective validation cohort. We collected data from breast cancer patients hospitalized at the First Affiliated Hospital of Chongqing Medical University (CMUFAH) between October 2020 and July 2024.The inclusion criteria were: (1) Patients were pathologically confirmed to have invasive breast cancer; (2) patients who underwent SLNB with evaluable SLNs; and (3) age > 18 years. A total of 1,741 patients met these criteria. Exclusion criteria were: (1) patients with missing laboratory, imaging, or pathological data (n = 225); and (2) patients who received neoadjuvant or other preoperative therapies (n = 30). Finally, 1,485 patients were included for model training and testing. The trained model was subsequently prospectively validated in CMUFAH between March 2025 and May 2025. To mitigate the potential impact of class imbalance on model training, a case-enriched sampling strategy was applied during data assembly to ensure adequate representation of SLN-positive cases from the eligible patient population. All patients were derived from routine clinical practice during the predefined study period according to the same inclusion and exclusion criteria. No oversampling, synthetic data generation, or artificial manipulation of individual patient data was performed. This adjustment was applied across the study period rather than being limited to specific time points, thereby maintaining representativeness of the patient population and avoiding systematic selection bias. Exploratory external validation was performed at the Third Affiliated Hospital of Chongqing Medical University (CMUTAH) between January 2018 and May 2025 (Supplementary Table S2). The inclusion and exclusion criteria were consistent with those applied in the primary cohort, and a total of 25 patients met the eligibility criteria and were included in the analysis.

This retrospective study was conducted in accordance with the principles of the Declaration of Helsinki. All data were anonymized, and the requirement for informed consent was waived in accordance with national legislation and institutional policies. The study protocol was reviewed and approved by the Ethics Committee of our hospital (approval number: 2024-468-01).

Feature collection and preprocessing

Data collection

Patient-level variables were extracted from inpatient electronic medical records (EMR). Data included: (1) demographic information (DM), such as age, family history of breast cancer, and menstrual history; (2) laboratory data (LB) obtained within one month prior to surgery, including complete blood counts, tumor markers, sex hormones, and breast function tests; (3) imaging studies (IS), such as MRI, CT, and ultrasound results; and (4) relevant pathological data (RP), including T stage (American Joint Committee on Cancer [AJCC], 8th edition), histological grade (assessed using the Nottingham histological score, NHS), biomarkers, and other tumor characteristics. Several imaging variables included in our model, such as tumor margin smoothness, spiculated margins, and malignant lymph node signs, are based on standardized BI-RADS descriptors that are widely used in both international and Chinese clinical guidelines.

The data were collected following patient history, clinical examination, SLN biopsy, and histopathological assessment. DM, LB, and IS data were obtained from EMR, whereas RP data were independently reviewed and evaluated by pathologists with more than ten years of diagnostic experience. To minimize bias from missing data, variables with more than 70% missing values were excluded from subsequent analyses. A total of 83 features were ultimately included in the study.

Missing value imputation

Missing data were handled using multiple imputation strategies, including multiple imputation by chained equations (MICE), K-nearest neighbors (KNN) imputation, and Generative Adversarial Imputation Networks (GAIN). These approaches represent traditional statistical, similarity-based, and deep learning-based imputation paradigms, respectively, to ensure robustness of the findings.In total, five imputed datasets were generated: GAIN_lp, KNN, RF, CART, and Plp.In addition, we performed visualization and sensitivity analyses across different imputation methods for missing data (see Supplementary Figure S1 and Additional file 2: Supplementary Imputation Methods).To evaluate the overall stability of different imputation methods, we defined a stability index (Stability):

1

Suppose there are M imputation methods, and each method can be assessed across K performance metrics (e.g., AUC, sensitivity, specificity, F1 score). For method i, the overall stability is defined as the average coefficient of variation (CV) across all metrics.

Preprocessing

Outliers in continuous variables were assessed. Outliers were defined as values greater than the upper quartile plus three times the interquartile range, or less than the lower quartile minus three times the interquartile range. Each identified outlier was winsorized by replacing it with the corresponding boundary value, thereby reducing the influence of extreme values and bringing them closer to the main data distribution. Hormone-related continuous variables, however, were not adjusted due to substantial distributional heterogeneity across populations. All laboratory indicators were retained as continuous variables without categorization. Categorical predictors were pre-specified prior to model development. Data were normalized using the StandardScaler method, which scales features to a mean of 0 and a standard deviation of 1. This procedure ensured consistency across features, mitigated scale-related bias, and improved the stability and accuracy of model training. Categorical variables were processed using one-hot encoding. To prevent data leakage, all preprocessing procedures were conducted independently in the training and test sets.

Feature selection

Feature selection was applied to identify the most predictive variables, thereby improving the model’s generalizability, interpretability, and robustness. In this study, we adopted a multi-step strategy for feature selection to enhance model performance and ensure stability. First, all variables were screened using both univariable and multivariable logistic regression analyses. Subsequently, mutual information (MI) and minimum redundancy maximum relevance (mRMR) methods were applied to balance information gain and redundancy control [25]. For continuous variables, the least absolute shrinkage and selection operator (LASSO) combined with recursive feature elimination (RFE) was employed, whereas categorical variables were assessed using random forest (RF) and gradient boosting decision tree (GBDT) algorithms. Finally, the candidate features identified by each method were integrated and included in model development, thereby ensuring comprehensive and robust feature selection.

Model development and comparison

The retrospectively collected BC cohort was randomly divided into a training set (80%) and a test set (20%). The training set was used for model development, while the test set was reserved for internal validation. Nine ML algorithms were applied for model training: multilayer perceptron (MLP), linear discriminant analysis (LDA), logistic regression (LR), SVM, RF, extreme gradient boosting (XGBoost), naive Bayes (NB), KNN, and decision tree (DT). To optimize predictive performance, the optimal feature subsets and hyperparameters were determined through a combination of 10 rounds of 10-fold cross-validation and grid search. Each model was refitted on the training set using the selected optimal feature subset and its corresponding final hyperparameters, ensuring robust model estimation. Finally, the trained ML models were tested in a prospective validation cohort. Their predictive performance was compared against a logistic regression model derived from radionuclide imaging results, providing a benchmark for evaluating the added value of ML-based approaches.

Model evaluation

The reliability of the predictive models was evaluated using several commonly used performance metrics, including accuracy, precision, recall, F1-score, ROC-AUC, negative predictive value (NPV), false discovery rate (FDR), specificity, Youden index, and Brier score. The F1-score represents the harmonic mean of precision and recall (Supplementary Table S1). The clinical utility of the final machine learning models was quantitatively evaluated using DCA [26], with net benefit normalized to facilitate comparison across models. Log-loss was employed to assess the discrepancy between predicted probabilities and observed outcomes, reflecting predictive accuracy. The DeLong test was applied to determine whether the differences in the areas under the AUC between models were statistically significant [27]. Additionally, Integrated Discrimination Improvement (IDI) and Net Reclassification Improvement (NRI) were calculated to evaluate the improvement in predictive performance of the new models compared with baseline models [28]. Collectively, these metrics guided the selection of the optimal predictive model.

Statistical analysis

All statistical analyses were performed using R (version 4.4.2) and Python (version 3.9.21). Continuous variables were summarized as mean ± standard deviation (SD) and compared between training and test sets using independent two-sample t-tests assuming unequal variances. Categorical variables were expressed as counts and percentages, and comparisons were performed using chi-square tests or Fisher’s exact tests where appropriate. Missing data were addressed using multiple imputation with the MICE R package (version 3.18.0), including predictive mean matching with logistic/multinomial regression (pmm_logreg_polyreg), RF, and Classification and Regression Trees (CART). KNN imputation was conducted using the VIM R package (version 6.2.2), whereas GAIN was implemented in Python [29].

Results

Patient characteristics

The retrospective dataset used for model training and internal validation included 1,485 patients with breast cancer. Of these, 1,188 patients were assigned to the training cohort (594 with SLNM [50.0%] and 594 without SLNM [50.0%]), and 297 patients to the testing cohort (149 with SLNM [50.2%] and 148 without SLNM [49.8%]). In the prospective validation cohort, a total of 103 patients were included, of whom 34 (33.0%) had SLNM and 69 (67.0%) did not. Baseline characteristics of the retrospective cohorts are presented in Table 1.

Table 1.

Baseline characteristics of patients between SLNM and Non-SLNM patients, and between training and test sets

Characteristic	SLNM0, N = 742	SLNM1, N = 743	P-value	Train set, N = 1188	Test set, N = 297	P-value
AJCC			0.058			0.263
I	47 (6.3%)	32 (4.3%)		58 (4.9%)	21 (7.1%)
II	526 (70.9%)	564 (75.9%)		872 (73.4%)	218 (73.4%)
III	169 (22.8%)	147 (19.8%)		258 (21.7%)	58 (19.5%)
Diameter, mean (SD)	2.1 (1.0)	2.4 (1.2)	< 0.001	2.3 (1.1)	2.3 (1.2)	0.648
Age, mean (SD)	54.5 (11.8)	51.7 (11.4)	< 0.001	53.2 (11.7)	52.5 (11.8)	0.372
Merche_Age, mean (SD)	13.2 (1.2)	13.3 (1.2)	0.159	13.3 (1.3)	13.2 (0.9)	0.188
Menstrual_cycle, mean (SD)	28.1 (2.4)	28.3 (4.7)	0.356	28.1 (3.5)	28.5 (4.4)	0.249
Diabetes			0.711			0.643
No	677 (91.2%)	682 (91.8%)		1089 (91.7%)	270 (90.9%)
Yes	65 (8.8%)	61 (8.2%)		99 (8.3%)	27 (9.1%)
Smoking			< 0.001			0.540
No	735 (99.1%)	712 (95.8%)		1159 (97.6%)	288 (97.0%)
Yes	7 (0.9%)	31 (4.2%)		29 (2.4%)	9 (3.0%)
Alcohol			< 0.001			0.393
No	735 (99.1%)	715 (96.2%)		1162 (97.8%)	288 (97.0%)
Yes	7 (0.9%)	28 (3.8%)		26 (2.2%)	9 (3.0%)
BMI			0.067			0.007
< 24	440 (59.3%)	405 (54.5%)		697 (58.7%)	148 (49.8%)
≥ 24	302 (40.7%)	338 (45.5%)		491 (41.3%)	149 (50.2%)
Family Tumor			0.282			0.192
0	642 (86.5%)	637 (85.7%)		1027 (86.4%)	252 (84.8%)
1	32 (4.3%)	28 (3.8%)		51 (4.3%)	9 (3.0%)
2	7 (0.9%)	16 (2.2%)		15 (1.3%)	8 (2.7%)
3	61 (8.2%)	62 (8.3%)		95 (8.0%)	28 (9.4%)
Gender			0.218			0.590
Female	738 (99.5%)	742 (99.9%)		1183 (99.6%)	297 (100.0%)
Male	4 (0.5%)	1 (0.1%)		5 (0.4%)	0 (0.0%)
Hyperlipidemia			0.724			0.883
No	706 (95.1%)	703 (94.6%)		1126 (94.8%)	283 (95.3%)
Yes	36 (4.9%)	40 (5.4%)		62 (5.2%)	14 (4.7%)
Hypertension			0.003			0.931
No	597 (80.5%)	641 (86.3%)		991 (83.4%)	247 (83.2%)
Yes	145 (19.5%)	102 (13.7%)		197 (16.6%)	50 (16.8%)
Reproductive_History			0.001			0.595
0	27 (3.6%)	45 (6.1%)		57 (4.8%)	15 (5.1%)
1	382 (51.5%)	334 (45.0%)		581 (48.9%)	135 (45.5%)
2	333 (44.9%)	364 (48.9%)		550 (46.3%)	147 (49.5%)
Tumor_history			0.323			0.142
0	702 (94.6%)	686 (92.3%)		1115 (93.9%)	273 (91.9%)
1	4 (0.5%)	5 (0.7%)		9 (0.8%)	0 (0.0%)
2	18 (2.4%)	23 (3.1%)		29 (2.4%)	12 (4.0%)
3	18 (2.4%)	29 (3.9%)		35 (2.9%)	12 (4.0%)
Menopause			< 0.001			0.511
No	270 (36.4%)	349 (47.0%)		490 (41.2%)	129 (43.4%)
Yes	472 (63.6%)	394 (53.0%)		698 (58.8%)	168 (56.6%)
Marital			0.001			0.569
Unmarried	16 (2.2%)	39 (5.2%)		45 (3.8%)	10 (3.4%)
Married	726 (97.8%)	704 (94.7%)		1143 (96.2%)	287 (96.6%)
LR			0.166			0.805
Right	352 (47.5%)	384 (51.7%)		592 (49.8%)	144 (48.5%)
Lleft	390 (52.5%)	359 (48.3%)		596 (50.2%)	153 (51.5%)
Type			0.019			0.723
1	654 (88.1%)	671 (90.3%)		1056 (88.9%)	269 (90.6%)
2	25 (3.4%)	34 (4.6%)		49 (4.1%)	10 (3.4%)
3	14 (1.9%)	11 (1.5%)		21 (1.8%)	4 (1.3%)
4	49 (6.6%)	27 (3.6%)		62 (5.2%)	14 (4.7%)
Location			0.049			0.641
1	373 (50.3%)	402 (54.1%)		609 (51.3%)	166 (55.9%)
2	126 (17.0%)	107 (14.4%)		193 (16.2%)	40 (13.5%)
3	160 (21.6%)	135 (18.2%)		240 (20.2%)	55 (18.5%)
4	58 (7.8%)	56 (7.5%)		91 (7.7%)	23 (7.7%)
5	25 (3.4%)	43 (5.8%)		55 (4.6%)	13 (4.4%)

AJCC: American Joint Committee on Cancer, 8th edition; Diameter: tumor size (mm); Age: age at diagnosis (years); Merche_Age: age at menarche (years); Menstrual_cycle: menstrual cycle length (days); Diabetes: history of diabetes mellitus; Smoking: smoking history; Alcohol: alcohol consumption; BMI: body mass index, < 24 = normal, ≥ 24 = overweight/obese (Chinese criteria); FamilyTumor: family history of malignancy (0 = none, 1 = breast cancer, 2 = ovarian/prostate cancer, 3 = other tumors); Hyperlipidemia: history of hyperlipidemia; Hypertension: history of hypertension; Reproductive_History: Reproductive and breastfeeding history (0 = Nulliparous, 1 = Single parity, 2 = Multiple parities); Tumor_history: personal history of malignancy (0 = none, 1 = ovarian/uterine cancer, 2 = other cancers, 3 = breast cancer); Menopause: menopausal status (0 = premenopausal, 1 = postmenopausal); LR: laterality of breast tumor; Type: histological type (1 = invasive ductal carcinoma/invasive carcinoma of no special type, 2 = invasive lobular carcinoma, 3 = mixed type, 4 = special type invasive carcinoma); Location: tumor location (1 = upper-outer, 2 = lower-outer, 3 = upper-inner, 4 = lower-inner, 5 = central/retro-areolar)

Overall, the distribution of most variables was comparable between the training and testing cohorts (most P values > 0.05). However, significant differences were observed between patients with SLNM and those without SLNM. Specifically, compared with non-SLNM patients, those with SLNM had larger tumor diameters (2.1 vs. 2.4 cm), younger age (54.5 vs. 51.7 years), and higher proportions of smoking (0.9% vs. 4.2%) and alcohol history (0.9% vs. 3.8%). Moreover, SLNM patients exhibited significantly higher CA53 (12.2 vs. 13.2 U/mL) and PRL levels (16.4 vs. 17.6 ng/mL), but lower FSH levels (46.6 vs. 43.0 mIU/mL). Certain IS and RP also showed significant intergroup differences (Supplementary Table S2). The overall framework of the study is illustrated in Fig. 1.

Fig. 1.

Flow chart of the study design. A Dataset import and integration. B Data preprocessing and feature selection. C Model development and performance evaluation. D Model interpretation and visualization. CMUFAH: Breast cancer patients treated at The First Affiliated Hospital of Chongqing Medical University; CV: cross-validation; SHAP: SHapley Additive explanation. LR: logistic regression; DT: decision tree; RF: random forest; KNN: k-nearest neighbors; SVM: support vector machine; NB: naive bayes; XGB: extreme gradient boosting; SGBT: stochastic gradient boosting; NNET: neural network; GBDT: gradient boosting decision tree

Imputation methods sensitivity analysis

To ensure the robustness of imputed data, we conducted a sensitivity analysis to evaluate the impact of five imputation methods on the predictive performance of nine machine learning models. The results indicated that the GAIN_lp method achieved the best average performance metrics across all nine models. Although its stability was slightly lower than that of the CART and Plp methods, the differences were minor, and overall it demonstrated good robustness and consistency (Fig. 2).

Fig. 2.

Comparison of imputation methods across mean test metrics

Each point represents the mean value of a test metric for a given imputation method, with error bars indicating ± standard deviation. The last column (Stability) shows the overall stability of different imputation methods.

In both the training and testing cohorts, GAIN_lp showed superior performance across all models, with particularly notable improvements in MLP, SVM, and LDA models (Supplementary Fig. S1). Furthermore, these three models maintained stable performance across different imputation strategies, whereas the performance of NB and DT models was more affected, suggesting greater sensitivity to the choice of imputation method. Based on the sensitivity analysis and overall performance, GAIN_lp was ultimately selected as the default imputation method.

Selection of feature

A multi-step feature selection strategy was applied (Fig. 1). Univariate and multivariate logistic regression identified 17 features, MI and mRMR analyses selected 20 features, LASSO + RFE retained five continuous features, and GBDT/RF identified four categorical features. This multidimensional approach minimized the risk of omitting important predictors (Supplementary Fig. S2). Finally, 34 core features were obtained by combining.

Model development and sensitivity analysis

Hyperparameters of all models were optimized using a combination of 10-round 10-fold cross-validation and grid search (Supplementary Tables S3).

Among the nine machine learning models trained on 34 core features, SVM, LDA, MLP, and LR demonstrated superior performance, with the SVM model achieving the best results on the test set (accuracy 0.845, recall 0.839, F1-score 0.845, ROC-AUC 0.904, negative predictive value 0.840, Youden index 0.690; Fig. 3A). Ten-fold cross-validation on the training set and bootstrap-based ROC analysis on the test set confirmed robust and conservative estimates of model performance, with SVM, LDA, MLP, and LR achieving consistently high AUCs in both datasets (Fig. 3B). Calibration curves and Brier scores further indicated that predicted probabilities of SVM, LDA, and MLP closely aligned with observed outcomes, with test set Brier scores around 0.123, markedly lower than other models (Fig. 3C). Although RF exhibited the lowest log loss in the training set (0.218), indicating excellent fit to the training data, the SVM model showed the closest log loss between training and test sets, reflecting superior stability and interpretability (Supplementary Tables S4). Decision curve analysis (DCA) demonstrated that SVM and LDA provided higher net benefit across clinically relevant threshold probabilities (0.5–0.7), and DeLong tests confirmed that SVM outperformed Decision Tree (ΔAUC = 0.0889, P = 0.0034), KNN (ΔAUC = 0.066, P = 0.0227), and Naive Bayes (ΔAUC = 0.0521, P = 0.0712, marginally significant), while showing comparable AUCs to LR, RF, XGBoost, MLP, and LDA, indicating its robust classification ability in this cohort (Supplementary Tables S5). NRI and IDI analyses suggested that LDA achieved the highest improvement in reclassification and overall discrimination compared to baseline models, with SVM performing second-best (Supplementary Tables S6). Sensitivity analyses using multiple imputation confirmed that SVM demonstrated stable performance across different imputed datasets in the training set, whereas RF maintained the most stable performance in the test set, followed by LDA and SVM (Supplementary Fig. S3; Tables S7 and S8).

Fig. 3.

Performance of nine machine learning models in the internal cohort (training and testing sets). A Comparative performance of nine models in the training and testing sets across multiple evaluation metrics, including accuracy, precision, recall, F1-score, ROC-AUC, negative predictive value (NPV), false discovery rate (FDR), specificity, Youden index, and normalized Brier score (1 – Brier). B Receiver operating characteristic (ROC) curve analysis. C Calibration curve analysis. D Decision curve analysis (DCA)

In summary, both LDA and SVM demonstrated superior performance in training and test sets. Considering that LDA assumes linear separability and normally distributed features, which may not fully capture complex clinical data, and that its probability outputs tend to be concentrated, SVM especially with an RBF-kernel can model nonlinear feature interactions and, after Platt scaling or isotonic calibration, produces well-calibrated, more dispersed probability distributions, facilitating stratification of low-, intermediate-, and high-risk patients. Taken together, while LDA demonstrated slight advantages in some statistical metrics (e.g., NRI/IDI), the SVM model was selected as the final predictive model due to its superior overall performance, stability, and clinical interpretability.

Feature optimization

We further optimized the SVM model using recursive feature elimination based on permutation feature importance (PFI). When retaining 26 final features, the model maintained stable performance in 10-fold cross-validation, with a mean ROC-AUC of 0.892, mean accuracy of 0.802, mean F1-score of 0.800, mean recall of 0.800, mean specificity of 0.804, and mean Youden index of 0.604, comparable to the full-feature model. Therefore, the SVM model incorporating these 26 features was selected as the final predictive model (Fig. 4A).

Fig. 4.

SVM model feature optimization process and final model performance. A Recursive feature elimination (RFE) based on permutation feature importance (PFI) was applied to the SVM model constructed with core features, with performance evaluated by accuracy at each step. B Receiver operating characteristic (ROC) curves of the final SVM model constructed with 26 selected features in the training and testing sets. C Decision curve analysis (DCA) of the final SVM model constructed with 26 selected features in the training and testing sets

The final SVM model demonstrated robust discriminative performance in both the internal training and testing sets, as illustrated by the ROC and decision curve analysis (DCA) curves (Fig. 4B and C). To further evaluate the predictive accuracy for SLNM in BC, a heatmap was generated to visualize the model’s performance across individual patients. Heatmap analysis indicated that the SVM model effectively stratified patients into high-risk and low-risk groups, with consistent performance observed in both the training and testing sets, suggesting strong predictive accuracy and generalizability. These results highlight the potential utility of the model for early clinical identification of BC patients at high risk of SLNM (Fig. 5).

Fig. 5.

Heatmap representation of the SVM model for predicting SLNM in BC

Each row represents a feature, and each column represents a sample. The color gradient encodes the relative values of the features. Since all 26 selected final features are ordinal or continuous variables, they share a unified color scale. Additional rows at the top display the predicted probability of SLNM, true labels, and training/testing cohort assignments, providing an intuitive overview of the model’s discriminative performance across all samples. The heatmap demonstrates that the SVM model effectively stratifies high- and low-risk patients with consistent performance in both training and testing datasets, indicating robust predictive accuracy and potential clinical utility.

Subgroup analysis and validation summary

(A) Receiver operating characteristic (ROC) curves showing the performance of the final SVM model and the baseline model in the prospective validation cohort from the First Affiliated Hospital of Chongqing Medical University. The baseline model identified SLNM positivity when either radioisotope-inactive sentinel lymph nodes or imaging-suspicious axillary lymph nodes were present. (B) exploratory external validation using the independent cohort from the Third Affiliated Hospital of Chongqing Medical University, evaluated by ROC analysis. (C) ROC curves comparing the final SVM model with the logistic regression model by Ceylan et al. in the test cohort. (D) Subgroup analysis, including Location (tumor location; UOQ = upper outer quadrant, Others = other quadrants), T stage (tumor size/stage), and TNBC (triple-negative breast cancer status; 0 = no, 1 = yes).

In the internal testing cohort, the final SVM model (AUC 0.892, 95% CI = 0.854–0.940) outperformed the traditional logistic regression model proposed by Ceylan et al. [13]. (AUC 0.796, 95% CI = 0.791–0.811; Fig. 6C).In the prospective validation cohort, the final SVM model achieved an AUC of 0.775 (95% CI: 0.667–0.865) (Fig. 6A), demonstrating strong generalizability and robustness across different sample distributions and data characteristics. Compared with the baseline model—constructed from radioisotope sentinel lymph node activity and imaging-suspicious axillary lymph node features (AUC = 0.65, 95% CI: 0.564–0.729)—the final SVM model exhibited a marked improvement in discriminative performance.

Fig. 6.

Model performance and exploratory external validation

Furthermore, an exploratory external validation was conducted using an independent cohort from CMUTAH (n = 25), yielding an AUC of 0.753 (Fig. 6B).
Given the limited sample size, this analysis primarily served to explore the model’s potential external applicability and trend of generalization, rather than to establish definitive external validation.

Subgroup analysis indicated that the model’s overall performance in non-triple-negative breast cancer (non-TNBC) patients (AUC 0.90, Recall 0.84) was slightly higher than in triple-negative breast cancer (TNBC) patients (AUC 0.86, Recall 0.70); however, the difference in AUC between the two groups was not statistically significant according to the DeLong test. Model performance was comparable across different tumor locations. In the T stage subgroup analysis, the T3 subgroup achieved a perfect AUC of 1.00, which was significantly higher than the T1 (AUC 0.88) and T2 (AUC 0.89) subgroups. (Fig. 6D) These results suggest that the model is generally robust across molecular subtypes and tumor locations, but performance may vary with tumor size.

To further evaluate the clinical applicability of the SLNM prediction model, we calculated the secondary axillary lymph node false-negative cases identified by completion ALND following or concurrent with SLNB in the test cohort. The results showed that the false-negative rate of the actual SLNB was 9.5%, while that of the SLNM prediction model was 11.1%, indicating a high level of agreement between the model-predicted and the actual SLN status, which could spare 58.9% of node-negative patients from SLNB. We further explored a safety-oriented threshold. When the probability threshold was set to 0.49, the FNR decreased to 7.9%, which falls within the clinically acceptable threshold (10%) and could spare 58.1% of node-negative patients from SLNB. At this operating point, the model maintained acceptable discriminative performance, achieving a favorable balance between sensitivity (0.852) and false-positive rate (1 − specificity = 0.216).

Model explanation

(A) Feature importance matrix and beeswarm plot. Features are ranked in descending order of mean SHAP values to indicate their contribution to model predictions. The probability of SLNM increases with higher SHAP values. Each point represents a patient’s SHAP value for a given feature, with orange indicating higher feature values and green indicating lower values. (B) SHAP dependence plots. Each plot illustrates how a single feature affects model output. For continuous variables, each point represents an individual patient; for categorical variables, boxplots are used. The y-axis represents SHAP values, and the x-axis represents the actual feature values. Features with SHAP values above zero push the prediction toward the “SLNM” class. (C–D) Force plots illustrating the SHAP contributions and decision pathways for representative SLNM-negative and SLNM-positive patients, providing an intuitive view of how individual features drive the final prediction.

To enhance interpretability of the predictive model, SHAP analysis was applied to quantify the contribution of each variable to the model output. At the global feature level, the summary plot (Fig. 7A) ranked feature importance, indicating that tumor edge smoothness (Edge_smooth), androgen receptor expression percentage (AR_percent), cellular atypia (cell_atypia), malignant imaging features of lymph nodes (Ino_malignt_signs), estrogen receptor expression percentage (ER_percent), spiculated margins (spiculated_margins), and patient age were the seven most influential features for the prediction of SLNM. The corresponding beeswarm plot further illustrated the direction of feature effects, showing that higher AR expression and older age were associated with lower predicted probabilities of SLNM in the model, whereas higher ER expression and an increased number of mitotic figures were associated with an elevated SLNM risk in the model.

Fig. 7.

Global and local SHAP interpretation of the final SVM model

The SHAP dependence plots (Fig. 7B) provided additional insights into feature effects. For continuous variables, each dot represented one patient. Increasing age and Ki-67 expression were associated with lower model-predicted SLNM probability in the SHAP analysis., whereas elevated carcinoembryonic antigen (CEA) levels and ER expression contributed toward increased model-predicted SLNM probability. For categorical variables, violin plots demonstrated that the presence of marked cellular atypia and spiculated tumor margins increased SLNM risk. Partial dependence (PD) and individual conditional expectation (ICE) curves (Supplementary Fig. S4) were used to visualize the global and local nonlinear effects of continuous variables on the model predictions, showing patterns consistent with the SHAP analysis. Restricted cubic spline (RCS) analysis further revealed nonlinear trends between continuous variables and SLNM risk in the model (Supplementary Fig. S5), which aligned with the SHAP-derived feature effects. The SHAP dependence plot with interaction (Supplementary Fig. S6) illustrates interactions between features. Specifically, when mitotic figures are high, younger age showed a stronger contribution toward the model prediction of the positive class.

At the individual level, local interpretation of model predictions was performed using SHAP force plots. For example, Fig. 7C illustrates the contributions of features to the prediction for patient #251 (SLNM-negative). In this case, ER expression at 0% and Ki-67 expression at 50% contributed strongly in the negative direction, while MRI score of 4 and unsmooth tumor edge contributed positively toward model-predicted probability of SLNM. The sum of contributions was − 0.41, leading to a final prediction of SLNM-negative, consistent with the true label. Conversely, Fig. 7D shows the feature contributions for patient #856 (SLNM-positive), where the cumulative SHAP contributions shifted the model output toward a positive prediction.

Web-based application

To facilitate clinical implementation, we developed a user-friendly web application that allows patient data to be entered and processed via desktop computers, laptops, or smartphones. Users simply input the patient’s clinical and pathological information to obtain the predicted probability of sentinel lymph node metastasis. This tool provides convenient risk assessment to offer clinical decision-support. The UTRL application is provided as a web-based tool. The web-based application does not perform real-time GAIN imputation. When limited variables are missing, predefined reference values derived from the training cohort are used as defaults. The reliability indicator may facilitate safer clinical interpretation when incomplete inputs are encountered (Multimedia Appendix 11).

Discussion

Breast cancer incidence has been increasing steadily, with a trend toward younger age at diagnosis [30, 31]. With advances in adjuvant therapies, surgical management of breast cancer is shifting from a “maximum tolerable” approach to a “minimum effective” strategy [32–34]. Although SLNB is less invasive than axillary lymph node dissection (ALND), it still carries postoperative risks, including lymphedema, pain, and impaired upper limb function [35–37]. Clinical cohort and retrospective studies have reported that the SLN positivity rate in early-stage T1–T2 breast cancer patients is corresponding to a negative rate of 70%–82% [6–9]. indicating that a substantial proportion of SLN-negative patients may not derive direct pathological benefit from invasive axillary procedures. Recent randomized trials and systematic reviews have suggested that, in carefully selected low-risk cN0 T1–T2 patients, omission of SLNB does not significantly affect invasive disease-free survival while reducing the incidence of complications such as lymphedema [10, 11, 38]. To enable such “careful selection,” several models have been proposed to predict SLNM. Traditional nomograms based on logistic regression (e.g., Hu et al. [12], Ceylan et al. [13], Abdulla et al. [14], Viale et al. [16]) mainly rely on tumor size, grade, and hormone receptor status. While convenient to use, these models have limited predictive performance and may exhibit heterogeneous performance or “algorithmic bias” across populations stratified by age, molecular subtype, or ethnicity. ML approaches can capture complex nonlinear relationships and help mitigate such biases. Shahriarirad et al. [17] applied a deep learning model to improve predictive performance, but it lacked interpretability, achieving an internal validation AUC of only 0.74. Zhao M et al. combined ultrasound imaging with peripheral blood T-cell subsets to construct an ML model, achieving an internal validation AUC of 0.79, but the requirement for manual pixel-wise segmentation limited clinical applicability. Most studies remain constrained by small sample sizes, limited feature dimensions, and a lack of prospective design or independent external validation, highlighting limitations in generalizability, data heterogeneity, and clinical feasibility.

In this study, we developed an interpretable ML model using routinely available preoperative clinical features to predict SLNM. The model demonstrated robust discriminative performance in the internal validation cohort (AUC 0.892), outperforming traditional nomograms [13] and text-based deep learning models [17]. Previous studies have suggested that “non-visualized” SLNs on lymphoscintigraphy may indicate functional impairment or tumor-related lymphatic blockage (“blocking effect”) [39–41]. Accordingly, we constructed a baseline model incorporating radionuclide tracer uptake and radiologically suspicious axillary lymph nodes, achieving an AUC of 0.65 (95% CI: 0.564–0.729). The final SVM model showed superior performance in the prospective validation cohort (AUC 0.775, 95% CI: 0.667–0.865).

SHAP analysis was applied to interpret feature contributions and improve model transparency. PD curves, ICE curves, and SHAP analysis consistently indicated that younger age and elevated CEA levels were associated with an increased risk of SLNM, which aligns with previous findings [42, 43]. Interestingly, our model suggested that higher ER expression was generally associated with an increased probability of SLNM. This trend does not fully align with all prior studies [44]. Qingyang Li et al. reported ER positivity as a risk factor for SLNM [45], whereas Ruihao Liu et al. observed an opposite association, suggesting a potential protective role of ER positivity [46]. Such variations across studies may be attributable to differences in molecular subtype distributions, covariate adjustments, or the interplay of clinicopathological features within different analytical frameworks. PD curves, ICE curves, and RCS plots in Multimedia Appendix 12 demonstrated that the probability of SLNM increased with higher ER expression, consistent with the univariable analysis in our study and with previous findings reported by Qingyang Li et al. Notably, the distribution of HER2 and PR positivity in our cohort aligns with clinical population data, whereas Ruihao Liu et al. reported atypically high HER2 positivity (88.8%) and low PR positivity (21%), likely reflecting selection or coding biases in their dataset. In our analysis, ER expression was treated as a continuous variable, minimizing information loss and allowing a more nuanced depiction of its relationship with SLNM risk. Subgroup analyses revealed that the impact of ER expression on SLNM varied across different hormone receptor and HER2 statuses, particularly in HER2-positive or AR-negative subgroups. Moreover, RCS curves indicated a non-linear relationship between ER expression and SLNM, highlighting the importance of capturing non-linear and interactive effects in risk modeling. SHAP dependence plots with interaction further elucidated these interactions. For instance, high Ki67 levels (red) amplified the negative SHAP values of low ER expression, partially attenuating the positive effect of ER, whereas low Ki67 (blue) mitigated this effect. AR expression also modulated the influence of ER, with low AR (blue) enhancing the negative effect of low ER (Supplementary Fig. S6, S7). Additionally, in P53 overexpressed tumors, high ER levels contributed to higher SHAP values, whereas in P53 wild-type tumors, the opposite trend was observed. These findings underscore that optimized and interpretable machine learning models can capture complex interactions among clinicopathological features, providing clinically actionable insights for axillary management in breast cancer.

In terms of clinical applicability, beyond demonstrating the model’s decision benefit through DCA, we evaluated the false-negative rates of the actual SLNB (9.5%) and the SLNM prediction model (11.1%). Although the model’s false-negative rate slightly exceeded the clinically acceptable threshold of 10% [47], the operating threshold of the model can be adjusted according to clinical priorities. When the probability threshold was set to 0.49, the FNR decreased to 7.9%, which falls within the clinically acceptable range while maintaining acceptable discriminative performance. These findings suggest that the model can be flexibly optimized to meet clinical safety requirements while preserving practical utility.

This study has several limitations. First, it is retrospective in design; despite strict inclusion and exclusion criteria, residual bias cannot be entirely excluded. Another limitation is that formal inter-observer agreement analysis could not be performed because the imaging features were retrospectively collected from routine clinical reports, where each examination was generally interpreted by a single radiologist in standard practice. Future prospective studies with independent multi-reader assessment are warranted to further evaluate the robustness and reproducibility of imaging feature extraction. Second, the model was developed using data from Chinese patients, and while it was validated prospectively, its generalizability to other populations requires further investigation. Because the prevalence of SLN metastasis in the retrospective training cohort was slightly higher than that observed in the general breast cancer population, caution is warranted when interpreting absolute predicted probabilities. Although model discrimination is theoretically independent of disease prevalence, probability calibration may be affected by prevalence differences. Importantly, we further evaluated the model in an independent prospective validation cohort with a naturally occurring SLN-positive proportion, where consistent discrimination performance was observed. Nevertheless, additional calibration and validation in larger multicenter cohorts are still necessary before routine clinical implementation.

In addition, although an independent cohort from another institution was included, the external validation sample size was relatively small. Therefore, the findings from this cohort should be interpreted as preliminary and exploratory rather than definitive evidence of model generalizability. The primary purpose of this analysis was to provide an initial assessment of cross-center feasibility and potential transportability of the model. The limited number of outcome events in the external cohort may also increase uncertainty in performance estimates. Larger multicenter and international external validation studies will be necessary to more reliably evaluate model stability, calibration, and clinical applicability across diverse populations and practice settings. Finally, only routinely available clinical features from electronic medical records were included to maximize usability, potentially omitting biologically or radiologically relevant markers.

In conclusion, we developed and prospectively validated an interpretable SVM model based on multidomain routine clinical data to predict SLNM in breast cancer patients. The model demonstrated high accuracy and good stability. Combined with its interpretability, this approach may serve as a practical and transparent risk-stratification and decision-support tool to assist individualized axillary management, particularly for patients who are minimally invasive candidates or ineligible for SLNB.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

ALND

Axillary lymph node dissection

AUC

Area under the curve

BC

Breast cancer

CART

Classification and regression tree

CMUFAH

First Affiliated Hospital of Chongqing Medical University

CMUTAH

Third Affiliated Hospital of Chongqing Medical University

DCA

Decision curve analysis

DM

Demographic information

DT

Decision tree

FDR

False discovery rate

GAIN

Generative adversarial imputation network

GBDT

Gradient boosting decision tree

ICE

Individual conditional expectation

IDI

Integrated discrimination improvement

IS

Imaging study

KNN

K-nearest neighbors

LASSO

Least absolute shrinkage and selection operator

LDA

Linear discriminant analysis

LB

Laboratory data

LR

Logistic regression

MI

Mutual information

MICE

Multiple imputation by chained equations

MLP

Multilayer perceptron

mRMR

Minimum redundancy maximum relevance

NB

Naive Bayes

NPV

Negative predictive value

PD

Partial dependence

pmm_logreg_polyreg

Predictive mean matching with logistic/multinomial regression

RCS

Restricted cubic spline

RFE

Recursive feature elimination

RF

Random forest

RP

Relevant pathological data

SHAP

Shapley additive explanations

SLN

Sentinel lymph node

SLNB

Sentinel lymph node biopsy

SLNM

Sentinel lymph node metastasis

SVM

Support vector machine

XGBoost

Extreme gradient boosting

Author contributions

Author Ruoyan Wang: Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft.Author Xiongwu Li: Data curation, Formal analysis, Investigation.Author Hongni Zhu and Author Tingting Xiong: Data curation, Formal analysis.Author Linjun Li and Author Fangdan Li: Investigation.Author Longke Ran and Author Yaying Yang: Conceptualization, Project administration, Supervision, Writing – review & editing.

Funding

This study was sponsored by Natural Science Foundation of Chongqing, China, CSTB2024NSCQ-MSX0178, and the Science and Health Joint Project of Dazu District Science and Technology Bureau, DZKJ2023JSYJ-KWXM1009.

Data availability

The datasets generated and analyzed during the current study contain sensitive patient information and are subject to hospital data protection policies, and therefore are not publicly available. Data access can be requested from the corresponding author upon reasonable request and with appropriate institutional approval.

Declarations

Ethics approval and consent to participate

This retrospective study was conducted in accordance with the principles of the Declaration of Helsinki. All data were anonymized, and the requirement for informed consent was waived in accordance with national legislation and institutional policies. The study protocol was reviewed and approved by the Ethics Committee of [the First Affiliated Hospital of Chongqing Medical University] (Approval No. 2024-468-01).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Conflict of interest

The authors declare that the SLNM prediction website mentioned in this study was developed for research use only and has no commercial purpose. None declared.