Abstract
Background
Breast cancer is one of the most common malignancies worldwide, with bone metastasis representing its most frequent distant metastatic form, significantly worsening patient prognosis. This study aims to develop a machine learning-based predictive model for accurately assessing the risk of bone metastasis in breast cancer patients, thereby enabling personalized risk stratification, early clinical intervention, and optimized treatment strategies.
Methods
This study utilized the Surveillance, Epidemiology, and End Results database as the primary data source to develop machine learning models for predicting bone metastasis risk in breast cancer patients. Initially, univariate and multivariate logistic regression analyses were conducted to screen key predictive variables; subsequently, eight machine learning algorithms were constructed based on the screening results.10-fold cross-validation employed for hyperparameter optimization. Following training, model performance was evaluated on an internal test cohort and externally validated on 342 real-world cases from an independent hospital cohort. Model assessment incorporated multiple metrics, including area under the curve (AUC), area under the precision-recall curve (AUPRC), decision curve analysis, and calibration curves. Additionally, SHAP analysis was applied to enhance model interpretability, and a web-based calculator was developed based on the optimal model to facilitate clinical application and decision support.
Results
Baseline characteristics across cohorts indicated that the majority of patients were aged over 50 years, female, predominantly with the HR+/HER2− molecular subtype, and exhibited a low incidence of bone metastasis. Univariate and multivariate logistic regression analyses identified key independent risk factors, including age >50 years, higher tumor grade, advanced T stage, N stage, clinical stage and HR-/HER2- subtype factors included radiotherapy, surgery, and married status. The LGB model demonstrated superior performance, achieving an AUC of 0.98 in the training set and 10-fold cross-validation (standard deviation=0.00), 0.98 in the internal validation set, and 0.91 in the external validation set; AUPRC values across the three cohorts were 0.96, 0.79, and 0.87, respectively; decision curve analysis showed excellent net clinical benefit within the 0.1-0.8 threshold range; calibration curves further confirmed high concordance between predicted probabilities and actual event rates. SHAP analysis highlighted surgery as the primary protective factor, followed by N stage, T stage, and radiotherapy as risk enhancers; for example, advanced N stage was associated with positive SHAP values, indicating a significant increase in bone metastasis risk.
Conclusions
This study developed an interpretable LGB model accompanied by a web-based calculator, thereby advancing personalized risk stratification, early detection of bone metastasis and optimized treatment strategies.
Keywords: bone, breast cancer, interpretable, machine learning, metastasis
Introduction
Breast cancer (BC) remains one of the most prevalent malignancies worldwide, with recent epidemiological data indicating a steady rise in incidence rates (1). In 2025, an estimated 316,950 new cases of invasive BC are projected among women in the United States alone, alongside 59,080 cases of ductal carcinoma in situ, reflecting a 1% annual increase over the past decade, particularly among younger women and those with higher body mass indices (2). Globally, the age-standardized incidence rate of breast cancer was approximately 47.8 cases per 100,000 women in 2022, with incidence rates increasing by 1-5% annually in many high-income countries, underscoring the growing public health burden (3). Bone metastasis represents a common and devastating complication, occurring in up to 70% of patients with advanced BC (4), with incidence rates ranging from 3.6% at initial diagnosis to 13.6% within 15 years for early-stage cases (5). Early prediction of bone metastasis is critical, as it enables timely interventions that can improve survival, reduce skeletal-related events such as fractures and pain, and enhance quality of life by facilitating targeted therapies like bone-modifying agents (6). This study addresses the research question of whether machine learning models can accurately predict bone metastasis risk in BC patients, hypothesizing that integrating diverse clinical features through advanced algorithms will outperform traditional methods in prognostic accuracy (7).
Despite substantial advances in BC research, existing predictive models for bone metastasis continue to exhibit important methodological and clinical limitations. Early predictive efforts primarily relied on traditional statistical approaches, particularly logistic regression models incorporating demographic and clinicopathological variables such as age, tumor grade, tumor size, and hormone receptor status. Although these models offer interpretability, they generally assume linearity and independence among predictors, limiting their ability to capture the complex, non-linear biological interactions underlying metastatic progression. As a result, their predictive accuracy and risk stratification performance remain suboptimal, especially in the context of high-dimensional clinical data (8). In recent years, machine learning–based models, including random forests, support vector machines, and gradient boosting algorithms, have been increasingly applied to predict bone metastasis in BC patients. These approaches demonstrate improved discrimination by modeling non-linear relationships and interactions among variables. However, a substantial proportion of published models are developed exclusively using public databases such as SEER, which may introduce selection bias and restrict population heterogeneity. Consequently, the generalizability and external validity of these models are often limited, as they may not adequately reflect real-world clinical diversity in patient characteristics, treatment strategies, and healthcare systems (9). Moreover, few studies incorporate independent external cohorts or cross-regional validation, further constraining clinical translation. The lack of integration of multi-institutional data and real-world clinical settings may lead to models that perform well in retrospective analyses but fail to maintain robustness across different clinical environments (10).
To address these gaps, the Surveillance, Epidemiology, and End Results (SEER) database offers substantial advantages, including comprehensive population-based data on cancer incidence, survival, and treatment for nearly 50% of the U.S. population, enabling robust analysis of large cohorts with long-term follow-up (11, 12). Machine learning further enhances prediction by automating feature selection, handling non-linear relationships, and improving accuracy through techniques like ensemble methods, which have demonstrated 15-25% better performance in cancer prognosis compared to conventional statistics (13).
In this study, we utilize SEER data to train multiple machine learning models for bone metastasis prediction, validate them externally using data from an independent hospital cohort, and select the optimal algorithm based on performance metrics. Additionally, we develop a web-based calculator to aid clinical decision-making. The subsequent sections outline the methodology, present results including model comparisons and validation, and discuss implications for personalized BC management.
Method
Study population and data selection
This retrospective study utilized data from the SEER database (2010-2022) and a Chinese medical institution. The SEER program, maintained by the U.S. national cancer institute, collects population-based cancer incidence and survival data from multiple registries that collectively cover approximately 48% of the U.S. population, ensuring strong representativeness and data reliability. Given that no personally identifiable information was used, patient informed consent and institutional review board approval were not required for this retrospective analysis. Breast cancer cases were identified using the International Classification of Diseases for Oncology, Third Edition (ICD-O-3) primary site codes C50.0–C50.9 and histological codes 8500–8543, which correspond to the main invasive breast carcinoma subtypes according to the world health organization (WHO) classification of tumors of the breast, 5th edition (2019). Patients diagnosed with breast cancer were extracted from the SEER database and subsequently divided into training and internal validation cohorts using a 7:3 ratio. For the Chinese cohort, data were retrospectively collected from the electronic medical records of the First People’s Hospital of Kashi between 2015 and 2023. Inclusion criteria were: (1) pathologically confirmed primary breast cancer; (2) complete clinicopathological and follow-up information; and (3) absence of other concurrent malignancies. Exclusion criteria included: (1) incomplete clinical or follow-up data; (2) diagnosis based solely on imaging without histological confirmation; and (3) presence of another primary cancer. All patient data were manually reviewed by two independent investigators to ensure accuracy and consistency. Clinical variables collected from the Chinese cohort included age, sex, tumor location, histological grade, TNM stage, molecular subtype, treatment modalities (surgery, radiotherapy, and chemotherapy), and bone metastasis status. Data extraction and cleaning were performed following standardized procedures to ensure consistency with the variable definitions used in the SEER database. Missing or ambiguous values were verified through cross-checking with pathology reports and follow-up records. To maintain consistency with SEER database classification, patients from the Chinese medical institution were categorized as “Other” under the race variable. Treatment modalities (surgery, radiotherapy, and chemotherapy) were dichotomized as either “yes” or “no” according to treatment administration status. This study was approved by the ethics committee of the first people’s hospital of Kashi (No. (2025)-47). All research was conducted in accordance with relevant guidelines/regulations and the Declaration of Helsinki, and informed consent was obtained from all participants.
Machine learning model construction and evaluation
Univariate and multivariate logistic regression analyses were employed to evaluate risk factors associated with bone metastasis in the training cohort (14). Odds ratios (ORs) with 95% confidence intervals (CIs) were calculated, with OR > 1 indicating potential risk factors. Variables that demonstrated statistical significance in the multivariate logistic regression analysis were subsequently incorporated as input features for machine learning model development. The machine learning methodology was implemented using Python (version 3.10) and the scikit-learn library (version 0.24). The dataset was randomly partitioned into training and testing sets at a ratio of 7:3. Eight machine learning algorithms were utilized for model construction: Logistic Regression (LR), Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGB), Random Forest (RF), Light Gradient Boosting Machine (LGB), Decision Tree (DT), and Naive Bayes (BNB). We conducted 10-fold cross-validation on the developed algorithm, while employing random search for hyperparameter optimization. The predictive performance of each model was comprehensively evaluated on the testing set using multiple metrics: area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPRC), decision curve analysis (DCA), and calibration curves. This multi-faceted evaluation approach enabled robust assessment of model discrimination, precision, clinical utility, and calibration.
Explainable machine learning methods
To enhance the interpretability of the machine learning model, we applied the Shapley Additive Explanations (SHAP) method, a game theory–based algorithm that quantifies the contribution of each variable to the model’s output (14). SHAP values were computed to determine the relative importance and direction of influence of each feature in the prediction process. The analysis was performed using the SHAP package in Python, generating global and individual-level visualizations of feature importance and interaction effects. This approach allowed for transparent assessment of model behavior and feature relevance across the dataset.
Statistical analysis
Data processing and statistical analyses were conducted using R software (version 4.0.5). Continuous variables were presented as median with interquartile range, while categorical variables were expressed as frequencies and percentages. For comparative analyses, continuous variables were evaluated using independent samples t-tests, and categorical variables were assessed using chi-square tests. All statistical tests were two-sided, with P<0.05 considered statistically significant. This analytical approach was selected to provide clinically meaningful insights into the relationships between independent and dependent variables within our study population.
Result
Baseline characteristics
The baseline characteristics of breast cancer patients in the training set (N = 151,662), internal validation set (N = 64,999), and external validation set (N = 342) are summarized in Table 1. Across all cohorts, the majority of patients were aged over 50 years, and females accounted for more than 99% of cases. Radiation therapy was administered to approximately half of the patients in the training and internal validation sets, but was slightly more common in the external set. Chemotherapy receipt rates were 39.2% and 39.5% in the training and internal validation sets, respectively, but markedly lower at 21.3% in the external set. Surgical intervention was performed in the vast majority of patients across cohorts. Histological grade II was the most prevalent. Stage I disease was the most frequent in the training and internal validation sets, whereas stage III was more common in the external set. The HR+/HER2- molecular subtype was the most common, while HER2 positivity was lower in the external set. The overall incidence of bone metastasis was low. The source code for this study has been made publicly available in Supplementary Material 1.
Table 1.
Baseline characteristics of breast cancer patients in training, internal validation, and external validation cohorts.
| Variables | Training | Internal validation (test) | External validation |
|---|---|---|---|
| N=151662 | N=64999 | N=342 | |
| Age | |||
| ≤50 | 30162 (19.9%) | 12960 (19.9%) | 9 (2.63%) |
| >50 | 121500 (80.1%) | 52039 (80.1%) | 333 (97.4%) |
| Sex | |||
| Male | 1046 (0.69%) | 474 (0.73%) | 1 (0.29%) |
| Female | 150616 (99.3%) | 64525 (99.3%) | 341 (99.7%) |
| Literal | |||
| Biliter | 19 (0.01%) | 4 (0.01%) | 0 (0) |
| Right | 76883 (50.7%) | 32816 (50.5%) | 161 (47.1%) |
| Left | 74760 (49.3%) | 32179 (49.5%) | 181 (52.9%) |
| Race | |||
| White | 103616 (68.3%) | 44737 (68.8%) | 0 (0.0%) |
| Black | 10885 (7.18%) | 4554 (7.01%) | 0 (0.0%) |
| Others | 37161 (24.5%) | 15708 (24.2%) | 342 (100.0%) |
| Radiation | |||
| No | 74685 (49.2%) | 32095 (49.4%) | 148 (43.3%) |
| Yes | 76977 (50.8%) | 32904 (50.6%) | 194 (56.7%) |
| Chemotherapy | |||
| No | 91792 (60.5%) | 39524 (60.8%) | 269 (78.7%) |
| Yes | 59870 (39.5%) | 25475 (39.2%) | 73 (21.3%) |
| Marriage | |||
| No | 67209 (44.3%) | 28925 (44.5%) | 170 (49.7%) |
| Yes | 84453 (55.7%) | 36074 (55.5%) | 172 (50.3%) |
| Surgery | |||
| No | 7715 (5.09%) | 3233 (4.97%) | 22 (6.43%) |
| Yes | 143947 (94.9%) | 61766 (95.0%) | 320 (93.6%) |
| Grade | |||
| I | 38754 (25.6%) | 16734 (25.7%) | 108 (31.6%) |
| II | 67295 (44.4%) | 28844 (44.4%) | 163 (47.7%) |
| III | 45514 (30.0%) | 19381 (29.8%) | 71 (20.8%) |
| IV | 99 (0.07%) | 40 (0.06%) | 0 (0) |
| T | |||
| T1 | 92267 (60.8%) | 39716 (61.1%) | 220 (64.3%) |
| T2 | 46593 (30.7%) | 19851 (30.5%) | 97 (28.4%) |
| T3 | 8605 (5.67%) | 3680 (5.66%) | 12 (3.51%) |
| T4 | 4197 (2.77%) | 1752 (2.70%) | 13 (3.80%) |
| N | |||
| N0 | 106069 (69.9%) | 45572 (70.1%) | 277 (81.0%) |
| SN1 | 33806 (22.3%) | 14434 (22.2%) | 48 (14.0%) |
| N2 | 7339 (4.84%) | 3076 (4.73%) | 10 (2.92%) |
| N3 | 4448 (2.93%) | 1917 (2.95%) | 7 (2.05%) |
| Stage | |||
| Stage I | 79946 (52.7%) | 34540 (53.1%) | 128 (37.4%) |
| Stage II | 51611 (34.0%) | 21952 (33.8%) | 19 (5.56%) |
| Stage III | 15349 (10.1%) | 6488 (9.98%) | 195 (57.0%) |
| Stage IV | 4756 (3.14%) | 2019 (3.11%) | 0 (0) |
| Breast-subtype | |||
| HR-/HER2- | 15991 (10.5%) | 6870 (10.6%) | 29 (8.48%) |
| HR-/HER2+ | 6256 (4.12%) | 2576 (3.96%) | 7 (2.05%) |
| HR+/HER2- | 114185 (75.3%) | 49072 (75.5%) | 288 (84.2%) |
| HR+/HER2+ | 15230 (10.0%) | 6481 (9.97%) | 18 (5.26%) |
| ER | |||
| Negative | 23595 (15.6%) | 10053 (15.5%) | 36 (10.5%) |
| Positive | 128067 (84.4%) | 54946 (84.5%) | 306 (89.5%) |
| PR: | |||
| Negative | 39709 (26.2%) | 16891 (26.0%) | 88 (25.7%) |
| Positive | 111953 (73.8%) | 48108 (74.0%) | 254 (74.3%) |
| HER2: | |||
| Negative | 130176 (85.8%) | 55942 (86.1%) | 317 (92.7%) |
| Positive | 21486 (14.2%) | 9057 (13.9%) | 25 (7.31%) |
| Bone-met | |||
| No | 148633 (98.0%) | 63693 (98.0%) | 322 (94.2%) |
| Yes | 3029 (2.00%) | 1306 (2.01%) | 20 (5.85%) |
Univariate and multivariate logistic regression analysis
In univariate logistic regression analysis, factors significantly associated with an increased risk of bone metastasis in patients with breast cancer included age >50 years, chemotherapy, higher grade, advanced T stage, lymph node involvement, advanced stage, HR-/HER2- subtype, and HER2 positivity (P < 0.001). Protective factors comprised male sex, radiotherapy, marital status, surgery, and PR positivity (P < 0.001). Laterality and race showed no significant associations (P > 0.05). In multivariate logistic regression analysis, independent risk factors included age >50 years (OR: 1.63, 95% CI 1.46 - 1.83, P < 0.001), marital status (OR: 1.39, 95% CI 1.28–1.51, P < 0.001), advanced T stage (T4: OR: 4.96, 95% CI 4.06-5.85, P < 0.001), lymph node involvement (N3: OR: 6.34, 95% CI 5.41-7.44, P < 0.001), advanced stage (stage IV: OR: 3.20, 95% CI 3.00–3.45, P < 0.001), and HR+/HER2+ subtype (OR: 2.29, 95% CI 1.92-2.73, P < 0.001). Protective factors comprised radiotherapy (OR: 0.15, 95% CI 0.13-0.17, P < 0.001) and surgery (OR: 0.03, 95% CI 0.03–0.04, P < 0.001). Gender, chemotherapy, race, and PR status showed no significant associations (P > 0.05) (Table 2).
Table 2.
Univariate and multivariate logistic regression analysis of risk factors for bone metastasis in breast cancer patients.
| Variable | P | OR (95%CI) | P | OR (95%CI) |
|---|---|---|---|---|
| Age | ||||
| ≤50 | 1.00 (Reference) | 1.00 (Reference) | ||
| >50 | <.001 | 1.45 (1.33 ~ 1.58) | <.001 | 1.63 (1.46 ~ 1.83) |
| Sex | ||||
| Male | 1.00 (Reference) | |||
| Female | 0.814 | 0.96 (0.67 ~ 1.37) | ||
| Literal | ||||
| Biliter | 1.00 (Reference) | |||
| Right | 0.057 | 0.51 (0.07 ~ 1.05) | ||
| Left | 0.058 | 0.51 (0.07 ~ 1.08) | ||
| Race | ||||
| White | 1.00 (Reference) | |||
| Black | 0.367 | 1.05 (0.95 ~ 1.16) | ||
| Others | 0.090 | 0.94 (0.87 ~ 1.01) | ||
| Radiation | ||||
| No | 1.00 (Reference) | 1.00 (Reference) | ||
| Yes | <.001 | 0.74 (0.70 ~ 0.79) | <.001 | 0.15 (0.13 ~ 0.17) |
| Chemotherapy | ||||
| No | 1.00 (Reference) | |||
| Yes | 0.059 | 1.07 (0.88 ~ 1.26) | ||
| Marriage | ||||
| No | 1.00 (Reference) | 1.00 (Reference) | ||
| Yes | <.001 | 0.85 (0.80 ~ 0.90) | <.001 | 1.39 (1.28 ~ 1.51) |
| Surgery | ||||
| No | 1.00 (Reference) | 1.00 (Reference) | ||
| Yes | <.001 | 0.02 (0.02 ~ 0.03) | <.001 | 0.03 (0.03 ~ 0.04) |
| Grade | ||||
| I | 1.00 (Reference) | |||
| II | 0.124 | 1.79(0.84 ~ 2.74) | ||
| III | 0.057 | 1.04 (0.96 ~ 1.12) | ||
| IV | 0.157 | 1.82 (0.85 ~ 3.68) | ||
| T | ||||
| T1 | 1.00 (Reference) | 1.00 (Reference) | ||
| T2 | <.001 | 5.03 (4.60 ~ 5.50) | <.001 | 2.44 (2.15 ~ 2.73) |
| T3 | <.001 | 13.99 (12.62 ~ 15.51) | <.001 | 3.52 (2.74 ~ 4.30) |
| T4 | <.001 | 45.46 (41.16 ~ 50.20) | <.001 | 4.96 (4.06~ 5.85) |
| N | ||||
| N0 | 1.00 (Reference) | 1.00 (Reference) | ||
| N1 | <.001 | 6.04 (5.60 ~ 6.50) | <.001 | 3.93 (3.52 ~ 4.38) |
| N2 | <.001 | 7.41 (6.67 ~ 8.23) | <.001 | 4.91 (4.17 ~ 5.79) |
| N3 | <.001 | 15.70 (14.20 ~ 17.35) | <.001 | 6.34 (5.41 ~ 7.44) |
| Stage | ||||
| Stage I | 1.00 (Reference) | 1.00 (Reference) | ||
| Stage II | <.001 | 1.67 (1.49 ~ 1.88) | <.001 | 1.10 1.03 ~ 1.17) |
| Stage III | <.001 | 7.86 (7.06 ~ 8.76) | <.001 | 2.35 (2.05 ~ 2.65) |
| Stage IV | <.001 | 24.98 (22.69 ~ 27.50) | <.001 | 3.20 (3.00 ~ 3.45) |
| Breast-subtype | ||||
| HR-/HER2- | 1.00 (Reference) | 1.00 (Reference) | ||
| HR-/HER2+ | <.001 | 1.74 (1.48 ~ 2.05) | 0.171 | 1.17 (0.94 ~ 1.45) |
| HR+/HER2- | 0.027 | 1.13 (1.01 ~ 1.26) | <.001 | 1.92 (1.65 ~ 2.24) |
| HR+/HER2+ | <.001 | 1.94 (1.71 ~ 2.20) | <.001 | 2.29 (1.92 ~ 2.73) |
| ER | ||||
| Negative | 1.00 (Reference) | |||
| Positive | 0.665 | 1.02 (0.94 ~ 1.11) | ||
| PR: | ||||
| Negative | 1.00 (Reference) | |||
| Positive | 0.576 | 0.98 (0.92 ~ 1.05) | ||
| HER2: | ||||
| Negative | 1.00 (Reference) | 1.00 (Reference) | ||
| Positive | 0.029 | 1.07 (1.01 ~ 1.14) | <.001 | 0.46 (0.32 ~ 0.60) |
Machine learning model construction
To develop a robust predictive model for bone metastasis risk in breast cancer patients, we employed 10-fold cross-validation for hyperparameter optimization. In the 10-fold cross-validation on the training set, the LGB model demonstrated the highest AUC value (AUC = 0.98, SD = 0.00) (Figure 1B). ROC analysis revealed that all models exhibited high discriminative performance on the training set, with AUC values ranging from 0.82 (DT) to 0.98 (LGB). A similar trend was observed on the test set, where AUC values varied from 0.78 (DT) to 0.98 (LGB) (Figures 2A–C). However, performance attenuated on the external validation set, with LGB achieving the highest AUC of 0.91. The PR curves highlighted LGB’s superiority in the precision-recall trade-off, with an average precision score of 0.96 on the training set, and slightly lower scores on the test and validation sets (0.79 and 0.87, respectively) (Figures 3A–C). In the training and test sets, LGB consistently yielded the highest net benefit across threshold ranges of 0.1 to 0.8. On the validation set, LGB maintained superior net benefit at clinically relevant thresholds (0.2-0.6), underscoring its utility in decision-making for high-risk patients (Figures 4A–C). Calibration curves indicated good agreement between predicted probabilities and observed outcomes for LGB across the training, test, and validation sets (Figures 5A–C). Based on its superior AUC, AUPRC, net benefit, and calibration performance across all datasets, particularly in the external validation set, LGB was selected as the optimal model.
Figure 1.
Correlation matrix of clinical variables and comparative performance of eight machine learning algorithms with 10-fold cross-validation for predicting bone metastasis in breast cancer. (A) Pearson correlation matrix illustrating the pairwise correlations among clinical and pathological variables. (B) AUC values for eight machine learning algorithms obtained through 10-fold cross-validation. ROC, receiver operating characteristic curve; AUC, average area under the curve.
Figure 2.
ROC curves of the eight model for predicting bone metastasis in breast cancer across training (A), internal validation (B), and external validation cohorts (C). ROC, receiver operating characteristic curve.
Figure 3.
PR curves of the eight model for predicting bone metastasis in breast cancer across training (A), internal validation (B), and external validation cohorts (C). PR, precision-recall curve.
Figure 4.
DCA curves of the eight model for predicting bone metastasis in breast cancer across training (A), internal validation (B), and external validation cohorts (C). DCA, decision curve analysis.
Figure 5.
Calibration curves of the XGBoost model for predicting bone metastasis in breast cancer across training (A), internal validation (B), and external validation cohorts (C).
Interpretation of machine learning models
Due to the black-box nature of the LGB model, we employed the SHAP method to enhance its interpretability. Figure 6 presents the SHAP summary plot for the LGB model, illustrating the relative importance of surgery-related features in predicting the risk of bone metastasis in breast cancer patients. In this plot, each dot represents a single prediction for an individual patient, with the color indicating the magnitude of the corresponding feature value, and the x-axis denoting the SHAP value, which reflects the direction and magnitude of the feature’s impact on the prediction outcome. Notably, whether surgery was performed exerts the most significant influence on the model’s predictions. For instance, features such as N stage, T stage, and radiotherapy rank prominently in the SHAP value ordering, underscoring their strong predictive capability in distinguishing the occurrence of bone metastasis. Specifically, advanced N stages are typically associated with positive SHAP values, suggesting a potential link to elevated bone metastasis risk; in contrast, radiotherapy correlates with negative SHAP values, implying a possible protective effect against distant metastasis. Overall, the SHAP analysis not only improves the model’s transparency but also provides quantitative insights for identifying high-risk populations for postoperative bone metastasis. In clinical practice, this interpretability enhances the model’s trustworthiness and applicability, particularly in scenarios requiring clear delineation of risk factors and decision-making support.
Figure 6.
SHAP analysis of feature importance and individual prediction explanations in the breast cancer bone metastasis prediction model. (A) SHAP summary plot showing the contribution of each clinical feature to the model output. (B) Mean absolute SHAP values ranking the overall importance of features in predicting bone metastasis. (C, D) SHAP force plots illustrating the contribution of individual features to the predicted risk for two representative patients.
Discussion
This study leveraged large-scale data from the SEER database to develop machine learning models, which were subsequently validated using real-world data to confirm their accuracy. Furthermore, we created a web-based calculator based on the final model, facilitating its clinical dissemination and enabling personalized decision-making for patients. The findings revealed significant associations between bone metastasis in breast cancer patients and several key clinical and demographic factors, including advanced age, marital status, surgical intervention, radiotherapy, breast cancer subtype, HER2 status, clinical staging, T staging, and N staging. These associations are consistent with established biological mechanisms and epidemiological patterns in breast cancer progression, thereby providing a robust foundation for enhancing risk stratification and tailoring individualized treatment strategies.
Advanced age emerges as a prominent risk factor for bone metastasis, with underlying mechanisms attributable to age-related declines in immune surveillance, hormonal imbalances, and the accumulation of genetic mutations-all of which facilitate metastatic dissemination (15, 16). This observation aligns with the multi-step carcinogenesis model, wherein prolonged exposure to carcinogenic stressors heightens the likelihood of epithelial-mesenchymal transition and seeds metastatic cells within the bone microenvironment, particularly in patients over 50 years old (17, 18). Furthermore, older patients often exhibit bone remodeling alterations driven by the senescence-associated secretory phenotype, which amplifies the RANKL/OPG signaling pathway to create a fertile niche for tumor cell colonization (19). Marital status, especially being unmarried or divorced, correlates with an elevated risk of bone metastasis (20–22). This association may operate through psychosocial mechanisms, such as chronic stress-induced cortisol elevation, which suppresses immune function and promotes tumor progression via the NF-κB pathway. According to the biopsychosocial model of cancer outcomes, marital support can buffer depression and enhance adherence to screening and treatment, thereby reducing the metastatic burden. In contrast, social isolation exacerbates inflammation and angiogenesis, fostering distant spread (23, 24). Overall, this underscores the role of social environmental factors in modulating epigenetic changes that influence metastasis.
Surgical interventions exhibit a protective effect against bone metastasis, likely by excising the primary tumor mass and disrupting the formation of pre-metastatic niches through reduced release of circulating tumor cells and cytokines (25, 26). According to Paget’s “seed and soil” hypothesis, resection of the primary lesion interrupts vascular and lymphatic dissemination pathways while achieving superior local control to prevent epithelial-mesenchymal transition induction. Similarly, radiotherapy’s association with lower metastatic risk stems from its induction of DNA damage in residual tumor cells, activation of anti-tumor immunity via immunogenic cell death, and alteration of the bone extracellular matrix to inhibit osteoclast-mediated homing (27). These therapies align with the oligometastasis theory, wherein localized treatments can eradicate micrometastases before they become clinically apparent.
Breast cancer subtypes exhibit distinct risks, with hormone receptor-positive subtypes demonstrating higher bone tropism due to estrogen-driven osteoclast activation and CXCL12/CXCR4 chemotaxis in the bone marrow (28). This is explained by Perou and Sorlie’s intrinsic subtype model, which differentiates subtypes based on gene expression profiles: luminal subtypes favor osteolytic metastases through PTHrP secretion, whereas triple-negative basal-like subtypes, while less bone-specific, are overall more aggressive. HER2 status further modulates this dynamic, as HER2-overexpressing tumors display enhanced invasiveness via PI3K/AKT pathway activation, promoting epithelial-mesenchymal transition and vascular endothelial growth factor-mediated angiogenesis, thereby facilitating bone colonization. Targeted therapies like trastuzumab disrupt this process by inhibiting HER2 dimerization, reducing metastatic potential in line with signal transduction models (29).
Advanced clinical staging, along with higher T and N stages, is strongly associated with bone metastasis (30, 31). This association mirrors the sequential progression outlined in the Halstedian paradigm, where larger primary tumors and lymph node involvement signal increased genomic instability and clonal evolution toward enhanced metastatic capability. From a mechanistic perspective, elevated T staging correlates with hypoxia-induced HIF-1α expression, which upregulates genes involved in invasion and bone mimicry, whereas N staging indicates lymphatic spread that seeds distant sites through tumor-draining lymph nodes (32). These elements are incorporated into the TNM staging system, which forecasts outcomes based on anatomical extent and biological aggressiveness (33).
Given the strong clinical and biological relevance of these staging-related variables, the model’s high discriminative performance is understandable. However, the AUC values approaching 0.98 should still be interpreted with caution, as such high estimates may partly reflect overfitting or dataset-specific patterns. Moreover, discrimination alone does not establish clinical utility. Variables such as surgery and radiotherapy should not be interpreted as causal protective factors against bone metastasis, since they may reflect earlier disease stage, lower tumor burden, or treatment selection bias. Therefore, the model should be regarded primarily as a tool for risk stratification rather than causal inference or direct treatment decision-making. Furthermore, patient-centered outcomes, particularly patient-reported outcomes such as symptom burden, quality of life, functional status, and treatment-related experiences, were not included in the present model, although these indicators are increasingly recognized as important evidence for oncology decision-making (34).
Regarding model performance, LGB outperforms other algorithms in predicting the risk of bone metastasis. This superiority can be attributed to its gradient boosting framework (35). Unlike simpler models, LGB incorporates regularization to prevent overfitting, automatic feature selection through histogram-based algorithms and exclusive feature bundling, as well as ensemble learning that aggregates weak learners to achieve better generalization (36, 37).
Limitation
Nevertheless, several limitations of this study should be acknowledged. First, the SEER database may introduce selection and information bias due to the absence or incomplete recording of important clinical variables, which may affect model accuracy. Second, although real-world data were used for external validation, the limited sample size and single-institution origin may restrict the generalizability of the findings. In addition, regional differences in healthcare systems and treatment patterns may influence patient characteristics. Therefore, despite the good performance observed across cohorts, application of this model in other clinical settings should be interpreted cautiously.
Funding Statement
The author(s) declared that financial support was received for this work and/or its publication. This study was supported partly by the Nature Science Foundation of Hunan Province (No. 2026JJ80464), the Anesthesia Talent Training Project of the Anesthesiology Branch of the Chinese Medical Association (No. Z-2017-24-2504), the "co-PI" project from the Third Xiangya Hospital of Central South University (No. 202405) and the Health Research Project of Hunan Provincial Health Commission (No. 20255998).
Footnotes
Edited by: Izidor Mlakar, University of Maribor, Slovenia
Reviewed by: Anna La Salvia, National Institute of Health (ISS), Italy
Ishak Pacal, Iğdır Üniversitesi, Türkiye
Data availability statement
The original contributions presented in the study are included in the article/SupplementaryMaterial. Further inquiries can be directed to the corresponding author.
Ethics statement
The studies involving humans were approved by the first people’s hospital of Kashi (No. (2025)-47). The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent for participation was not required from the participants or the participants’ legal guardians/next of kin in accordance with the national legislation and institutional requirements.
Author contributions
CF: Writing – review & editing, Conceptualization, Writing – original draft, Data curation, Methodology. MT: Conceptualization, Writing – review & editing, Writing – original draft. ZD: Writing – original draft, Investigation, Software, Writing – review & editing. JP: Investigation, Writing – review & editing, Software, Writing – original draft. GY: Writing – original draft, Writing – review & editing. MF: Software, Writing – review & editing, Conceptualization, Writing – original draft. AT: Software, Writing – original draft, Investigation, Writing – review & editing. BQ: Data curation, Software, Supervision, Writing – original draft, Conceptualization, Writing – review & editing. XZ: Supervision, Methodology, Writing – review & editing, Software, Conceptualization, Investigation, Writing – original draft, Formal Analysis, Project administration.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that generative AI was not used in the creation of this manuscript. Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Supplementary material
The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fonc.2026.1692909/full#supplementary-material
References
- 1. Siegel RL, Miller KD, Fuchs HE, Jemal A. Cancer statistics, 2021. CA Cancer J Clin. (2021) 71:7–33. doi: 10.3322/caac.21654 [DOI] [PubMed] [Google Scholar]
- 2. Hu D, Li Z, Zheng B, Lin X, Pan Y, Gong P, et al. Cancer-associated fibroblasts in breast cancer: Challenges and opportunities. Cancer Commun (Lond). (2022) 42:401–34. doi: 10.1002/cac2.12291 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3. Kim J, Harper A, McCormack V, Sung H, Houssami N, Morgan E, et al. Global patterns and trends in breast cancer incidence and mortality across 185 countries. Nat Med. (2025) 31:1154–62. doi: 10.1038/s41591-025-03502-3 [DOI] [PubMed] [Google Scholar]
- 4. Nie H, Yuan Y, Li J, Xu F, Wang T, Bian L, et al. Occurrence and distribution of bone metastases in 984 metastatic breast cancer patients. Transl Breast Cancer Res. (2021) 2:4. doi: 10.21037/tbcr-20-64 [DOI] [Google Scholar]
- 5. Liede A, Jerzak KJ, Hernandez RK, Wade SW, Sun P, Narod SA. The incidence of bone metastasis after early-stage breast cancer in Canada. Breast Cancer Res Treat. (2016) 156:587–95. doi: 10.1007/s10549-016-3782-3 [DOI] [PubMed] [Google Scholar]
- 6. Hao Y, Zhang F, Ma Y, Luo Y, Zhang Y, Yang N, et al. Potential biomarkers for the early detection of bone metastases. Front Oncol. (2023) 13:1188357. doi: 10.3389/fonc.2023.1188357 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Wang S, Wu W, Lin X, Zhang KM, Wu Q, Luo M, et al. Predictive and prognostic biomarkers of bone metastasis in breast cancer: Current status and future directions. Cell Biosci. (2023) 13:224. doi: 10.1186/s13578-023-01171-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Ye LJ, Suo HD, Liang CY, Zhang L, Jin ZN, Yu CZ, et al. Nomogram for predicting the risk of bone metastasis in breast cancer: A SEER population-based study. Transl Cancer Res. (2020) 9:6710–9. doi: 10.21037/tcr-20-2379 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Zhou X, Zhang J, Wang Y, Cao Z. Survival analysis in male breast cancer with bone metastasis based on the SEER database. Front Oncol. (2022) 12:659812. doi: 10.3389/fonc.2022.659812 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Duwe G, Mercier D, Wiesmann C, Kauth V, Moench K, Junker M, et al. Challenges and perspectives in use of artificial intelligence to support treatment recommendations in clinical oncology. Cancer Med. (2024) 13:e7398. doi: 10.1002/cam4.7398 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Penberthy L, Friedman S. The SEER program's evolution: Supporting clinically meaningful population-level research. J Natl Cancer Inst Monogr. (2024) 2024:110–7. doi: 10.1093/jncimonographs/lgae022 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Kolasseri AE, B V. Comparative study of machine learning and statistical survival models for enhancing cervical cancer prognosis and risk factor assessment using SEER data. Sci Rep. (2024) 14:22203. doi: 10.1038/s41598-024-72790-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Wang K, Hong T, Liu W, Xu C, Yin C, Liu H, et al. Development and validation of a machine learning-based prognostic risk stratification model for acute ischemic stroke. Sci Rep. (2023) 13:13782. doi: 10.1038/s41598-023-40411-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Li C, Liu M, Li J, Wang W, Feng C, Cai Y, et al. Machine learning predicts the prognosis of breast cancer patients with initial bone metastases. Front Public Health. (2022) 10:1003976. doi: 10.3389/fpubh.2022.1003976 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Zhong X, Lin Y, Zhang W, Bi Q. Predicting diagnosis and survival of bone metastasis in breast cancer using machine learning. Sci Rep. (2023) 13:18301. doi: 10.1038/s41598-023-45438-z [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Pan Y, Lin Y, Mi C. Clinicopathological characteristics and prognostic risk factors of breast cancer patients with bone metastasis. Ann Transl Med. (2021) 9:1340. doi: 10.21037/atm-21-4052 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Purushotham A, Shamil E, Cariati M, Agbaje O, Muhidin A, Gillett C, et al. Age at diagnosis and distant metastasis in breast cancer--a surprising inverse relationship. Eur J Cancer. (2014) 50:1697–705. doi: 10.1016/j.ejca.2014.04.002 [DOI] [PubMed] [Google Scholar]
- 18. Yao YB, Zheng XE, Luo XB, Wu AM. Incidence, prognosis and nomograms of breast cancer with bone metastases at initial diagnosis: A large population-based study. Am J Transl Res. (2021) 13:10248–61. [PMC free article] [PubMed] [Google Scholar]
- 19. Anwar SL, Avanti WS, Dwianingsih EK, Cahyono R, Suwardjo S. Risk factors, patterns, and distribution of bone metastases and skeletal-related events in high-risk breast cancer patients. Asian Pac J Cancer Prev. (2022) 23:4109–17. doi: 10.31557/APJCP.2022.23.12.4109 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Aizer AA, Chen MH, McCarthy EP, Mendu ML, Koo S, Wilhite TJ, et al. Marital status and survival in patients with cancer. J Clin Oncol. (2013) 31:3869–76. doi: 10.1200/JCO.2013.49.6489 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. Zhu S, Lei C. Association between marital status and all-cause mortality of patients with metastatic breast cancer: A population-based study. Sci Rep. (2023) 13:9067. doi: 10.1038/s41598-023-36139-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Qian L, Chen Y, Peng M, Xia Y, Zhou T, Hong J, et al. The importance of marital status in the morbidity and prognosis of lung metastasis in newly diagnosed ovarian cancer. J Cancer. (2023) 14:1024–38. doi: 10.7150/jca.83017 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Jin JJ, Wang W, Dai FX, Long ZW, Cai H, Liu XW, et al. Marital status and survival in patients with gastric cancer. Cancer Med. (2016) 5:1821–9. doi: 10.1002/cam4.758 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Chen N, McGrath CB, Ericsson CI, Vaselkiv JB, Rencsok EM, Stopsack KH, et al. Marital status, living arrangement, and survival among individuals with advanced prostate cancer in the International Registry for Men with Advanced Prostate Cancer. Cancer Epidemiol Biomarkers Prev. (2024) 33:419–25. doi: 10.1158/1055-9965.EPI-23-1207 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. Pérez-Fidalgo JA, Pimentel P, Caballero A, Bermejo B, Barrera JA, Burgues O, et al. Removal of primary tumor improves survival in metastatic breast cancer. Does timing of surgery influence outcomes? Breast. (2011) 20:548–54. doi: 10.1016/j.breast.2011.06.005 [DOI] [PubMed] [Google Scholar]
- 26. Zhu E, Zhang L, Wang J, Hu C, Jing Q, Shi W, et al. Personalized surgical recommendations and quantitative therapeutic insights for patients with metastatic breast cancer: Insights from deep learning. Cancer Innov. (2024) 3:e119. doi: 10.1002/cai2.119 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27. Feyer PC, Steingraeber M. Radiotherapy of bone metastasis in breast cancer patients - current approaches. Breast Care (Basel). (2012) 7:108–12. doi: 10.1159/000338724 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. Xiao W, Zheng S, Yang A, Zhang X, Zou Y, Tang H, et al. Breast cancer subtypes and the risk of distant metastasis at initial diagnosis: A population-based study. Cancer Manag Res. (2018) 10:5329–38. doi: 10.2147/CMAR.S176763 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Siregar KB, Al Anas M. Unveiling bone metastasis: Exploring histological subtypes of breast cancer in Indonesia's tertiary referral hospital. Cancer Treat Res Commun. (2023) 37:100764. doi: 10.1016/j.ctarc.2023.100764 [DOI] [PubMed] [Google Scholar]
- 30. Karatas M, Zengel B, Durusoy R, Tasli F, Adibelli Z, Simsek C, et al. Clinicopathologic features of single bone metastasis in breast cancer. Med (Baltimore). (2021) 100:e24164. doi: 10.1097/MD.0000000000024164 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31. Koizumi M, Yoshimoto M, Kasumi F, Ogata E. Comparison between solitary and multiple skeletal metastatic lesions of breast cancer patients. Ann Oncol. (2003) 14:1234–40. doi: 10.1093/annonc/mdg348 [DOI] [PubMed] [Google Scholar]
- 32. Valsecchi ME, Pomerantz SC, Jaslow R, Tester W. Reduced risk of bone metastasis for patients with breast cancer who use COX-2 inhibitors. Clin Breast Cancer. (2009) 9:225–30. doi: 10.3816/CBC.2009.n.038 [DOI] [PubMed] [Google Scholar]
- 33. Kinne DW. Staging and follow-up of breast cancer patients. Cancer. (1991) 67:1196–8. doi: 10.1002/1097-0142(19910215)67:4+<1196::aid-cncr2820671515>3.0.co;2-4 [DOI] [PubMed] [Google Scholar]
- 34. Bellino S, La Salvia A. The importance of patient reported outcomes in oncology clinical trials and clinical practice to inform regulatory and healthcare decision-making. Drugs R D. (2024) 24:123–7. doi: 10.1007/s40268-024-00478-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35. Zhang WH, Tan Y, Huang Z, Tan QX, Zhang YM, Chen BJ, et al. Development and validation of AI models using LR and LightGBM for predicting distant metastasis in breast cancer: A dual-center study. Front Oncol. (2024) 14:1409273. doi: 10.3389/fonc.2024.1409273 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36. Wang W, Sheng R, Liao S, Wu Z, Wang L, Liu C, et al. LightGBM is an effective predictive model for postoperative complications in gastric cancer: A study integrating radiomics with ensemble learning. J Imaging Inform Med. (2024) 37:3034–48. doi: 10.1007/s10278-024-01172-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37. Du G, Lv H, Liang Y, Zhang J, Huang Q, Xie G, et al. Population-based colorectal cancer risk prediction using a SHAP-enhanced LightGBM model. Front Oncol. (2025) 15:1575844. doi: 10.3389/fonc.2025.1575844 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The original contributions presented in the study are included in the article/SupplementaryMaterial. Further inquiries can be directed to the corresponding author.