Iranian Journal of

Introduction: Breast cancer metastasis is a major cause of cancer-related death. Accurate prediction helps doctors make better decisions. This study developed and evaluated tree-based machine learning models to predict metastasis in Iranian women, using real clinical data with substantial missing values.
Methods: They looked at clinical records of 8,148 breast cancer patients in Tehran from 1997 to 2020. Variables with over 50% missing data were removed, leaving 4,310 complete cases. They compared Decision Tree, Random Forest, and XGBoost (which handles missing data well) against K-NN and Naïve Bayes (which need data imputation). They used stratified 10-fold cross-validation to check for overfitting and class imbalance, and then tested the best models on a separate hold-out set. Performance was measured using AUC, sensitivity, specificity, accuracy, and F1 score.
Results: Tree-based models worked better than the others. XGBoost had the best discrimination (AUC = 0.96, accuracy = 99.4%, F1 = 0.96), and Decision Trees were highly interpretable (sensitivity = 94%, specificity = 96.9%). Even though key predictors such as tumor size were excluded, other variables, such as hormone receptor status and age at menarche, allowed for strong predictions. K-NN had very low sensitivity (6%), and Naïve Bayes was inconsistent.
Conclusion: Decision trees and similar models can reliably predict breast cancer metastasis using incomplete, imbalanced real-world data, provided they are properly validated. These models are good for places with fewer resources. Future work should focus on better data collection and imputation methods. This study demonstrates the utility of interpretable machine learning for cancer applications in underrepresented populations.