A Machine Learning Model for Predicting Breast Cancer Recurrence and Supporting Personalized Treatment Decisions Through Comprehensive Feature Selection and Explainable Ensemble Learning

Introduction

Breast cancer remains the most common malignancy among women, with approximately 2.3 million new cases annually, accounting for over 15% of cancer-related deaths¹. Despite advancements in diagnosis and treatment, local recurrence remains a challenge, significantly affecting survival and quality of life.^2,3 Breast cancer recurrence can occur many years after initial treatment, especially in patients with larger tumors or lymph node involvement.⁴ Accurately identifying high-risk patients is crucial for guiding personalized treatment strategies.

Breast cancer is highly heterogeneous, and the risk of local recurrence varies significantly among patients.⁵ Conventional risk models relying on limited clinical and pathological features often lack accuracy and interpretability, especially in complex cases. Managing high-dimensional clinical data while ensuring model robustness remains a key challenge.⁶

Precision medicine has emerged as a strategy to optimize treatment through individualized risk assessment.⁷ However, traditional models struggle to integrate diverse clinical and biological factors, leading to suboptimal treatment decisions. High-risk patients may not receive timely interventions, while low-risk patients could be overtreated, increasing side effects and costs.

Machine learning (ML) offers a data-driven approach to modeling complex interactions and improving recurrence risk prediction. However, traditional statistical models face challenges such as overfitting and limited interpretability. The “black box” nature of many ML algorithms further reduces physician trust.⁸

To address these issues, this study develops a breast cancer recurrence risk prediction model that enhances both accuracy and interpretability. We employ LASSO regression for feature selection to identify key clinical variables while mitigating overfitting.⁹ A stacking ensemble learning approach integrates multiple ML models to enhance predictive performance and stability.¹⁰

To improve model interpretability, we incorporate Shapley additive explanations (SHAP) to quantify the contribution of each feature to recurrence risk.^11,12 Unlike traditional feature importance methods, SHAP provides both global (overall risk factors) and local (individual patient influence) interpretability, bridging the gap between ML models and real-world clinical applications.

This model improves recurrence risk prediction while providing clinicians with transparent decision support, facilitating personalized, evidence-based breast cancer treatment.¹³ The ultimate goal is to develop a clinical decision support system (CDSS); however, further validation through multi-center trials is necessary to ensure clinical applicability.

Materials and Methods

This study aimed to develop a machine learning-based breast cancer local recurrence risk prediction model by integrating extensive clinical data and advanced data analysis techniques, with the goal of improving risk assessment and supporting personalized treatment decisions. The overall research framework (Figure 1) covers data collection, preprocessing, feature selection, model development, and evaluation to ensure the reliability and clinical applicability of the results.

Figure 1 Research flow chart. Abbreviations: LASSO, Least Absolute Shrinkage and Selection Operator; LR, Logistic Regression; SVM, Support Vector Machine; RF, Random Forest; XGB, eXtreme Gradient Boosting; KNN, K-Nearest Neighbors; SHAP, Shapley additive explanation.

Data Collection

Clinical data were retrospectively obtained from 1,131 breast cancer patients recorded in the Kaohsiung Medical University Chung Ho Memorial Hospital Cancer Registry. The dataset included 1,056 patients without recurrence and 75 patients with recurrence, diagnosed and treated between 2001 and 2022. This study was conducted in accordance with the declaration of Helsinki. IRB approval was granted by the Kaohsiung medical university institutional review board (KMUHIRB-E(I)20220126), and the requirement for informed consent was waived due to the retrospective nature of the study, which utilized pre-existing, de-identified clinical data.

Inclusion & Exclusion Criteria

● Inclusion Criteria:

1. Female patients diagnosed with primary breast cancer.

2. Underwent mastectomy or lumpectomy as the primary treatment.

3. Complete medical records with documented recurrence status.

● Exclusion Criteria:

1. Incomplete clinical records or missing follow-up data.

2. Patients diagnosed with metastatic breast cancer at baseline.

This study adheres to retrospective cohort study design principles, using only historical, de-identified patient records without additional data collection or direct patient interaction, aligning with ethical considerations.

Variables Collected

A total of 16 clinical variables were collected, covering:

• Demographics: Age at diagnosis, BMI.
• Pathological features: Tumor size, lymph node status, histological grade, molecular subtype (ER, PR, HER2, Ki-67).
• Treatment details: Type of surgery
• Other factors: Smoking behavior, primary tumor site, laterality.

These data form a robust foundation for analyzing breast cancer recurrence risk factors and model development.

Data Preprocessing

To ensure data quality and model robustness, a multi-step data preprocessing strategy was implemented:

1. Standardization: Continuous variables were standardized to a mean of 0 and a standard deviation of 1 to ensure comparability across features. The formula (1) used was:

(1)

where x represents a data point, μ the mean, and σ the standard deviation.

2. Categorical variable encoding (one-hot encoding): Categorical features were transformed into binary vectors, ensuring that categorical data were processed without implying ordinal relationships.

3. Handling class imbalance (SMOTEENN): Due to the significant imbalance between recurrence (75 cases) and non-recurrence (1,056 cases), resampling was performed using SMOTEENN, a hybrid approach combining:^14,15

• SMOTE (synthetic minority oversampling technique) to generate synthetic minority class samples.
• Edited nearest neighbors (ENN) to remove redundant majority class samples, refining the decision boundary and reducing noise.

Justification for Choosing SMOTEENN

• Compared to pure SMOTE or random undersampling, SMOTEENN provides superior balance while avoiding over-reliance on synthetic data.^16–18
• Compared to ADASYN or borderline-SMOTE, which may oversample near decision boundaries, SMOTEENN ensures more robust and representative augmentation.^16–18
• Final class distribution post-SMOTEENN was adjusted to 0.7: 1 (not fully 1: 1) to enhance model generalization while maintaining sensitivity to minority cases.

SMOTEENN’s effectiveness in handling clinical data imbalance enhances predictive performance and model robustness, ensuring that recurrence cases are adequately represented in the training process.

Feature Selection

Feature selection was performed using LASSO (least absolute shrinkage and selection operator) regression, which applies L1 regularization to eliminate irrelevant or redundant variables, preventing overfitting and improving model generalizability.⁹

Feature Set & Selection Process

● Initial Features (16 total):

○ Demographic: Age at diagnosis, BMI

○ Tumor characteristics: Tumor size, histology behavior, grade, laterality, primary site

○ Lymph node and vascular involvement: Regional lymph node positivity, lymphovascular invasion (LVI)

○ Molecular markers: ER (estrogen receptor), PR (progesterone receptor), HER2, Ki-67

○ Treatment & lifestyle: Surgery method, smoking behavior

● LASSO-selected features (Final 6):

○ Regional lymph node positivity

○ ER status

○ Ki-67

○ Lymphovascular invasion (LVI)

○ Tumor size

○ Age at diagnosis

The LASSO penalty parameter (λ) was optimized using 10-fold cross-validation, testing 50 logarithmically spaced λ values (ranging from 10⁻³ to 10¹), selecting the one minimizing cross-validation mean squared error (MSE).

LASSO was chosen over stepwise regression and other methods because it:¹⁹

• Provides stable feature selection, avoiding the sensitivity to random errors observed in stepwise methods.
• Handles multicollinearity effectively by shrinking redundant predictors to zero, which simplifies the model and improves interpretability.

Machine Learning Model Development

Baseline Models

Five commonly used ML models were implemented:

• Logistic regression (LR)

• Support vector machine (SVM)

• Random forest (RF)

• Extreme gradient boosting (XGB)

• K-Nearest neighbors (KNN)

The dataset was randomly split into 70% training and 30% testing, with 10-fold cross-validation used for hyperparameter tuning.

Integrated Learning Model

To further improve prediction accuracy and model stability, this study used ensemble learning techniques. Ensemble methods include bagging, boosting, and stacking.^20,21 These methods reduce the limitations of a single model by integrating the prediction results of multiple models, thereby generating a more powerful prediction model.

The ensemble learning architecture is shown in Figure 2. The training set data is input into the base classifiers for training, and the predicted probabilities from each classifier are obtained. The dataset was initially divided into training and testing sets in a 7:3 ratio, and the meta-classifier was trained using the predicted probabilities from the base classifiers as input. The final prediction result is generated after training the meta-classifier.

Figure 2 Flowchart of ensemble learning. Abbreviations: LR, Logistic Regression; SVM, Support Vector Machine; RF, Random Forest; XGB, eXtreme Gradient Boosting; KNN, K-Nearest Neighbors.

To improve model performance and stability, we employed an ensemble learning approach utilizing stacking, which integrates multiple base models into a meta-classifier:

• First layer (base models): Predictions were generated using LR, SVM, RF, XGB, and KNN.
• Second layer (meta-classifier): The meta-classifier was trained on the base model outputs to produce final predictions.

Model Evaluation

To comprehensively assess model performance, we employed the following metrics:

● Area under the curve (AUC) – Measures overall classification performance.

● Accuracy (ACC) – Evaluates overall prediction correctness.

● Recall (sensitivity) – Ensures effective detection of recurrent cases.

● 10-fold cross-validation – Used to reduce bias and confirm generalizability.

SHAP Value Analysis

To enhance clinical interpretability, we applied Shapley additive explanations (SHAP) analysis, which provides:

1. Global Interpretability: Identifies which features contribute most to recurrence prediction.
2. Local Interpretability: Assesses how features impact individual patient predictions.

• Consistent, theoretically justified explanations (based on cooperative game theory).
• Both global and local interpretability, enabling personalized clinical decision-making.
• More robust feature explanations, reducing instability compared to alternative methods.

Since LASSO already filtered redundant features, SHAP was applied only to the final selected features to provide clear and clinically meaningful insights.

Results

This study successfully combined advanced machine learning techniques with comprehensive clinical data to develop effective predictive models for breast cancer local recurrence. The clinical data of 1,056 patients without recurrence and 75 patients with recurrence (Table 1) were analyzed, including variables such as age at diagnosis, BMI, and tumor size. These variables not only revealed key risk factors for breast cancer recurrence but also emphasized the significant correlation between tumor size and recurrence risk, further highlighting the importance of careful monitoring.

Table 1 Clinical Characteristics of Patients

Key Predictors

A total of 16 clinical characteristics were initially considered for model development. Using LASSO regression for feature selection, we identified six key features that were most predictive of breast cancer local recurrence (Figure 3a). These features, ranked by their importance to recurrence prediction, were: Regional lymph node positivity, ER status, LVI, Ki-67 index, Age at diagnosis, Tumor size.

Figure 3 (a) Ranking diagram of the importance of feature factors of LASSO. The ranking of important factors selected by LASSO is as follows: Regional Lymph Nodes Positive, ER, LVI, ki-67, Diagnosis Age, Tumor Size; (b) Comparison of the AUC of a single model and the integration of all the features and the LASSO selected features. Abbreviations: BMI, Body Mass Index; LVI, Lymphovascular Invasion; ER, Estrogen Receptor; PR, Progesterone Receptor; HER2, Human Epidermal Growth Factor Receptor 2; LASSO, Least Absolute Shrinkage and Selection Operator; LR, Logistic Regression; SVM, Support Vector Machine; RF, Random Forest; XGB, eXtreme Gradient Boosting; KNN, K-Nearest Neighbors.

These features are critical in predicting local recurrence due to their strong correlation with tumor behavior. For example, regional lymph node positivity indicates potential cancer spread beyond the breast, ER status affects hormone therapy responsiveness, and Ki-67 reflects tumor proliferation rates, all of which are known indicators of aggressive cancer.

Model Performance Evaluation

Figure 3b demonstrates the performance comparison between individual models and the ensemble model, with different feature combinations evaluated using the AUC (area under the curve) metric. The ensemble model consistently outperformed the individual models, particularly when using the LASSO-selected features. The ensemble model achieved an AUC of 0.817, which was significantly higher than the best-performing individual model with an AUC of 0.711. This highlights the superior accuracy and stability of the ensemble model in predicting local recurrence.

Table 2 provides a detailed performance comparison of the five different models (logistic regression, support vector machine, random forest, extreme gradient boosting, and k-nearest neighbors) when using all features and LASSO-selected features. After applying the LASSO feature selection, the ensemble model’s accuracy improved to 73%, with balanced performance in terms of AUC, accuracy (ACC), and recall. These results underscore the ensemble model’s potential for clinical decision support due to its reliability and balanced performance metrics.

Table 2 Integrated Model Evaluation Indicators of the Five Models Created by the Combination of All the Features and the LASSO Features

SHAP Value Analysis

To enhance the interpretability of the model, SHAP value analysis was conducted to evaluate the contribution of each feature to the model’s predictions. SHAP values provide a framework for understanding how individual features impact each prediction, thereby offering greater transparency.

Figure 4a presents the average SHAP values for each LASSO-selected feature, illustrating the importance of regional lymph node positivity, ER status, and Ki-67 as the most influential factors in recurrence prediction. Figure 4b provides a bee colony plot showing the SHAP values for individual samples across various feature ranges, highlighting how these features interact and their specific impact on prediction outcomes. For instance, patients with high SHAP values for regional lymph node positivity were more likely to experience recurrence, while those with lower SHAP values had reduced risk.

Figure 4 (a) Bar plot and (b) Beeswarm plot of feature importance in the ensemble machine learning model. Abbreviations: ER, Estrogen Receptor; LVI, Lymphovascular Invasion.

By using SHAP analysis, actionable insights can be drawn from the model’s predictions. For example, in one patient case with high Ki-67 and positive regional lymph nodes, the SHAP value analysis revealed that these factors significantly increased the predicted recurrence risk, informing the clinical decision to recommend more aggressive treatment. This case highlights how SHAP analysis makes the model more interpretable and useful for personalized treatment planning.

Feature Combination and Model Optimization

The breast cancer local recurrence prediction model developed in this study demonstrated high predictive accuracy and stability, particularly when combining LASSO feature selection with ensemble learning techniques. The ensemble approach, combined with key clinical features, significantly improved the model’s performance. Additionally, SHAP value analysis enhanced the model’s interpretability, making it more transparent and applicable for clinical use. These findings not only offer a strong scientific foundation for clinical decision-making but also reinforce the model’s potential for real-world medical applications by providing clinicians with an effective tool to identify high-risk patients and develop personalized treatment strategies.

Discussion

Key Findings and Contributions

This study developed a local risk prediction model for breast cancer recurrence by combining LASSO feature selection and ensemble learning technology. The model demonstrated significant predictive performance, especially in improving accuracy and model interpretability. The key results included the following: (1) Six key clinical variables were screened by LASSO and ranked in order of importance: regional lymph node positivity, ER status, lymphovascular invasion (LVI), ki-67, and diagnostic age. (2) Ensemble learning technology significantly improved the stability and accuracy of the model, with the AUC value reaching 0.817. (3) SHAP analysis enhanced the interpretability of the model, enabling clinicians to better understand the significance of the prediction result logic to develop a more personalized treatment plan. These technological innovations demonstrate the potential application value of this model in the personalized treatment of breast cancer.

The six factors screened by LASSO cover many aspects of breast cancer recurrence risk, including the biological behavior of the tumor (ER, ki-67), indicators of cancer spread (positive regional lymph nodes, LVI), size and stage of the tumor (tumor size), and individual characteristics (age at diagnosis) of the patients. The analysis of the clinical significance of these six important factors is explained below.

Positive regional lymph nodes are among the most important prognostic indicators of breast cancer.²² When cancer cells metastasize to lymph nodes, the risk of recurrence and distant metastasis increases significantly. Patients with positive lymph nodes usually require more aggressive treatment, such as adjuvant chemotherapy or radiotherapy, to reduce the risk of recurrence. The importance of this variable is that it directly reflects the extent of cancer spread and predicts a poor prognosis.²³

The estrogen receptor (ER) status is the most common hormone receptor status in breast cancer patients, and it is usually associated with a better treatment response and a lower risk of recurrence.²⁴ ER-positive patients respond better to endocrine therapy (such as tamoxifen or aromatase inhibitors). The ER status plays a key role in treatment decisions and is closely related to the biological behavior of the disease.²⁵

Lymphovascular invasion (LVI) refers to the invasion of cancer cells into lymphatic vessels or blood vessels and is an early sign of tumor metastasis. LVI-positive patients have a greater risk of recurrence because cancer cells may metastasize to other locations through the lymphatic or vascular system. Therefore, LVI is an important predictor for determining the need for adjuvant therapy.²⁶

ki-67 is a cell proliferation marker that reflects the proliferation rate of tumor cells. High ki-67 levels usually predict faster tumor growth and a greater risk of recurrence and are usually associated with poor prognosis. Clinically, ki-67 is used to predict whether patients need more aggressive treatment strategies. High ki-67 usually indicates that a tumor has more invasive characteristics.²⁷

Age at diagnosis is an important variable for the prognosis of patients with breast cancer. Young patients (usually less than 40 years old) usually have more invasive breast cancer and a higher risk of recurrence. In contrast, tumors in elderly patients usually grow slowly, but their treatment tolerance is lower; therefore, individualized treatment is needed. Age at diagnosis helps to determine the intensity of treatment and the evaluation of prognosis.²⁸

Tumor size is an important factor for the prognosis of patients with breast cancer. Larger tumors are generally associated with a greater risk of recurrence and poorer prognosis.²² Patients with tumors larger than 2 cm are considered at greater risk and may require more aggressive treatment, such as adjuvant chemotherapy or radiotherapy after surgery. Tumor size reflects the invasiveness and progression stage of cancer at the time of diagnosis.²⁹

Analysis of the Key Influencing Factors

In this study, the most important influencing factors were regional lymph node positivity and ER status. These features have been confirmed as important predictors of breast cancer recurrence in most studies^30,31 and had the highest weight in the LASSO feature selection process in the present study; in particular, positive regional lymph nodes contributed the most to model prediction. In contrast, other variables, such as ki-67, tumor size, diagnosis age and lymphovascular invasion (LVI), still play important roles in prediction, but their relative weights are low, indicating that the impact on the final forecast result is relatively small.

Highlights

This study used LASSO feature selection technology to screen six clinical features with significant predictive power for the risk of local recurrence of breast cancer, which significantly improved the accuracy of the model and reduced the risk of overfitting. The combination of ensemble learning technology further improved the stability and performance of the model and finally reached an AUC value of 0.817, which was significantly better than those of the other single models. Moreover, SHAP value analysis was introduced in the present study, which enhances the transparency and interpretability of the model, enables clinicians to more clearly understand the logic of the prediction results, and is helpful for clinical application. This predictive model can help clinicians accurately identify high-risk patients to develop individualized treatment strategies and promote the development of personalized treatments for patients with potential clinical practicability.

Comparison With Existing Literature

In comparison with existing studies, the innovation of the present study lies in the significant improvement in the accuracy and stability of breast cancer recurrence risk prediction through the combination of LASSO feature selection and ensemble learning techniques. While many studies have already established the correlation between individual factors, such as ER status, positive regional lymph nodes, LVI, Ki-67, tumor size, and age at diagnosis, and breast cancer recurrence,^32–34 this study is unique in integrating these factors into a machine learning framework. Furthermore, this study goes beyond simply identifying these factors by quantifying their contribution to model predictions using SHAP analysis. This approach not only highlights the predictive power of individual variables like ER status and lymph node positivity but also provides transparency and explainability that were lacking in previous models.

Previous studies have also explored the integration of LASSO-based feature selection with ensemble learning for breast cancer recurrence prediction. For example, Lee et al (2023) developed a radiomics-based machine learning model to predict locoregional recurrence (LRR) using LASSO for feature selection and an ensemble stacking approach, achieving an AUC of 0.78, which was higher than individual models (0.61–0.70).³⁵ This aligns with our findings, as our LASSO-selected features combined with ensemble learning yielded an AUC of 0.817, further demonstrating the advantage of integrating LASSO with ensemble techniques for improving predictive performance and stability.

Unlike traditional statistical models, which often struggle with high-dimensional clinical data, this study demonstrates the capability of ensemble learning to process such data efficiently while yielding actionable clinical predictions. Previous studies have demonstrated the effectiveness of ensemble learning in medical AI.^10,36 By integrating multiple classifiers, ensemble models improve both predictive accuracy and robustness, making them especially valuable for analyzing complex clinical datasets. Moreover, the use of SHAP analysis allows clinicians to understand how specific factors contribute to the risk of recurrence on an individual patient basis, offering insights that were previously unavailable. This combination of predictive accuracy and interpretability marks a significant advancement over prior studies, providing a more powerful and clinically relevant prediction framework.

Analysis of Key Figures or Tables

Diagram analysis in this study revealed that the ensemble learning model had a significant advantage in processing and assessing the risk of breast cancer recurrence. In particular, the AUC curve shown in Figure 3b reveals that the AUC value of the ensemble model reaches 0.817, which is significantly higher than the performance of the other single models (the highest AUC value is 0.711). In addition, the effects of different feature combinations on the model are shown in detail in Table 2, which validates the ability of the feature set screened by LASSO to improve the prediction accuracy.

Linking With Practical Applications

The design of this prediction model fully addresses the practical needs of clinical applications, particularly in terms of interpretability. The model allows clinicians to understand the logic behind the predictions, which is critical for clinical decision-making. The introduction of SHAP values enhances transparency, helping doctors quickly identify high-risk patients and facilitating the implementation of personalized treatment strategies. Moreover, the model’s ability to handle large volumes of heterogeneous clinical data makes it suitable for large-scale breast cancer risk prediction.

By using SHAP analysis, this study significantly improves clinical decision-making by making the model’s predictions more interpretable SHAP values break down the overall prediction, revealing the individual contribution of features such as ER status, lymph node positivity, and Ki-67 to a specific patient’s recurrence risk. For instance, in a patient with high regional lymph node positivity and elevated Ki-67 levels, SHAP values highlight the dominant role these factors play in the predicted high recurrence risk.

This interpretability is crucial for personalized treatment, as it enables doctors to tailor interventions based on a patient’s unique risk profile. Understanding the exact contribution of each risk factor allows clinicians to make informed decisions on whether to escalate treatment, increase monitoring, or adjust therapy. The ability to translate complex predictions into clear, actionable insights supports precision medicine, leading to more targeted and effective treatments.

Research Limitations and Challenges

Although this study achieved good prediction results, it is still limited. First, the data in this study came from a single hospital and may have geographic and population specificity, which affects the generalizability of the model. Second, the sample size of the present study is relatively limited, especially in terms of data from relapsed patients, which may affect the application of the model in larger datasets. Future studies should consider increasing the sample size and carrying out multicenter cooperation to improve the extensive applicability of the model. In addition, although the SHAP value enhances the interpretability of the model, it is still a postmodel explanation method. In the future, we can explore the construction of a model framework with interpretability itself. Another important factor to consider is that advancements in breast cancer surgical techniques and treatment strategies from 2001 to 2022 may have influenced recurrence risk. Changes in systemic therapies, radiotherapy techniques, and surgical approaches over time could introduce variability in recurrence patterns, which this study could not fully account for. Future research should consider stratifying patients by treatment era to assess the impact of evolving clinical practices on recurrence prediction. Lastly, before clinical implementation, real-world validation through large-scale prospective studies and multi-center trials is essential. Ensuring the model’s robustness across diverse healthcare settings will confirm its generalizability and clinical utility. Without external validation, the model risks being overfitted to a specific dataset, limiting its broader applicability. Future research should prioritize external validation to establish its reliability in real-world clinical practice.

Future Research Directions

Future research directions include the following: (1) expand data sources and carry out multicenter data collection to improve the generalizability of the model; (2) introduce multiomics data such as genetic data and image data to further improve the accuracy of prediction; (3) explore more machine learning technologies based on interpretability so that model construction has transparency in the process rather than relying on later explanation methods; and (4) explore more advanced deep learning technologies and reinforcement learning technologies to further improve the performance of the model in dealing with heterogeneous data. (5) Compare resampling techniques, such as ADASYN and Borderline-SMOTE, to improve class balance and model generalizability.

Conclusion

This study successfully developed a machine learning model integrating LASSO feature selection with ensemble learning, significantly improving the prediction accuracy of local recurrence risk in breast cancer patients. By leveraging the strengths of multiple models, the ensemble learning approach demonstrated superior stability and predictive performance compared to individual models, achieving an AUC of 0.817. These findings underscore the effectiveness of multimodel integration in handling complex clinical data and enhancing predictive reliability.

Beyond predictive accuracy, this study also prioritized interpretability through SHAP analysis, enabling clinicians to understand the contribution of each feature to recurrence risk predictions. This interpretability facilitates the identification of high-risk patients and supports the development of personalized treatment strategies, reinforcing the model’s potential as a decision-support tool in clinical practice. The ability to provide both robust predictive performance and explainability strengthens its applicability in real-world oncology settings.

Future research should aim to enhance the model’s generalizability by incorporating additional data sources, such as genomic and imaging data, and validating its performance across multicenter datasets. Additionally, further exploration of advanced ensemble learning techniques or deep learning architectures may further optimize the accuracy of breast cancer recurrence risk prediction, ultimately supporting more precise and personalized treatment strategies for breast cancer management.