DOI: 10.37349/edht.2026.1011104 ISSN: 2996-9409

Boruta-based feature selection and machine learning classifiers for predicting two-year postoperative breast cancer recurrence in a Tanzanian cohort: a comparative evaluation

Tella Marvel Gyeyock, Chernor Sulaiman Bah, Zexiang Bao, Asha Khatib Iddi, Sufficiency Elijah, Natasha Kangololo, Bianca Tankie Baakile, Sylvanus Favour Osahie, Jianjun Zou
Show PDF Cite
Aim: Breast cancer recurrence (BCR) remains a major determinant of poor outcomes in low- and middle-income countries (LMICs), where context-specific predictive tools are limited. We aimed to evaluate Boruta-based feature selection for identifying relevant predictors and compare multiple machine learning (ML) classifiers for predicting two-year postoperative BCR in a Tanzanian cohort. Methods: In this retrospective methodological study, data from 199 surgically treated breast cancer (BC) patients were analyzed. Boruta feature selection was applied to identify relevant predictors, followed by model development using six classifiers: logistic regression (LR), random forest, naïve Bayes, multilayer perceptron, support vector machine, and LightGBM. Models were trained using 10-fold cross-validation and evaluated on an internal hold-out test set (n = 60). Performance was assessed using AUROC, AUPRC, calibration, and decision curve analysis (DCA). Results: Boruta identified six key predictors, including tumour grade, lymph node involvement, and tumour size. Discrimination was comparable across models. LR demonstrated the strongest calibration and clinical utility, achieving an AUROC of 0.91 and an accuracy of 0.87. Although more complex models achieved comparable discrimination, they did not consistently outperform LR. DCA confirmed the superior clinical utility of the LR model across all relevant thresholds. Conclusions: In resource-constrained settings, a parsimonious LR model combined with Boruta-selected features provides robust, interpretable, and clinically applicable prediction of BCR, pending external validation in larger, diverse cohorts. These findings highlight that increased model complexity does not necessarily translate into improved performance in small clinical datasets and support the adoption of transparent, context-appropriate tools in LMICs.

More from our Archive