Introduction

Breast cancer (BC) remains the most commonly diagnosed malignancy and the leading cause of cancer-related mortality among women worldwide, accounting for approximately 2.3 million new cases and 685,000 deaths in 2020 [1]. Despite substantial advances in early detection and multimodal treatment that have improved survival in high-income countries, the burden of BC continues to rise disproportionately in low- and middle-income countries (LMICs), where healthcare resources are often limited [2, 3]. In sub-Saharan Africa, patients frequently present at advanced stages, and survival outcomes remain poor due to limited access to early detection and treatment services [4–6]. Tanzania exemplifies this disparity, with BC representing the second most prevalent cancer among women, contributing to over 3,000 new cases and 1,300 deaths in 2018, and projected to increase by more than 80% by 2030 [7]. BC survival in these settings is significantly lower compared with high-income countries, and a considerable proportion of patients experience recurrence within five years, with high-risk individuals relapsing within the first two years, often reflecting aggressive tumour biology and resistance to therapy [8]. A major clinical challenge in this context is postoperative BC recurrence (BCR), defined as the reappearance of disease following a period of remission after surgical treatment [7].

Numerous machine learning (ML) models have been developed to predict BCR using data from cohorts in North America, Europe, and East Asia [9, 10]. For instance, Zeng et al. [11] (2023) utilized a long short-term memory (LSTM) network within a Chinese population and achieved promising discrimination. Similarly, Wu et al. [12] (2017) developed and validated personalized prognostic models for BC recurrence and overall survival using multidimensional data from patients treated at MD Anderson Cancer Center in the United States. They also developed race-specific and receptor subtype-specific models, highlighting the potential value of incorporating population characteristics into individualized prognostic assessment. However, models trained predominantly on Caucasian or Asian cohorts may perform sub-optimally in African patients, who often present with more aggressive subtypes, younger age at diagnosis, and distinct genetic backgrounds [13, 14]. These differences underscore the importance of developing and validating population-specific predictive tools.

To address this gap, Iddi et al. [15] (2025) introduced the inaugural ML-based risk prediction model for postoperative BCR specifically tailored to the Tanzanian population. Utilizing retrospective data from 199 patients treated at the Mnazi Mmoja Hospital and Ocean Road Cancer Institute, the authors conducted a comparative analysis of six ML algorithms: logistic regression (LR), extreme gradient boosting (XGBoost), random forest (RF), multilayer perceptron (MLP), light gradient boosting machine (LightGBM, or LGBM), and categorical boosting (CatBoost). Through univariate screening and least absolute shrinkage and selection operator (LASSO) regression, they identified five predictors of two-year BCR: lymph node metastasis, tumour grade, number of involved lymph nodes, tumour size, and marital status. The MLP demonstrated the highest test-set performance, achieving an area under the receiver operating characteristic curve (AUROC or ROC-AUC) of 0.935 [95% confidence interval (CI): 0.878–0.992], an accuracy of 0.850, a sensitivity of 0.875, and a specificity of 0.800. Subsequently, the model was implemented as a web-based risk calculator for clinical application in Tanzania.

Despite this advance, the exclusive use of LASSO regression for feature selection warrants further investigation. LASSO may overlook complex nonlinear relationships and interactions between variables [16]. The Boruta algorithm provides an alternative approach for feature selection in this context [17]. We evaluate whether Boruta-selected features, when integrated with various classifiers, achieve robust predictive performance in this Tanzanian cohort. Furthermore, we aim to assess whether increased model complexity offers a meaningful predictive advantage over an LR-based nomogram in the context of small sample sizes and limited resources. This comparison has direct implications for clinical implementation decisions in environments where interpretable and accessible tools are essential.

Materials and methods

Study design and data source

This retrospective methodological comparison study employed the dataset initially described by Iddi et al. [15] for BC patients who underwent surgical procedures at Mnazi Mmoja Hospital (Zanzibar) and Ocean Road Cancer Institute (Dar es Salaam, Tanzania) from 2019 to 2023. In accordance with the data availability statement in the original publication, a fully anonymized dataset was procured from the corresponding authors. All patient data, encompassing demographics, clinical characteristics, pathological reports, and two-year follow-up recurrence status, were preserved in their original form to ensure direct comparability of results (Figure 1). This study adhered to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) guidelines [18].

Display full size

Figure 1. Study flowchart.

Results

Feature selection using Boruta

Feature selection was performed using the Boruta algorithm as described in Feature selection.

In this study, Boruta was applied to the full set of candidate predictors (approximately 50 variables), as illustrated in Figure 2. To prevent information leakage and ensure methodological rigor, the feature selection procedure was conducted exclusively within the training dataset. The test dataset remained completely unseen during both feature selection and subsequent model development.

Display full size

Figure 2. Boruta feature selection plot.

Of the approximately 50 candidate predictors assessed, Boruta retained six that consistently outperformed the best shadow feature across all 100 iterations: tumour grade, lymph node metastasis, tumour size at diagnosis, tumour size within two years, number of lymph nodes involved, and receipt of hormone therapy before menopause. The remaining candidates were rejected as unimportant. These six predictors represent well-established prognostic factors in BC [15] and were carried forward for model development.

Model performance

In the internal hold-out test dataset, all six ML models demonstrated acceptable predictive performance (Table 2). A classifier that predicted recurrence for all patients (i.e., a majority-class baseline) would achieve 66.7% accuracy on the internal hold-out test set (40 of 60 patients), providing a contextual benchmark for the reported model performances.

Table 2. Performance comparison of ML models on the internal hold-out test set (n = 60; 40 events, 20 non-events).

Model	Accuracy	Bal. Acc.	F1-score	Kappa	MCC	NPV (95% CI)	PPV (95% CI)	PR-AUC (95% CI)	ROC-AUC (95% CI)	Sensitivity	Specificity (95% CI)
LR	0.87	0.85	0.90	0.70	0.70	0.80 (0.61–0.92)	0.90 (0.86–0.97)	0.94 (0.89–0.98)	0.91 (0.84–0.96)	0.90	0.80 (0.62–0.94)
LGBM	0.87	0.86	0.90	0.71	0.71	0.77	0.92	0.94	0.89 (0.80–0.95)	0.875	0.85
RF	0.83	0.83	0.87	0.63	0.64	0.73	0.89	0.92	0.88 (0.77–0.94)	0.85	0.80
SVM	0.77	0.81	0.79	0.54	0.59	0.59	0.96	0.96	0.90 (0.82–0.97)	0.675	0.95
MLP	0.82	0.81	0.86	0.60	0.61	0.70	0.89	0.97 (0.90–0.99)	0.93 (0.86–0.97)	0.825	0.80
NB	0.80	0.79	0.85	0.56	0.56	0.68	0.87	0.93	0.88 (0.78–0.95)	0.825	0.75

Display full size

Bal. Acc.: balanced accuracy; CI: confidence interval (bootstrap, 1,000 resamples); LR: logistic regression; LGBM: light gradient boosting machine; MCC: Matthews correlation coefficient; MLP: multilayer perceptron; NB: Naïve Bayes; NPV: negative predictive value; PPV: positive predictive value; RF: random forest; SVM: support vector machine. CIs are those obtained in the primary analysis. Where point estimates were recomputed, the original bootstrap intervals are retained.

Discrimination was comparable across the six models. LR and LGBM showed the most balanced profiles: LR achieved an accuracy of 0.87, F1-score of 0.90, sensitivity of 0.90, and specificity of 0.80, while LGBM achieved an accuracy of 0.87, F1-score of 0.90, sensitivity of 0.875, and specificity of 0.85. SVM showed the highest specificity (0.95) but the lowest sensitivity (0.675), indicating a trade-off toward negative-class predictions. NB showed comparatively lower performance across most metrics. MLP achieved the highest discrimination (ROC-AUC 0.93, PR-AUC 0.97). Because LR additionally demonstrated the strongest calibration (Figure 3A) and the greatest net benefit in DCA (Figure 3B), and because it is the most parsimonious and interpretable of the six models, LR was selected as the most suitable model for clinical application (Figure 4A and 4B).

Display full size

Figure 3. Calibration and DCA of ML models. (A) Calibration plots for the ML models on the internal hold-out test dataset. The calibration curves illustrate agreement between predicted probabilities and observed outcomes. (B) DCA for the ML models on the internal hold-out test dataset. DCA evaluates the net clinical benefit across a range of threshold probabilities. DCA: decision curve analysis.

Display full size

Figure 4. Curves comparing the performance of ML models on the internal hold-out test dataset. (A) Receiver operating characteristic (ROC) curves. (B) Precision-recall (PR). AUROC: area under the receiver operating characteristic curve; AUPRC: area under the precision-recall curve; LR: logistic regression; RF: random forest; MLP: multilayer perceptron; LightGBM: light gradient boosting machine; SVM: support vector machine; NB: naïve Bayes.

Furthermore, the calibration analysis demonstrated that the LR model achieved good agreement between predicted probabilities and observed outcomes, with the calibration curve closely approximating the ideal diagonal line across most risk thresholds (Figure 3A). Only minor deviations were observed at higher probability ranges. DCA further supported the clinical utility of the LR model (Figure 3B). LR consistently provided a higher net benefit compared with the “treat all” and “treat none” strategies across a broad range of threshold probabilities, particularly within the clinically relevant interval of approximately 0.2 to 0.8.

Model development

A nomogram was developed based on the LR model, incorporating six key predictive variables identified through Boruta feature selection: tumour grade, tumour size at diagnosis, lymph node metastasis, number of lymph nodes, tumour size within 2 years, and hormone therapy before menopause. The nomogram provides a quantitative and intuitive tool for estimating individual risk by assigning weighted scores to each predictor according to its relative contribution to the model. The nomogram represents the same model parameters as the LR model but in a graphical format suitable for clinical use (Figure 5).

Display full size

Figure 5. Nomogram and dynamic risk prediction tool based on the LR model. (A) Static nomogram. (B) Online web-based dynamic nomogram predicting clinical outcomes. The model provides graphical, numerical, and model summaries, with predictions visualized as 95% confidence intervals across a probability range (0.00 to 1.00).

To further enhance clinical applicability, a dynamic, web-based risk calculator was developed based on the final LR model (https://baozexiang.shinyapps.io/bca_dynnomapp/). This interactive tool allows users to input patient-specific variables and obtain real-time risk predictions, potentially assisting clinicians in identifying high-risk patients and informing tailored management strategies.

SHAP analysis

To elucidate the contribution of individual variables to the LR model, SHAP analysis was performed (Figure 6). The summary plot presents each feature ranked by importance, with the horizontal axis representing the corresponding SHAP values reflecting the direction and magnitude of each feature’s contribution. SHAP analysis identified tumour grade, tumour size at diagnosis, and lymph node metastasis as the most influential predictors, with higher values consistently associated with increased predicted risk. Tumour size within 2 years showed a less consistent and more heterogeneous effect, which may reflect measurement variability or its relationship with baseline tumour characteristics. The number of lymph nodes exhibited a modest positive contribution, whereas hormone therapy before menopause was associated with reduced predicted risk.

Display full size

Figure 6. SHAP-based feature importance and interpretation. (A) SHAP feature importance based on mean absolute SHAP values. (B) SHAP summary plot illustrating the distribution and direction of feature effects.

Discussion

In this study, we conducted a comprehensive comparative analysis of ML classifiers for predicting two-year postoperative BC recurrence in a Tanzanian cohort. Our findings demonstrate that while multiple ML models achieved strong discriminative performance, LR consistently provided the most balanced performance across discrimination, calibration, and clinical utility metrics. Notably, Boruta-based feature selection confirmed a clinically coherent subset of well-established prognostic predictors, reinforcing the importance of tumour burden (tumour size, nodal status) and tumour biology (grade) in recurrence risk stratification.

The balanced and stable performance of LR in our study is particularly noteworthy. Although MLP achieved slightly higher discrimination (ROC-AUC = 0.93 vs. 0.91), LR demonstrated superior calibration and decision curve performance, suggesting greater reliability for probability estimation. This finding aligns with prior evidence indicating that, in small to moderate-sized clinical datasets, simpler models often match or outperform complex algorithms due to reduced risk of overfitting and enhanced interpretability [21–23].

From a methodological standpoint, the comparison between LASSO (Iddi et al. [15]) and Boruta feature selection offers contextual insights into predictor identification [24]. It is important to note that this comparison is contextual rather than a controlled head-to-head analysis, as LASSO was not re-implemented within our analytical framework. The LASSO-driven framework yielded a parsimonious set of five variables, while Boruta identified a slightly broader, clinically coherent set including tumour size within two years and premenopausal hormone therapy alongside the core pathological features. Importantly, both approaches converged on the key determinants: lymph node metastasis, tumour grade, and tumour size, underscoring their central biological role in driving recurrence. Future work should systematically compare feature selection methods within a unified analytical pipeline.

Despite these methodological differences, overall predictive patterns remained highly consistent. The broader feature inclusivity achieved by Boruta did not translate into a substantial improvement in discrimination compared with the LASSO-based benchmark (AUROC 0.935 for Iddi et al. [15]’s MLP vs. 0.91 for our LR model); rather, it shifted the balance toward modest gains in interpretability when combined with LR. We also conducted a sensitivity analysis excluding the “tumour size within two years” variable to address potential concerns about temporal leakage. The AUROC decreased by only 0.01–0.02 across all models, with relative model rankings and conclusions regarding LR’s balanced performance remaining unchanged, indicating that our main findings are robust to its exclusion. Nevertheless, precise temporal definitions of predictors are important in clinical prediction modeling, and prospective validation should confirm the clinical utility of this predictor. The SHAP analysis revealed that this variable had a less consistent and more heterogeneous effect on predictions compared to other core predictors, consistent with this interpretation.

The SHAP analysis also revealed that pre-menopausal hormone therapy was associated with reduced predicted risk. This finding should be interpreted cautiously, as it may reflect confounding by indication: patients receiving hormone therapy may have had more favorable tumour characteristics (e.g., hormone receptor positivity) associated with lower recurrence risk. Alternatively, it may reflect the protective effect of endocrine therapy in appropriate candidates. Prospective studies with detailed treatment indications are needed to disentangle these effects.

Our findings have direct implications for clinical practice in LMICs. In resource-constrained settings such as Tanzania, predictive models must balance accuracy with interpretability and ease of implementation. The LR-based nomogram developed in this study provides a transparent and accessible tool for individualized risk prediction, aligning with global recommendations emphasizing context-appropriate models in LMIC healthcare systems [25].

Several strengths of this study should be emphasized. First, this is among the few studies to systematically evaluate classifier performance within an African BC cohort. Second, the use of an internal hold-out test set provides unbiased performance estimates. Third, the incorporation of multiple evaluation metrics, including calibration and DCA, ensures comprehensive assessment beyond discrimination alone.

However, several limitations must be acknowledged. The relatively small sample size may limit generalizability and increase the risk of model instability, particularly for more complex algorithms. The retrospective design introduces potential selection bias and limits causal inference. The effective EPV in this study warrants consideration: with 34 non-events in the training set and 9 model parameters after one-hot encoding, the non-EPV ratio was 3.8, well below the conventional minimum of 10 recommended for LR. While L2 regularization was applied to mitigate overfitting, this low EPV for the minority class reinforces the need for caution in interpreting model performance. Additionally, the internal hold-out test set (n = 60) contains only 20 non-events; a single misclassification in the non-recurrence class shifts specificity by 5 percentage points, making minority-class metrics [specificity, negative predictive value (NPV)] inherently unstable. While the LR model’s accuracy of 0.87 exceeds the 66.7% accuracy of a naïve ‘predict recurrence for all’ rule on the same test set, claims regarding robustness and generalizability must be tempered. External validation in independent cohorts, particularly across different African populations and healthcare settings, is essential before widespread clinical deployment.

Future research should focus on prospective validation and external validation across diverse African populations, as well as integration of additional data modalities such as genomic or imaging features. Implementation studies are needed to evaluate the real-world clinical impact of the dynamic nomogram in guiding treatment decisions.

This study suggests that Boruta-based feature selection, when integrated with LR, may provide clinically useful prediction of two-year postoperative BC recurrence in a Tanzanian cohort, pending external validation in larger, diverse cohorts. The Boruta algorithm confirmed well-established prognostic factors, particularly tumour grade, lymph node metastasis, and tumour size, as the dominant predictors in this population. Despite the availability of advanced ML algorithms, no meaningful performance advantage over LR was observed, underscoring the importance of model parsimony, interpretability, and robustness in small, resource-limited datasets.

The developed nomogram and dynamic web-based tool offer a practical and scalable approach for individualized risk stratification, with the potential to support clinical decision-making in LMIC settings. Future work should prioritize external validation across diverse African populations and prospective implementation studies to evaluate real-world clinical impact.