Explainable machine learning versus known nomogram for predicting non-sentinel lymph node metastases in breast cancer patients: A comparative study.

INTRODUCTION

Axillary lymph node dissection (ALND) is the standard of care for breast cancer patients with positive sentinel lymph nodes (SLN), which are the first lymph nodes that drain the breast. However, many patients with positive SLNs may not have additional positive nodes, making the prediction of non-sentinel lymph node (NSLN) metastasis challenging. Reliable prognostic tools are essential for accurately assessing NSLN metastasis. The Memorial Sloan Kettering Cancer Center (MSKCC) nomogram has demonstrated effectiveness in this context, but it requires further evaluation within the Iranian breast cancer population. While ALND remains the gold standard, its unnecessary application in patients without evidence of additional positive nodes raises concerns due to potential complications such as lymphedema, nerve injury, and shoulder joint dysfunction. Furthermore, integrating Artificial Intelligence (AI) and Machine Learning (ML) techniques presents an opportunity to enhance the precision of NSLN metastasis predictions.

METHOD

This study conducts an extensive comparative analysis between the MSKCC nomogram and various ML models to predict NSLN metastasis, utilizing a dataset of Iranian breast cancer patients. Employing eXplainable Artificial Intelligence (XAI) methodologies, we analyzed 16 clinical features across a cohort of 183 patients. Our methodology includes rigorous statistical evaluations and the training and validation of ML models to assess the precision and robustness of these models compared to the MSKCC nomogram.

RESULTS

Our analysis revealed that the Random Forest (RF) model outperformed the MSKCC nomogram, achieving an accuracy of 72.2 % and an AUC of 0.77, compared to the nomogram's AUC of 0.73. Logistic Regression (LR) also demonstrated competitive performance with an accuracy of 65 % and an AUC of 0.73. The RF model exhibited high sensitivity (75 %) and precision (73 %), effectively identifying critical predictors of NSLN metastasis, including the presence of ductal carcinoma in situ (DCIS) and tumor characteristics such as type and grade. Explainable AI techniques, particularly SHAP values, provided insights into feature importance, enhancing model interpretability.

CONCLUSION

Our study offers a comprehensive comparison between ML models and the MSKCC nomogram for predicting NSLN metastasis among Iranian breast cancer patients. These findings contribute valuable insights to the discourse on personalized treatment approaches, emphasizing the need for tailored prognostic tools across diverse populations. The implications of this research extend to clinical decision-making, potentially improving the accuracy and efficacy of breast cancer management within the Iranian healthcare landscape.