2.1. Ethics statement

This study utilized a de-identified, publicly available dataset and did not involve direct interaction with human subjects. Institutional review board (IRB) approval was not required per institutional policy for analyses of publicly accessible, anonymized data. The WBCD contains no personally identifiable information, demographic variables, or protected health information. All specimens were de-identified at source prior to public release. The study was conducted in accordance with the principles of responsible AI research, including transparency of methods, reproducibility of results, and open dissemination of analytical tools.

2.2. Study design and analytical framework

This study employed a retrospective, cross-sectional analytical design utilizing a publicly available benchmark dataset. The analytical pipeline comprised four sequential phases: (i) data acquisition and preprocessing, (ii) multi-algorithmic model development with hyperparameter optimization, (iii) comprehensive performance evaluation using discriminative and probabilistic metrics, and (iv) post-hoc explainability analysis using SHAP. The study adheres to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines and follows TRIPOD-ML extensions for machine learning-based prediction studies. A completed TRIPOD checklist is provided as supplementary material. All analyses were conducted in Python 3.10, leveraging scikit-learn (v1.3), XGBoost (v2.0), LightGBM (v4.1), and SHAP (v0.43) libraries. The complete analytical codebase and all supplementary materials are publicly available via a dedicated GitHub repository (see Data Availability Statement) to ensure full computational reproducibility in accordance with FAIR (Findable, Accessible, Interoperable, Reusable) data principles.

2.3. Data source and characteristics

We utilized the Wisconsin Breast Cancer Diagnostic Dataset (WBCD), originally curated by Dr. William H. Wolberg at the University of Wisconsin Hospitals and publicly accessible through the UCI Machine Learning Repository and scikit-learn library [27]. The WBCD comprises 569 instances representing digitized fine-needle aspirate (FNA) specimens from breast masses, with 212 malignant (37.3%) and 357 benign (62.7%) cases. Each specimen is characterized by 30 real-valued nuclear morphometric features computed from digitized cell nucleus images. These features capture ten distinct morphological properties of cell nuclei, each reported as mean, standard error (SE), and “worst” (mean of the three largest nuclei) values: radius, texture, perimeter, area, smoothness, compactness, concavity, concave points, symmetry, and fractal dimension [27].

The selection of this dataset was deliberate and methodologically motivated. The WBCD derives from FNAC, the cytological technique most commonly employed in LMICs for breast mass evaluation. Its morphometric features correspond to quantifiable nuclear properties assessable through digital microscopy, which is increasingly available even in peripheral health facilities. The dataset has been extensively validated in the ML literature, enabling meaningful benchmarking, while its manageable dimensionality (30 features) supports interpretable modeling without requiring the computational infrastructure demanded by high-dimensional omics data or deep learning on raw imaging data.

2.4. Data preprocessing and partitioning

The dataset contained no missing values. All 30 continuous features were standardized using z-score normalization (zero mean, unit variance) via the StandardScaler transformation, fitted exclusively on training data to prevent information leakage. The dataset was partitioned into training (80%, n = 455) and hold-out test (20%, n = 114) subsets using stratified random sampling (random seed = 42) to preserve the original class distribution in both partitions. The 80:20 ratio was selected based on statistical considerations appropriate to this dataset size; the resulting training set provides sufficient observations to support stable parameter estimation across all eight classifiers and 30 features, while the hold-out set (42 malignant, 72 benign) provides adequate statistical power for reliable performance estimation. This ratio complements the primary internal validation strategy of stratified 10-fold cross-validation applied to the training partition.

Prior to any model development, we verified the comparability of the two partitions through two checks: (i) class distribution confirmation, the malignant-to-benign ratio was preserved at approximately 37:63 in both the training (37.1% malignant) and test (36.8% malignant) subsets through stratified sampling; and (ii) feature distribution equivalence, Mann–Whitney U testing across all 30 features between the training and hold-out subsets revealed no statistically significant distributional differences (all p > 0.05 after Bonferroni correction), confirming that the hold-out partition constitutes a representative and unbiased sample of the source data prior to classifier development.

2.5. Classification algorithms

Eight supervised learning algorithms representing distinct inductive biases and learning paradigms were selected for systematic comparison (S1 Table provides the complete hyperparameter specifications for all models):

Logistic Regression (LR): A maximum-likelihood linear classifier with L2 regularization (C = 1.0) and LBFGS optimizer, serving as a transparent baseline with inherent interpretability.

Random Forest (RF): A bagging-based ensemble of 300 decision trees with maximum depth of 10, minimum samples per split of 5, and minimum samples per leaf of 2, providing implicit feature importance through impurity-based measures.

Extreme Gradient Boosting (XGBoost): A regularized gradient boosting framework with 200 estimators, maximum depth of 5, learning rate of 0.05, and column and row subsampling at 80%, incorporating L1 and L2 regularization.

Light Gradient Boosting Machine (LightGBM): A histogram-based gradient boosting algorithm with 200 estimators, maximum depth of 5, learning rate of 0.05, and 31 leaves, optimized for computational efficiency.

Support Vector Machine (SVM): A kernel-based classifier employing the Radial Basis Function (RBF) kernel with regularization parameter C = 10 and automatic gamma scaling, with Platt scaling for probability estimation.

Gradient Boosting Classifier (GBC): A sequential ensemble of 200 shallow decision trees (maximum depth = 4) with learning rate of 0.05 and stochastic subsampling at 80%.

Decision Tree (DT): A single CART classifier with maximum depth of 5 and minimum samples per split of 5, serving as an interpretable but potentially high-variance baseline.

K-Nearest Neighbors (KNN): A distance-based classifier with 7 neighbors, distance-weighted voting, and Minkowski metric, representing a non-parametric instance-based approach.

Hyperparameter Configuration and Optimization: The hyperparameter configurations were determined through a two-stage approach. In stage one, initial parameter ranges were informed by published algorithm-specific benchmarks for structured tabular data and developer recommendations [28,29]. In stage two, these ranges were refined through systematic grid search conducted exclusively on the training partition (n = 455) using 5-fold stratified cross-validation with AUC-ROC as the primary optimization criterion. For boosting algorithms, the search space encompassed: n_estimators ∈ {100, 200, 300, 500}, max_depth ∈ {3, 5, 7, 10}, and learning_rate ∈ {0.01, 0.05, 0.10, 0.20}. The hold-out test set was strictly withheld from all hyperparameter selection procedures to prevent information leakage. Final configurations represent those maximizing cross-validated AUC while maintaining training-to-validation performance stability, as evidenced by the learning curve analysis (S1 Fig).

2.6. Model evaluation strategy

Model performance was assessed through a dual evaluation protocol. Internal validation employed stratified 10-fold cross-validation (CV) on the training partition, reporting mean and standard deviation for accuracy, F1-score, and AUC-ROC. Hold-out test set evaluation utilized the pre-specified 20% partition, with the following metrics computed: accuracy, precision (positive predictive value), recall (sensitivity), F1-score, area under the receiver operating characteristic curve (AUC-ROC), Matthews Correlation Coefficient (MCC), Brier score, log loss, and average precision (area under the precision-recall curve). The selection of multiple complementary metrics was intentional. In cancer diagnostics, sensitivity (recall) is prioritized to minimize false negatives (missed malignancies), while precision and specificity govern false-positive rates that drive unnecessary procedures. MCC provides a balanced assessment robust to class imbalance, while Brier score and calibration analysis evaluate the probabilistic reliability essential for clinical risk communication [30].

We explicitly note that the hold-out test set, while derived through a pre-specified, strictly separated partitioning procedure, shares the same source dataset (WBCD) as the training partition and therefore constitutes internal hold-out validation, not true external validation as defined by TRIPOD-ML criteria. True external validation, requiring evaluation on an independent cytological dataset from a separate cohort, institution, or geographic context, was not feasible within the scope of this study and is designated as a priority for future work (Section 4.6).

2.7. SHapley Additive exPlanations (SHAP) explainability analysis

Post-hoc model interpretability was operationalized using SHAP (SHapley Additive exPlanations), a unified framework grounded in Shapley values from cooperative game theory [16]. For tree-based models (XGBoost, Random Forest, LightGBM), the optimized TreeExplainer algorithm was employed, which computes exact Shapley values in polynomial time by exploiting the recursive partitioning structure of decision trees.

The SHAP analysis comprised four components: (i) Global importance ranking, computed as the mean absolute SHAP value across all test instances for each feature, quantifying the average magnitude of each feature’s contribution to model output. (ii) Cross-model consensus analysis, comparing SHAP-derived feature rankings across XGBoost, Random Forest, and LightGBM to identify algorithmically robust predictive features versus model-dependent predictors. (iii) SHAP dependence analysis, examining the functional relationship between individual feature values and their SHAP contributions for the six most influential features, with quadratic trend estimation to characterize nonlinear effects. (iv) Probability calibration assessment using calibration curves (reliability diagrams) to evaluate the correspondence between predicted probabilities and observed outcome frequencies.

2.8. Clinical contextualization for resource-limited settings

To bridge the computational-clinical translation gap, model selection criteria incorporated considerations beyond predictive accuracy: (i) computational parsimony, favoring algorithms deployable on standard computing hardware without GPU requirements; (ii) interpretability depth, prioritizing models amenable to feature-level explanation; (iii) calibration quality, essential for communicating diagnostic certainty in settings with limited specialist oversight; and (iv) biological plausibility, evaluating whether SHAP-identified features align with established cytopathological diagnostic criteria.

2.9. Statistical analysis

Differences in feature distributions between malignant and benign classes were assessed using the Mann–Whitney U test, with significance defined as p < 0.05 after Bonferroni correction. Cross-model feature ranking concordance was evaluated using Spearman rank correlation. All confidence intervals are reported at the 95% level.

3.1. Dataset characteristics and feature distribution

All 15 top SHAP-ranked features demonstrated highly significant distributional differences between diagnostic classes (Mann–Whitney U test, all p < 0.001, Bonferroni correction; Table 1), supporting their discriminative potential. Notably, “worst” (extreme-value) features exhibited the largest between-group effect sizes; worst perimeter showed the most pronounced malignant-benign separation (p < 2.58 × 10⁻⁸⁰), consistent with the cytopathological recognition that malignant nuclei exhibit marked size heterogeneity [31]. Detailed descriptive statistics are provided in Table 1.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive statistics of the top 15 SHAP-ranked features by diagnostic class.

https://doi.org/10.1371/journal.pdig.0001706.t001

3.2. Comparative model performance

On the hold-out test set (Table 2), AUC values ranged from 0.916 (Decision Tree) to 0.996 (XGBoost, SVM), with all ensemble methods exceeding 0.988 (Fig 1). Logistic Regression achieved the highest accuracy (0.983), precision (0.986), and MCC (0.962), while KNN attained perfect recall (1.000), the latter at the cost of three false positives, a trade-off requiring careful clinical consideration. The performance differential between the six top-performing models was narrow (accuracy: 0.956-0.983; AUC: 0.988-0.996), indicating that model selection should be guided primarily by interpretability and calibration rather than marginal accuracy differences. In 10-fold cross-validation, performance rankings were consistent with hold-out results (Table 2).

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Comparative performance of eight machine learning classifiers on the hold-out test set and stratified 10-fold cross-validation.

https://doi.org/10.1371/journal.pdig.0001706.t002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Receiver Operating Characteristic (ROC) and Precision-Recall (PR) curves for all eight classification algorithms evaluated on the independent test set (n = 114).

(A) Representative ROC curves with area under the curve (AUC) values, and (B) precision-recall curves with average precision (AP) scores. The dashed diagonal line in Panel A represents random classifier performance (AUC = 0.5). All ensemble methods achieved AUC values exceeding 0.98, demonstrating near-ceiling discriminative performance.

https://doi.org/10.1371/journal.pdig.0001706.g001

3.3. SHAP-based explainability analysis

3.3.1. Global feature importance.

Worst perimeter (mean |SHAP| = 0.932), worst concave points (0.884), and worst area (0.778) were the three most important predictors in XGBoost (Fig 2A). Higher values of these features were associated with positive SHAP contributions (increased malignancy probability), consistent with cytopathological hallmarks of nuclear enlargement and irregular nuclear contours in malignant lesions [32].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. SHAP-based explainability analysis.

(A) SHAP beeswarm plot for XGBoost showing the distribution of SHAP values for the top 15 features across all test instances. Each point represents one patient; color indicates the feature value (red = high, blue = low); horizontal position indicates the magnitude and direction of the feature’s impact on model output. (B) Comparative mean absolute SHAP values across XGBoost, Random Forest, and LightGBM, demonstrating cross-model feature importance consensus.

https://doi.org/10.1371/journal.pdig.0001706.g002

3.3.2. Cross-model SHAP consensus.

The three top-ranked features (worst perimeter, worst concave points, worst area) maintained consistent top-3 positions across all three models (Table 3; Fig 2B), with all three consensus metrics (arithmetic mean, geometric mean, and median rank) placing them identically. In contrast, worst texture showed pronounced model dependence (ranks: 4, 12, 6 across XGBoost/RF/LightGBM), and mean smoothness showed the greatest discordance (ranks: 9, 20, 10). Pairwise Spearman correlations of SHAP ranking vectors were XGBoost vs. LightGBM ρ = 0.86, XGBoost vs. RF ρ = 0.82, and RF vs. LightGBM ρ = 0.67, confirming that the two gradient-boosted architectures share greater feature importance concordance than either shares with the bagging-based RF.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. SHAP-derived feature importance rankings across three tree-based algorithms.

https://doi.org/10.1371/journal.pdig.0001706.t003

3.4. Confusion matrix analysis

Logistic Regression achieved the most favorable error profile for clinical deployment, with only 1 false negative (missed malignancy) and 1 false positive across 114 test cases (Fig 3). Full sensitivity, specificity, PPV, and NPV values for all eight models are provided in S2 Table.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Confusion matrices for the four gradient boosting variants (A: XGBoost, B: Random Forest, C: LightGBM, D: Gradient Boosting) on the independent test set.

Each cell displays the raw count and percentage of the true class. Darker shading indicates higher proportions. All models demonstrated strong true positive and true negative rates, with false negative counts (missed malignancies) ranging from 1 to 2.

https://doi.org/10.1371/journal.pdig.0001706.g003

3.5. Model calibration and multi-metric performance comparison

Logistic Regression demonstrated the best calibration (Fig 4A), with predicted probabilities closely tracking observed frequencies. XGBoost showed modest mid-range overconfidence, and Random Forest exhibited characteristic sigmoidal miscalibration. The radar chart (Fig 4B) confirms that no single model dominates all six performance dimensions simultaneously, reinforcing context-specific selection guided by clinical priorities, Logistic Regression offering the most balanced overall profile.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Model calibration and multi-metric performance comparison.

(A) Probability calibration curves (reliability diagrams) for the five top-performing models. The dashed diagonal line represents perfect calibration; deviations indicate systematic over- or under-confidence in predicted probabilities. Logistic Regression demonstrated the best calibration fidelity. (B) Radar chart comparing six performance dimensions (Accuracy, Precision, Recall, F1-Score, AUC, normalized MCC) across the top five models.

https://doi.org/10.1371/journal.pdig.0001706.g004

3.6. SHAP dependence analysis

SHAP dependence plots (Fig 5) revealed that worst perimeter showed a strongly monotonic positive relationship consistent with a size threshold effect for malignancy, while worst concave points exhibited a steeper sigmoid-like profile reflecting its near-binary discriminative nature. Area error demonstrated a nonlinear relationship with heteroscedasticity at extreme values, potentially reflecting the cellular heterogeneity characteristic of malignant lesions.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. SHAP dependence plots for the six most influential predictive features identified by XGBoost.

Each panel displays the relationship between individual feature values (x-axis) and their SHAP contributions to model output (y-axis). Points are colored by feature value magnitude (red = high, blue = low). Dashed curves represent quadratic trend fits. The horizontal gray line at SHAP = 0 delineates the transition between features pushing predictions toward benign (below) versus malignant (above) classification.

https://doi.org/10.1371/journal.pdig.0001706.g005

4.1. Principal findings and methodological contributions

This study demonstrates that a multi-algorithmic ML framework, augmented with SHAP-based explainability, achieves near-ceiling discriminative performance for FNAC-based breast cancer classification while providing transparent, clinically interpretable feature attributions. The principal methodological contribution of this study, relative to prior SHAP-based analyses of breast cancer classifiers, are fourfold: (i) the introduction of cross-model SHAP consensus analysis, using arithmetic mean, geometric mean, and median rank alongside pairwise Spearman correlations, as a method for distinguishing algorithmically invariant biological signals from model-specific artifacts; (ii) integration of probability calibration analysis as a deployment-critical evaluation dimension alongside discriminative metrics; (iii) systematic mapping of SHAP-derived feature importance to established cytopathological grading criteria (Masood Classification); and (iv) contextualization of all findings within the specific operational constraints and clinical priorities of LMIC diagnostic settings, a combination not previously addressed in the breast cancer SHAP literature [33,34].

While the WBCD has been extensively utilized in ML benchmarking studies, the vast majority of prior work has focused exclusively on maximizing classification accuracy without addressing clinical interpretability [33,35]. Key studies, including those by Agarap (2018) and Salod and Singh (2019), evaluated multiple classifiers but did not incorporate any explainability analysis, did not assess probability calibration, and made no attempt to contextualize their findings for clinical deployment in resource-limited settings. Our study advances beyond this body of work on the four dimensions outlined above. Our finding that Logistic Regression achieved the highest test accuracy (0.983) and MCC (0.962) while providing inherent interpretability calls into question the common assumption in computational oncology that model complexity is necessary for high diagnostic accuracy [35].

4.2. SHAP-derived clinical signatures: Biological plausibility and diagnostic relevance

The SHAP analysis identified worst perimeter, worst concave points, and worst area as the three most important predictive features across all tree-based models. These findings are consistent with the cytomorphological criteria codified in the Masood Classification System and other standardized FNAC grading schemas [36,37]. The dominance of “worst” features (mean of three largest nuclei) over “mean” features is biologically interpretable. Malignant tumors are characteristically heterogeneous, with subpopulations of cells exhibiting extreme morphological deviations. The “worst” features capture this intra-tumoral heterogeneity, effectively functioning as computational surrogates for the cytopathological practice of focusing on the most abnormal cells in a specimen [38]. This concordance between SHAP-identified feature importance and established diagnostic heuristics supports the biological plausibility and clinical relevance of the ML framework.

4.3. Implications for resource-limited settings and translational equity

We operationalize ‘Translational Equity’ as a three-component construct for evaluating AI-assisted diagnostics in resource-limited settings. The first component is technical accessibility: the AI system must be deployable within the computational and infrastructural constraints of the target setting, including standard hardware, offline functionality, and low inference latency. The second component is interpretive accessibility: the system must provide outputs that clinicians without specialist AI training can critically evaluate and meaningfully integrate, a requirement that calibrated probability outputs and SHAP-based explanations directly address. The third component is contextual alignment; the features identified as predictive must correspond to biological phenomena clinically recognizable and measurable within the diagnostic workflows of the target setting. A system that achieves high discriminative accuracy but fails on any one of these three components does not achieve Translational Equity, even on benchmark datasets. This distinction is critical because the primary bottleneck to equitable AI in global oncology is not the availability of algorithms but the absence of frameworks that make those algorithms trustworthy, usable, and contextually appropriate for the clinicians who most need them.

The translation of ML diagnostics to LMIC contexts requires careful consideration of several operational constraints that are frequently overlooked in algorithm development. First, computational infrastructure: all models evaluated in this study, including the ensemble methods, run efficiently on standard laptop hardware without GPU requirements, with inference times under 100 milliseconds per specimen, making them deployable on the computing resources available in peripheral health facilities [39].

Second, specialist scarcity: in many LMIC settings, a single cytopathologist may serve populations exceeding 500,000 [40]. An interpretable ML system that provides not only a diagnostic classification but also a transparent explanation of which morphometric features drove the prediction can serve as a “virtual second opinion,” potentially reducing inter-observer variability and supporting less-experienced technicians in preliminary screening. The SHAP-based explanation framework provides precisely this capability, highlighting for the user which specific cellular features were most influential in each individual prediction.

Third, calibration and risk communication: our calibration analysis demonstrates that Logistic Regression provides the most reliable probability estimates, which is critical for triage decisions in settings where confirmatory testing may require referral to distant facilities. A calibrated probability of 92% malignancy conveys meaningfully different clinical urgency than 60%, enabling more rational resource allocation in capacity-constrained environments [41].

Fourth, algorithmic accountability: the deployment of ML diagnostics in underserved settings carries heightened ethical obligations, as patients may have limited access to appeal mechanisms if algorithmic errors occur. SHAP-based explanations provide an audit trail that enables post-hoc review of individual predictions, supporting the principles of algorithmic accountability and the right to explanation that are increasingly codified in AI governance frameworks [42].

Honest acknowledgment of implementation barriers is essential to avoid creating unrealistic expectations. In truly resource-limited settings, the proposed three-stage pathway faces significant structural challenges beyond technical feasibility. First, retrospective validation (Stage i) requires access to locally annotated FNAC datasets, a resource absent in most LMIC cytopathology laboratories where digital archiving is rarely standardized. Second, integration into laboratory information management systems (Stage ii) presupposes electronic health record infrastructure that many peripheral facilities lack. Third, digitization of FNAC specimens requires calibrated digital microscopes or flatbed scanners, equipment often inaccessible at sub-district facilities. Fourth, regulatory pathways for AI-assisted diagnostics in most LMICs remain undefined, creating approval uncertainty. These challenges reinforce rather than diminish the importance of the Translational Equity framework: future development must engage implementation scientists, health systems specialists, and LMIC clinician communities from the earliest stages.

4.4. Cross-model SHAP consensus as a robustness indicator

A methodological contribution of this study is the systematic comparison of SHAP-derived feature rankings across three tree-based architectures. The high concordance of the top three features across XGBoost, Random Forest, and LightGBM, confirmed across arithmetic mean, geometric mean, and median consensus metrics, provides evidence that these features represent algorithmically invariant signals rather than artifacts of a particular model’s inductive bias. Conversely, the model-dependent variability observed for features such as worst texture and mean smoothness suggests that these features may interact with algorithmic assumptions (e.g., feature splitting criteria, regularization strategies) in model-specific ways, warranting cautious interpretation [28].

Pairwise Spearman correlation analysis revealed that XGBoost and LightGBM (both gradient-boosted architectures) demonstrated higher mutual concordance (ρ = 0.86) than either shared with Random Forest (XGBoost vs. RF: ρ = 0.82; RF vs. LightGBM: ρ = 0.67), suggesting that algorithmic architecture, specifically the boosting versus bagging paradigm, partially determines feature ranking concordance for this dataset. The geometric mean and median consensus ranks are fully consistent with arithmetic mean rankings for the top 5 features, while the divergence observed for lower-ranked features (e.g., compactness error: arithmetic rank 18 vs. geometric rank 17.6) reflects the sensitivity of the arithmetic mean to extreme RF rank values (rank 28), confirming the value of the robust consensus metrics.

This consensus-based approach to SHAP analysis offers a practical framework for clinical validation: features that maintain high importance across multiple algorithmic paradigms are more likely to reflect genuine biological signals and are therefore more appropriate targets for clinical decision rules.

4.5. Utility, accessibility, and open science commitment

The complete codebase, encompassing data preprocessing, model training, SHAP analysis, visualization generation, and calibration assessment, is publicly deposited in a GitHub repository (https://github.com/drbamodu/explainable-ml-breast-cancer-lmic) under open-source MIT license and archived with a persistent DOI on Zenodo (https://doi.org/10.5281/zenodo.18928723). All dependencies are version-pinned, and the analysis can be reproduced on standard computing hardware (no GPU required) within approximately 15 minutes, lowering the barrier to replication by researchers in resource-constrained institutions.

From a practical deployment perspective, the framework is designed to be modular and extensible. The trained Logistic Regression model, identified as the optimal candidate for LMIC deployment due to its combination of near-ceiling accuracy, superior calibration, inherent interpretability, and minimal computational requirements, can be exported as a serialized object and integrated into existing digital pathology workflows via a lightweight application programming interface (API). We envision a three-phase implementation pathway: (i) retrospective validation on locally collected FNAC datasets to assess domain generalizability; (ii) integration into laboratory information management systems (LIMS) as a computer-aided detection (CADx) module; and (iii) prospective evaluation in a clinical effectiveness trial comparing ML-augmented versus standard cytopathological assessment. The SHAP explanation layer can be rendered as a visual overlay highlighting which morphometric features contributed most to each individual prediction, enabling cytotechnicians and trainee pathologists to engage critically with algorithmic outputs rather than treating them as opaque directives.

With respect to reporting standards and responsible AI, this study adheres to the TRIPOD-ML guidelines for transparent reporting of prediction model studies. We acknowledge the broader ethical imperative that ML systems deployed in healthcare settings, particularly those serving vulnerable populations, must be subject to continuous monitoring for performance drift, demographic bias, and calibration degradation over time. While the WBCD does not contain demographic variables that would enable formal fairness auditing, we emphasize that prospective deployments of this framework must incorporate systematic bias assessment as a prerequisite for ethical implementation.

4.6. Limitations and future directions

Several limitations warrant acknowledgment. First, the WBCD, while well-validated and FNAC-derived, originates from a single institution in the United States, and its generalizability to cytological specimens processed in LMIC laboratories with potentially different preparation protocols, staining techniques, and digitization equipment remains to be established through prospective multi-site validation studies [43]. Second, the binary classification framework (malignant vs. benign) does not capture the diagnostic complexity of indeterminate or atypical cytological findings, which represent a significant proportion of FNAC specimens and pose the greatest diagnostic challenge [44]. Third, while SHAP provides post-hoc interpretability, it does not establish causal relationships between features and outcomes; the identified associations, though biologically plausible, require prospective mechanistic validation.

Fourth, a key limitation of this study is the absence of true external validation. The hold-out test set utilized for performance evaluation was derived from the same WBCD dataset as the training partition and therefore does not fulfill TRIPOD-ML criteria for external validation, which requires evaluation on an independent dataset from a separate cohort, institution, or geographic context. The performance estimates reflect internal generalizability to this dataset’s distributional properties rather than generalizability to cytological specimens from LMIC laboratories. External validation on independent FNAC datasets from LMIC facilities represents the single most critical methodological priority before any clinical deployment can be responsibly considered. We are currently pursuing collaborative partnerships with cytopathology laboratories in Sub-Saharan Africa and South Asia for prospective collection and analysis.

Fifth, while hyperparameter optimization was conducted through cross-validated grid search on the training partition, the bounded search space and single optimization criterion (AUC-ROC) may not have identified globally optimal configurations for all performance dimensions. For hyperparameter-sensitive algorithms such as XGBoost and SVM, more exhaustive strategies (e.g., Bayesian optimization) may yield marginal gains, though the near-ceiling performance observed suggests these would be clinically inconsequential for this dataset. Sixth, sensitivity analyses across alternative split ratios (e.g., 70:30, 75:25) were not formally conducted; however, the primary 10-fold cross-validation substantially mitigates split-ratio dependency concerns.

Future research should prioritize external validation using cytological datasets from LMIC facilities, extension to multi-class classification incorporating indeterminate categories, integration with digital pathology workflows through application programming interfaces (APIs), and prospective clinical trials evaluating the impact of SHAP-augmented ML on diagnostic accuracy and turnaround time in resource-limited settings. Additionally, the development of model-embedded explanations that provide real-time interpretability during clinical use, rather than post-hoc analysis, represents a critical frontier for clinical XAI [22].