An explainable predictive ML model for axillary lymph node metastasis in breast cancer based on multimodal data

J Transl Med 2025 Metastasis Prediction 10 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Why Lymph Node Status Matters in Breast Cancer

Breast cancer is the most commonly diagnosed malignancy in women worldwide, affecting approximately 2.1 million women annually. Accurate preoperative assessment of axillary lymph node metastasis (ALNM) is critical for choosing the right treatment approach and predicting patient outcomes.

The traditional method for evaluating lymph node involvement, axillary lymph node dissection (ALND), carries significant risks including lymphedema, restricted shoulder movement, and nerve damage -- all of which reduce quality of life. A less invasive alternative, sentinel lymph node biopsy (SLNB), is now the standard of care but still carries risks: around 7% of patients develop lymphedema and 8.7% experience sensory deficits.

Current non-invasive imaging methods such as MRI and ultrasound (US) are commonly used to evaluate lymph node status but have important limitations. MRI has a false-negative rate of roughly 18%, and both methods depend heavily on radiologist expertise, which introduces variability in results.

TL;DR: Accurate, non-invasive prediction of axillary lymph node spread is essential in breast cancer to guide treatment while minimizing surgical complications.
Page 2
The Limits of Existing AI Approaches

In recent years, machine learning (ML) and deep learning (DL) have shown remarkable potential in medical imaging and healthcare diagnostics. Many current AI-based models rely on radiomics, which automatically extracts large numbers of quantitative features from medical images.

However, radiomics-based approaches face significant practical limitations. Feature extraction typically requires complex preprocessing pipelines that are time-consuming and difficult for clinicians to interpret. The resulting models often behave as 'black boxes', providing predictions without explaining the reasoning behind them.

These transparency issues limit adoption in clinical settings, where physicians need to understand why a model is making a prediction before acting on it. There is a pressing need for predictive tools that are both accurate and interpretable using data already collected in standard clinical care.

TL;DR: Existing AI diagnostic models often lack interpretability and rely on complex pipelines that hinder their routine use in clinical practice.
Pages 2-3
Study Design and Patient Population

This retrospective study analyzed clinical data from 401 breast cancer patients treated at a university hospital between January 2019 and May 2024. All patients had pathologically confirmed breast cancer and had undergone preoperative ultrasound and MRI examinations before axillary lymph node assessment.

Patients were randomly divided into a training cohort (70%, 280 patients) and a validation cohort (30%, 121 patients). The training cohort included 161 patients without lymph node spread and 119 with confirmed metastasis, while the validation cohort had 68 negative and 53 positive cases.

Variables collected included patient age, tumor size and location, ultrasound and MRI imaging features, and molecular biomarkers including estrogen receptor (ER), progesterone receptor (PR), HER2, and Ki-67. Pathological confirmation via biopsy or surgical resection served as the gold standard for lymph node status.

TL;DR: A retrospective dataset of 401 breast cancer patients with combined imaging and pathological data was split into training and validation cohorts.
Page 3
Feature Selection and Model Building

To identify the most predictive variables, the researchers used LASSO regression (Least Absolute Shrinkage and Selection Operator) combined with 10-fold cross-validation. This approach penalizes model complexity and helps select only the most relevant predictors from the initial set of 11 candidate features.

LASSO analysis retained five key variables: lymph node parenchymal thickness, lymph node enlargement on MRI, tumor width, and the molecular markers ER and PR. These five features then served as inputs to train all predictive models.

Ten different machine learning classifiers were tested, including Naive Bayes, logistic regression (LR), decision trees (DT), random forest, support vector machines (SVM), and XGBoost. Models were evaluated using standard metrics including the area under the ROC curve (AUC), accuracy, precision, recall, and F1-score.

TL;DR: LASSO regression narrowed 11 candidate variables to 5, which were then used to train and compare 10 different machine learning classifiers.
Pages 3-5
Model Performance: Naive Bayes Leads

Among all ten models tested, the Naive Bayes classifier achieved the highest predictive performance. On the validation cohort, it reached an AUC of 0.902 (95% CI: 0.839-0.965), with accuracy of 82.6%, precision of 83.4%, and an F1-score of 82.0%. This outperformed logistic regression (AUC 0.842) and random forest (AUC 0.815).

Additional analyses including Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) further confirmed the Naive Bayes model's superiority over logistic regression (NRI = 105.1%, IDI = 0.213). The Naive Bayes model significantly outperformed most other models in AUC comparisons (P = 0.000-0.006).

One notable finding was that the Naive Bayes model showed poorer calibration than logistic regression on the Hosmer-Lemeshow goodness-of-fit test, suggesting its probability estimates may not perfectly match observed outcomes. However, the authors note this test should be interpreted alongside other metrics including decision curve analysis (DCA), which showed XGBoost providing the highest net clinical benefit.

TL;DR: The Naive Bayes classifier achieved the best overall discrimination performance with an AUC of 0.902, outperforming nine other machine learning models.
Page 5
SHAP Analysis: Interpreting the Model

SHapley Additive exPlanations (SHAP) were applied to make the Naive Bayes model's predictions transparent and interpretable. SHAP values quantify how much each input feature contributes to a specific prediction, either pushing it toward higher or lower risk.

Global SHAP analysis ranked the features by overall importance: lymph node parenchymal thickness was the most influential predictor, followed by lymph node enlargement on MRI, tumor width, ER, and PR. Higher values of lymph node thickness and tumor width were consistently associated with greater probability of lymph node metastasis.

At the individual patient level, SHAP provided local explanations -- showing exactly how each feature contributed to that specific patient's predicted risk. This dual global-and-local interpretability makes the model transparent for clinicians, enabling personalized risk communication and supporting trust in the tool's recommendations.

TL;DR: SHAP analysis revealed that lymph node parenchymal thickness and tumor width are the dominant predictors, with the model offering both population-level and patient-specific interpretations.
Pages 6-8
Why Simple Features Can Outperform Complex Models

The study's findings challenge the assumption that model complexity correlates with performance. The Naive Bayes model, using only five clinically accessible features, outperformed deep learning radiomics models from prior studies that relied on hundreds of automatically extracted image features and complex neural networks.

A key insight was the diagnostic value of lymph node cortical thickness, using a threshold of greater than 3 mm. This criterion is biologically grounded: tumor cells initially deposit near the marginal sinus of the lymph node, causing cortical thickening, before growing inward. This mechanistic rationale supports the feature's strong predictive power and contrasts with older criteria that used the ratio of lymph node axes.

The study also found that tumor width -- rather than the more commonly studied tumor length -- predicted lymph node metastasis, suggesting that lateral spread into breast tissue may be a more meaningful indicator of tumor invasiveness. Additionally, while ER and PR status was not independently significant in logistic regression, including them still improved overall model performance.

TL;DR: A simple five-feature model using routinely available clinical data can match or outperform complex radiomics-based deep learning models.
Pages 8-10
Addressing the Black Box Problem in Clinical AI

One of the most significant barriers to adopting machine learning in clinical medicine is the 'black box' problem -- models that produce predictions without explaining their reasoning. Clinicians are understandably hesitant to rely on opaque tools when making decisions that directly affect patient care.

This study addresses that concern by integrating SHAP interpretability into the predictive workflow. By providing both global (which features matter most overall) and local (how features influenced a specific patient's result) explanations, the model becomes a transparent decision-support tool rather than an unexplainable oracle.

All input features -- lymph node measurements, tumor dimensions, and molecular markers -- are routinely collected during standard hospitalization, meaning the model requires no additional tests or specialized software. This practicality makes it well-suited for integration into existing clinical workflows across different healthcare settings.

TL;DR: Combining SHAP interpretability with clinically routine data transforms this model into a transparent, practical tool that clinicians can trust.
Pages 6, 10
Clinical Impact: Avoiding Unnecessary Surgery

Accurate preoperative prediction of axillary lymph node status has direct implications for surgical planning. In patients predicted to be lymph node-negative, avoiding ALND can prevent lymphedema and other long-term complications. This is especially important for patients with clinically node-negative (cN0) breast cancer, where definitive lymph node status cannot be determined by physical examination alone.

The model provides a non-invasive alternative or complement to SLNB, potentially reducing the number of unnecessary biopsies and surgical procedures. By offering individualized risk scores, it can help stratify patients into those who need aggressive axillary intervention versus those who may safely forgo it.

Physicians can also use the tool to guide conversations with patients about treatment options, presenting personalized risk assessments supported by transparent feature contributions. This level of explainability aligns with the principles of shared decision-making in modern oncology care.

TL;DR: The model could help spare breast cancer patients from unnecessary lymph node surgery by providing accurate, non-invasive, and explainable preoperative risk predictions.
Pages 10-11
Conclusions and Future Directions

This study successfully developed and validated an explainable machine learning model for predicting axillary lymph node metastasis in breast cancer. The Naive Bayes classifier, trained on five clinically accessible features, achieved an AUC of 0.902 in the validation cohort -- demonstrating that predictive accuracy and interpretability can coexist.

Several study limitations must be acknowledged. The retrospective single-center design limits generalizability, and the model used only five features, potentially omitting important predictors such as histological type. Additionally, sample size imbalance between positive and negative cases could affect model robustness.

Future work will aim to incorporate a broader set of clinical characteristics, additional imaging features such as Doppler flow and contrast-enhanced ultrasound, and multi-center datasets to improve the model's generalizability. The research team also plans to validate the model prospectively to better establish its real-world clinical utility.

TL;DR: The Naive Bayes model achieves strong predictive performance with full interpretability, with plans to expand it using more data sources and multi-center validation.
Citation: Open Access, 2025. Available at: PMC12341081.