International Journal of Management and Leadership Studies
2026; 6(ii): 1959-1979
ISSN:
2311 7575
IMPLEMENTING AN EXPLAINABLE STACKING ENSEMBLE FOR MULTI-CLASS MALARIA AND SICKLE CELL ANEMIA CLASSIFICATION IN WESTERN KENYA
Spencer Waliaula Mabonga, Laura Cheptegei, Andrew Kipkebut
Published:
01 November, 2026
Volume:
6
Issue:
ii
Keywords:
malaria, sickle cell anemia, stacking ensemble, explainable AI, SHAP, Western Kenya
Malaria and sickle cell anemia (SCA) are highly co-endemic in Western Kenya, and their overlapping clinical presentations complicate differential diagnosis, particularly for co-infected patients. Existing machine learning models are predominantly single- disease or binary and lack clinician-evaluated interpretability. The objective is to develop and evaluate an explainable heterogeneous stacking ensemble for four-class classification of malaria, SCA, co-infection, and negative cases using structured clinical data from Western Kenya. A retrospective cohort of 379 patient records from five facilities (January 2020–February 2026) was analyzed. The ensemble comprised Random Forest, Support Vector Machine, and XGBoost base learners with a logistic regression meta-learner. Missing data were handled using Multiple Imputations by Chained Equations, and class imbalance was addressed with SMOTE applied only within training folds. Performance was estimated using nested cross-validation. SHAP analysis provided global and local explanations, which were evaluated by 12 clinicians. The stacking ensemble achieved a nested cross-validation macro F1-score of 0.948 (SD = 0.012; 95% CI: 0.924–0.972), significantly outperforming individual base learners. On the hold-out test set (N = 58), accuracy was 98.3%, with perfect recall for co-infection. SHAP identified HbS fraction, hemoglobin level, and malaria RDT as the dominant predictors, accounting for approximately 85% of explanatory power. Overall, 83.3% of clinicians rated explanations as clear or moderately clear and expressed confidence in the model’s recommendations. Four-class malaria–SCA classification is feasible using structured clinical data alone, and heterogeneous stacking provides a statistically significant performance advantage. However, findings are formative given the retrospective design and small hold-out set; prospective validation is required before clinical deployment.