Unified interpretable machine learning for multiclass disease prediction from clinical and eco-medical data

Interpretable feature selection is important for developing reliable machine learning models for biomedical prediction, where predictive performance must be considered alongside transparency and computational efficiency. This study presents a systematic comparison of SHapley Additive exPlanations (SHAP) and Mutual Information (MI) for feature selection in two multiclass biomedical classification tasks: an Eco-Medical dataset involving environmental and clinical attributes and a Liver Health dataset comprising biochemical and demographic variables. Following stratified data partitioning, polynomial interaction features were generated using the training data, and the top 60 features identified by each selection method were evaluated with Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), and Gradient Boosting (GB). Model performance was assessed using accuracy, macro-averaged precision, recall and F1-score, balanced accuracy, Cohen’s kappa, Matthews correlation coefficient, and ROC-AUC, with sensitivity analysis used to examine alternative feature-selection sizes. The results demonstrate a clear dataset-dependent effect. On the Eco-Medical dataset, SHAP-based selection consistently outperformed MI across all four classifiers, improving test accuracy by 19.5–21.6% points and macro-F1 by 23.7–24.6% points. These improvements were statistically significant for all classifiers using exact paired McNemar tests with Holm correction. In contrast, on the Liver Health dataset, MI was competitive and achieved higher test accuracy and macro-F1 for LR, RF, and GB, while SHAP achieved slightly higher performance for DT; none of the paired SHAP–MI differences was statistically significant. MI also required less feature-selection time on both datasets, with total computational times of 1.679 s versus 3.100 s for Liver Health and 24.057 s versus 30.290 s for Eco-Medical. These findings indicate that neither feature-selection strategy is universally superior. SHAP provides model-specific attribution and valuable interpretability, but its additional computational cost and model dependence should be considered. MI offers a computationally lighter alternative that can remain highly competitive. These findings support empirical, dataset- and model-aware selection of feature-selection strategies for transparent biomedical decision-support systems.