How Can Robot Bees and Shark Algorithms Aid Dengue Diagnosis?

Dengue is a mosquito‑borne viral fever that affects hundreds of millions each year. Detecting it early is key to quick treatment and lower mortality. Researchers tested three nature‑inspired algorithms – Golden Jackal Optimization, Fox Optimizer and Sea Lion Optimization – to pick the most useful data and fine‑tune an AI model called XGBoost. The outcome? Almost perfect accuracy (99.89%) using just two blood parameters: white‑blood‑cell count (WBC) and platelet count.\n\nThe trick lies in data reduction: fewer variables mean faster, easier‑to‑interpret models that can be deployed in clinics. SHAP, an explanation tool, highlighted WBC count as the top driver: low WBC points to dengue, high WBC suggests a negative case. DiCE further shows how minimal tweaks to the input can turn a wrong prediction into a right one, offering clues for better clinical tests.\n\nThese findings pave the way for rapid, low‑cost, transparent diagnostic tools ready for the regions most hit by dengue.

The proposed approach merges animal‑inspired metaheuristics (GJO, FO, SLO) with the power of XGBoost, a gradient‑boosting decision‑tree algorithm. The metaheuristics operate on two levels: feature selection and hyper‑parameter tuning (number of trees, max depth, learning rate). Reduction to two features is feasible due to the strong correlation between WBC, platelets and infection status: dengue causes thrombocytopenia (low platelets) and leukopenia (low WBC).\n\nSHAP (SHapley Additive exPlanations) computes each feature's Shapley value, providing local and global contributions to the prediction. In our case, the average SHAP value for WBC is negative for positive cases, confirming its importance. DiCE (Diverse Counterfactual Explanations) generates minimal counterfactuals showing how slight changes in WBC or platelets can flip the predicted class, useful for interpreting errors and suggesting clinical thresholds.\n\nTen‑fold cross‑validation ensures robustness: each fold holds about 10% of samples, reducing overfitting risk. The F‑score of 99.92% indicates excellent balance between precision and recall, the latter being perfect (100%). The combination of feature reduction, SHAP interpretability, and DiCE correction makes the model not only accurate but also transparent and potentially integrable into clinical decision‑support systems, especially in resource‑limited settings.