Breast cancer (BC) is one of the most common malignancies among women worldwide, with an incidence that continues to rise. Recent studies indicate that advancements in screening technologies and comprehensive treatment strategies have significantly improved survival rates in breast cancer patients (1).However, increasing attention is being paid to postoperative complications.

Breast cancer-related lymphedema (BCRL) is among the most prevalent chronic complications following breast cancer surgery, affecting approximately one-fifth of patients (2). BCRL primarily results from impaired lymphatic drainage, leading to the accumulation of protein-rich fluid in the interstitial space and the development of secondary lymphedema (3, 4). Clinically, patients often present with upper limb swelling, heaviness, pain, progressive fibrosis, and skin hardening (5).

Numerous studies have identified several risk factors associated with BCRL, including axillary lymph node dissection (ALND), adjuvant radiotherapy and chemotherapy, tumor staging, obesity, and age (6, 7). Although these findings have provided valuable insights into the pathophysiology of BCRL, there remains a lack of comprehensive predictive tools that integrate multiple variables to assess individual risk.

In recent years, advanced technologies such as machine learning (ML) have emerged as powerful tools for modeling cancer treatment outcomes and prognostic parameters (8). Compared to traditional statistical approaches, ML methods are better suited for handling high-dimensional data with multiple predictors, offering improved accuracy in personalized risk prediction and demonstrating strong potential for clinical application (9).

This study aims to construct multiple ML-based models to predict the individualized risk of developing BCRL after surgery. In addition, the interpretability of the models is enhanced using Shapley Additive Explanations (SHAP), which enables clinicians to identify key risk factors more clearly, thereby supporting postoperative management and intervention strategies to reduce the incidence of BCRL.

The primary objective of this study is to develop a predictive model for breast cancer-related lymphedema (BCRL) using machine learning algorithms. Twelve clinical variables were selected from an initial set of 38 using LASSO regression to assess the risk of lymphedema in postoperative breast cancer patients. Multiple machine learning models were subsequently implemented and optimized. After comparing performance across AUC, DCA, calibration curves, and PR curves, the logistic regression model demonstrated superior performance compared to other ML models. Furthermore, SHAP values were employed for model interpretability, identifying several key predictors of BCRL. Among them, BMI, axillary lymph node dissection (ALND), and lymph node status showed the highest SHAP values, indicating a greater influence on model prediction and highlighting their clinical relevance.

These findings are consistent with prior research. Multiple retrospective cohort studies have identified BMI as the most significant patient-related risk factor for BCRL and emphasized weight reduction as a preventive strategy (12, 18). Additionally, another study demonstrated that preoperative BMI influences both the incidence and timing of BCRL. Once BCRL occurs, patients with higher BMI tend to experience slower recovery, with lymphedema persisting in 50.0% of cases after 5 years (19).Mechanistically, obesity has been found to negatively affect lymphatic vessel density, endothelial cell proliferation, lymphatic drainage, collecting vessel pumping efficiency, and macromolecular clearance, all of which contribute to impaired lymphatic absorption (20, 21).

In this study, the second most significant risk factor is the level of axillary lymph node dissection. As a surgical procedure that disrupts the lymphatic network, ALND remains strongly associated with BCRL even after adjusting for the number of dissected lymph nodes (22). Despite advances in early diagnosis and treatment of breast cancer, ALND still results in upper limb lymphedema in 30–50% of patients. This condition is characterized by the abnormal accumulation of protein-rich interstitial fluid, leading to swelling and tissue remodeling (23). The use of breast lymphatic level (BLL)-based ALND, involving stepwise excision of lymph nodes in patients with positive axillary involvement, has been shown to reduce surgical extent while preserving oncological outcomes and potentially lowering BCRL incidence (24).

The influence of lymph node status on the development of BCRL remains a topic of ongoing debate. Our findings indicate that both the number of positive lymph nodes and the number of nodes removed during surgery are important risk factors. According to Huang et al., axillary lymph node status significantly impacts treatment and prognosis in breast cancer, likely due to broader dissection ranges, increased numbers of positive nodes, and the use of radiotherapy or chemotherapy (25). These observations are consistent with our results. A high number of positive lymph nodes (>8) may serve as a surrogate marker for extensive lymphatic metastasis and treatment-related injury, often necessitating more aggressive ALND, which increases the risk of BCRL (26, 27). Multivariate analysis conducted by Kwan et al. further supports the prognostic value of pathological lymph node counts in predicting BCRL severity. These identifiable risk factors enable reliable prediction of BCRL and support more tailored treatment planning by clinicians (6, 28).

Additionally, this study is the first to incorporate postoperative health behavior factors—such as rehabilitation exercises, limb management, and lifestyle interventions—into model development, revealing their potential protective effects. Behavioral factors such as “not ignoring upper limb swelling,” “avoiding strenuous activity of the affected limb,” “avoiding lifting heavy objects with the affected limb,” and “avoiding fatigue of the affected limb” are significantly associated with reduced BCRL risk. This may be due to excessive use of the affected limb causing increased muscle tension, which impairs lymphatic return and triggers lymphedema (29). Previous studies have shown that early, structured rehabilitative exercise enhances blood and lymphatic circulation, reducing interstitial lymph accumulation and the severity of lymphedema (30).The current study further verifies these findings using quantitative machine learning models.

In this study, we observed a cumulative incidence of BCRL of 26.63%, which aligns closely with reports from multiple high-quality studies. For example, a systematic review and meta-analysis including 21 studies on BCRL reported a wide range of incidence rates from 6.4% to 76.3% (7). Recent large-scale cohort studies, such as those by Li et al. (31), Wu et al. (32) and Martínez-Jaimez et al. (18), reported pooled incidence rates of approximately 25%–30%, which are highly consistent with our findings. However, some studies focusing on modern surgical techniques, such as the widespread adoption of SLNB, reported lower incidence rates (6.4%–12.5%) (33). The relatively higher incidence in our cohort can be reasonably explained by several key factors. First, the surgical population composition: a considerable proportion of patients in our cohort underwent ALND, which remains a major risk factor for BCRL. Although SLNB has become the standard procedure for early-stage patients, ALND remains the routine approach for patients with clinically positive lymph nodes or metastatic sentinel nodes (34). The proportion of ALND in our study is consistent with other studies focusing on populations requiring this procedure. Second, the sensitivity of diagnostic criteria: we used arm circumference measurement (≥2.0 cm difference) as the diagnostic criterion for BCRL, a method widely used in clinical practice due to its simplicity and good clinical relevance. Compared with stricter objective methods, such as bioimpedance spectroscopy (BIS) or measurements using <2.0 cm thresholds, this method is more sensitive and may capture mild to moderate cases (35). Third, the impact of radiotherapy: a relatively high proportion of patients in our cohort received adjuvant radiotherapy, particularly targeting the axillary or supraclavicular regions. Regional lymph node irradiation is known to significantly increase the risk of lymphatic injury and fibrosis, acting synergistically with surgery to elevate BCRL risk (36). Finally, BCRL is a late-onset complication, with risk persisting for several years postoperatively. The relatively long follow-up period in our study ensured adequate capture of late-onset cases, which may be underestimated in studies with shorter follow-up durations. Taken together, the 26.63% incidence reported in our study accurately reflects the disease burden of BCRL in a modern cohort that includes patients at intermediate to high risk (e.g., those undergoing ALND and radiotherapy), employs sensitive clinical diagnostic criteria, and has been followed longitudinally. This comparison underscores the importance of considering differences in patient populations, treatment modalities, diagnostic criteria, and follow-up strategies when interpreting and comparing BCRL incidence across studies.

Compared with traditional statistical approaches, machine learning models demonstrated superior predictive performance in this study. Among nine tested models, logistic regression showed the best balance between predictive accuracy and classification ability, likely due to the strict inclusion criteria that minimized model noise. Notably, SHAP analysis allowed for the quantification of each predictor’s contribution, enabling the model to be not only predictive but also interpretable. This transparency is essential for clinical risk assessment and facilitates the development of individualized follow-up and intervention plans.

Nevertheless, several limitations should be acknowledged. First, as a retrospective study, selection bias may exist. Second, the dataset was derived from a single center with a relatively small sample size, which may limit generalizability. Third, some variables—such as frequency of limb usage and subjective sensory scores—were based on patient records and follow-up interviews and may be prone to measurement bias. Lastly, the model did not include imaging indicators such as lymphangiography or elastography, which could provide additional objective data and further improve predictive accuracy and clinical relevance in future research.

To address the concern regarding the generalizability of findings derived from a single-center cohort, we compared the main demographic and clinical characteristics of our study population with those reported in large national and international breast cancer cohorts. The mean BMI values in the Non-lymphedema group (24.24 kg/m²) and the Lymphedema group (25.96 kg/m²) in our study were comparable to those reported in two large prospective Chinese cohorts (37, 38) and a multicenter case–control study conducted in the United States (39). In addition, the prevalence of axillary lymph node dissection (ALND) in our cohort was consistent with the rates reported in a recent systematic review and meta-analysis of regional therapies for breast cancer (40), reflecting contemporary surgical practice in which ALND remains applied to selected patients despite the increasing adoption of sentinel lymph node biopsy (SLNB). Meanwhile, the distribution of clinical stages and the proportion of patients receiving adjuvant chemotherapy or radiotherapy were also in line with the characteristics described in a recent retrospective study of breast cancer populations (12). Collectively, these comparisons indicate that although our study was conducted within a single institution, the sample shares similar demographic and clinical features with broader breast cancer populations, supporting the potential external validity and generalizability of our predictive model.

Therefore, future studies should aim to conduct multi-center, prospective cohort investigations and integrate multi-dimensional data, including imaging, genomics, and behavioral monitoring. Embedding such predictive models into clinical follow-up systems to enable dynamic postoperative risk evaluation and personalized alerts will be critical for real-world implementation.

In this study, a machine learning–based individualized prediction model for postoperative lymphedema in breast cancer patients has been successfully developed, demonstrating robust predictive performance and promising clinical applicability. Additionally, SHAP-based interpretation facilitates personalized risk assessment for BCRL among breast cancer patients. This effective, computer-assisted approach may serve as an important tool for precision rehabilitation management and represents a shift from empirical prevention strategies toward data-driven clinical decision-making.