Identifying low-risk breast cancer patients for axillary biopsy exemption: a multimodal preoperative predictive model

Abstract

Background: As the most prevalent female malignancy worldwide, breast cancer frequently involves axillary lymph node metastasis (ALNM), which critically affects therapeutic algorithms. Current guidelines mandate preoperative ultrasound-guided axillary biopsy for suspicious lymph nodes, potentially exposing some low-risk patients with negative results to invasive risks. To optimize the utilization of biopsy, this study established a multimodal predictive framework that preoperatively assesses axillary lymph node (ALN) status, thereby triaging candidates for ultrasound-guided axillary biopsy.

Methods: We conducted a retrospective single-center analysis of 703 breast cancer patients who underwent ultrasound-guided axillary biopsy with subsequent definitive surgery at the First Affiliated Hospital of Xi'an Jiaotong University (07/2020-05/2023). Following rigorous application of the inclusion/exclusion criteria, 439 eligible patients were randomized into training (n = 307, 69.9%) and validation (n = 132, 30.1%) cohorts. Axillary surgical pathology served as the reference standard for categorizing lymph node status. Multivariable predictors identified through the least absolute shrinkage and selection operator (LASSO) and logistic regression informed the construction of a clinically deployable nomogram. Model discrimination was quantified via receiver operating characteristic (ROC) analysis with area under the curve (AUC) calculations. The optimal threshold was determined using the maximum Youden index.

Results: LASSO, univariate, and multivariate logistic regression analyses revealed that estrogen receptor (ER) status (P = 0.007), ALN cortical-medullary boundary (P = 0.012), ALN cortical thickness (P < 0.001), short-axis diameter (P = 0.032), and the BI-RADS category on magnetic resonance imaging (MRI) (P = 0.021) were independent predictors of non-ALNM. A nomogram was constructed based on these factors. The multimodal model demonstrated excellent discrimination with AUCs of 0.955 (95% CI 0.926-0.983) and 0.905 (95% CI 0.832-0.978) for the training and validation cohorts, respectively. The model achieved a maximum Youden index of 0.7789 with an optimal threshold of 0.3958.

Conclusion: Our multimodal predictive model integrates clinicopathological profiles with imaging biomarkers (ultrasound and magnetic resonance imaging). This model holds promise for preoperative axillary risk stratification in breast cancer patients, thereby identifying candidates suitable for axillary biopsy exemption, while its application serves as a reference for personalized and refined axillary management.

Keywords: Axillary lymph node status; Multimodal model; Noninvasive; Ultrasound-guided biopsy.

Fig. 1

Flowchart of patient inclusion and exclusion criteria in the study. US: ultrasound; MRI: magnetic resonance imaging; ALN: axillary lymph node; UpN0: ALNs were negative on ultrasound-guided biopsy; UpN + : ALNs were positive on ultrasound-guided biopsy; SLNB+: pathology of sentinel lymph node biopsy was positive; SLNB–: pathology of sentinel lymph node biopsy was negative; ALNM: axillary lymph node metastasis; non-ALNM: without axillary lymph node metastasis

Fig. 2

Proportion of non-ALNM and ALNM (a). Percentages reflect proportions in the ALNM subgroup (b). ALNM: axillary lymph node metastasis; LN + : positive lymph node(s)

Fig. 3

Cross-validation plot for the penalty term (a). Plots for LASSO regression coefficients over different values of the penalty parameter (b)

Fig. 4

A nomogram for predicting non-ALNM. Variables include estrogen receptor (ER) status, axillary lymph node (ALN) cortical–medullary boundary, ALN cortical thickness, ALN short-axis diameter, and MRI BI-RADS classification. Each variable is assigned points on the upper axis (0–100), which are summed to calculate the total points (0–240) on the lower axis. The final probability of non-ALNM is projected onto the corresponding scale (0.1–0.9), with higher total points indicating a lower likelihood of metastasis

Fig. 5

Predictive performance and clinical utility of the model. ROC curve for the training group, The AUC is 0.955 (95% CI 0.926–0.983) (a). ROC curve for the validation group, with an AUC of 0.905 (95% CI 0.832–0.978) (b). DCA for the training group (c). DCA for the validation group (d). AUC: area under the ROC curve; DCA: decision curve analysis; Net benefit, calculated as the weighted difference between true positives and false positives, adjusted for risk threshold preferences