An Evidential Neural Network Framework with Gaussian Random Fuzzy Numbers for Multi-Disease Risk Prediction and Uncertainty Quantification

Abstract Nigeria faces an increasing dual burden of communicable and non-communicable diseases. Cardiovascular disorders, diabetes, malaria, and kidney disease together account for more than two-thirds of hospital admissions. Although interest in advanced predictive systems is growing, many current models overlook correlated clinical features and fail to quantify prediction uncertainty. This study presents the Evidential Neural Network with Gaussian Random Fuzzy Numbers (EVNN-GRFN), a hybrid framework designed to model correlation, fuzziness, and uncertainty in diverse medical data. The method converts clinical indicators into Gaussian fuzzy sets, applies penalties to redundant correlations, and integrates probabilistic evidence using Dempster–Shafer theory. Four model variants (M1–M4) were tested on benchmark and Nigerian datasets—UCI Heart, PIMA Diabetes, Malaria Symptoms, and chronic kidney disease using accuracy, F1 score, AUC-ROC, RMSE, and inference delay as evaluation metrics. The EVNN-GRFN-M4 model achieved 97.8% accuracy, an AUC of 0.992, an RMSE of 0.06, and a 42 ms delay, outperforming Random Forest, SVM, and CNN baselines. Statistical testing (Friedman χ²(6) = 16.92, p = 0.009) confirmed the improvements as significant. By providing interpretable, uncertainty-aware outputs, the framework supports clinicians in assessing prediction confidence, consistent with WHO’s trustworthy-health technology guidelines and Nigeria’s Digital Health Strategy. Overall, EVNN-GRFN demonstrates that integrating fuzzy correlation control with evidential reasoning produces reliable, transparent, and efficient diagnostic tools suitable for resource-constrained healthcare environments.