Pareto-Optimized Deep Neural Network Framework for Multi-Objective Dengue Severity Classification

Abstract

Early and accurate prediction of dengue severity is critical for timely clinical intervention and reducing disease-related mortality. Traditional machine learning methods often struggle to simultaneously achieve high predictive accuracy, computational efficiency, and model scalability. To overcome these limitations, we propose a Pareto-Optimized Deep Neural Network (PO-DNN) framework that employs multi-objective optimization to minimize validation loss, computational cost, and model complexity. The framework generates a set of Pareto-optimal architectures, enabling a trade-off between predictive performance and operational efficiency. Using a preprocessed clinical dataset with 25 selected features, the PO-DNN was trained and evaluated via 10-fold cross-validation. Experimental results show that PO-DNN attains 96.8% accuracy, 95.7% recall, 97.1% precision, 96.4% F1-score, and an AUC of 0.975, outperforming baseline classifiers including SVM (87.3%), KNN (89.1%), Random Forest (91.0%), and conventional DNN (92.6%). Additionally, model optimization reduced average inference time to 1.9 ms per instance and compressed model size by 32%, demonstrating suitability for real-time deployment. These results highlight PO-DNN as a robust, efficient, and clinically applicable tool for dengue severity prediction.

KEYWORDS:

Dengue, prediction, classification, deeplearning, optimization, pareto, accuracy, efficiency, robustness, and scalability