Rethinking hyperparameter optimization for efficient and explainable machine learning in civic health...

Malaria remains one of the most persistent infectious diseases worldwide, with hundreds of millions of cases annually. Machine learning has emerged as a promising tool for supporting clinical decision-making in malaria care, but the practical deployment of such models is constrained by hyperparameter optimization complexity. This study provides a comprehensive comparison of stochastic and deterministic hyperparameter optimization strategies for supervised machine learning models applied to the binary classification of malaria severity (severe vs. not-severe disease) among symptomatic patients.

We evaluated five optimization strategies (Grid Search, Random Search, Bayesian Optimization, Genetic Algorithms, and Hyperband Racing) across six supervised learning algorithms (Random Forest, Neural Networks, Support Vector Machines, Logistic Regression, XGBoost, and K-Nearest Neighbors). The classification task was binary severity prediction (Severe vs. Not-Severe) among symptomatic malaria patients. Models were trained on synthetic datasets with varying complexity and a real clinical malaria dataset comprising 337 patients with 16 clinical features (34.4% severe cases). Performance was assessed using 10-fold cross-validation repeated five times, with evaluation metrics including F1-score, accuracy, ROC AUC, Matthews Correlation Coefficient, and balanced accuracy. Class imbalance was addressed using standard SMOTE oversampling applied exclusively within training folds, and model interpretability was examined through LIME, SHAP, and Permutation Feature Importance analyses.

Stochastic and deterministic optimization strategies showed no statistically or clinically meaningful difference in severity classification performance, with mean CV F1-scores of 0.693 and 0.694 respectively (difference [Formula: see text]) (formally confirmed by equivalence testing with TOST, [Formula: see text], [Formula: see text]). Random Search demonstrated the highest efficiency (17.992 evaluations/minute), followed by Grid Search (16.734), while Genetic Algorithms exhibited the lowest efficiency (5.952), requiring over 41 minutes on average. Algorithm selection and dataset characteristics dominated performance variability, accounting for 5.5% and 90.6% of total variance respectively, while optimization strategy contributed approximately 0% (variance component collapsed to zero under formal mixed-effects decomposition, consistent with negligible influence). After class balancing and hyperparameter tuning, Random Forest achieved the highest performance on the severity classification task (F1-score: 0.831, ROC AUC: 0.870). Explainability analyses consistently identified age, coca-cola urine, prostration, hyperpyrexia, and convulsions as key predictive features across multiple methods.

Simple hyperparameter optimization strategies may be sufficient for model development in malaria severity classification applications under the tested conditions, achieving performance comparable to more complex methods with substantially lower computational requirements. Clinical utility, however, remains to be confirmed through prospective validation. The findings support a pragmatic modeling approach prioritizing data quality, algorithm selection, and computational efficiency over optimization complexity, particularly relevant for resource-constrained healthcare settings where malaria burden is highest.