HybridNet-SEVIT: Multilabel Classification for Epidemic Risk Management and Public Health Surveillance

Abstract

Effective public health monitoring and epidemic prevention are crucial in mitigating the impact of infectious diseases. The dynamic nature of disease transmission necessitates a comprehensive understanding of demographic, health, environmental, and behavioral data. Despite advances in data collection, conventional methods often fail to accurately model disease dynamics, leading to suboptimal predictive capabilities. To address these challenges, we integrate a dataset of 43,689 entries from multiple sources and employ innovative preprocessing techniques such as Adaptive Distribution Recalibration, Contextual Outlier Filtering, and Multiscale Variance Modulation to enhance prediction accuracy and dataset integrity. Our proposed HybridNet-SEVIT model, which incorporates depthwise separable convolutions, dense connectivity, and ghost feature generation, achieves superior classification performance, with an accuracy of 97.9% and AUC values ranging from 96% to 99%. Comparative analysis demonstrates that HybridNet-SEVIT outperforms SOTA models, including ResNet (90.0% accuracy), DenseNet (84.1%), and SVM (84.8%), showing a significant 7.9% improvement in accuracy over the best-performing baseline model. Additionally, novel evaluation metrics—Adaptive Variability Index for Classes (AVIC), Stability of Prediction Dynamics Measure (SPDM), and Confidence Level Weighted Score (CLWS)—offer deeper insights into model robustness and predictive confidence. This study contributes to enhancing risk detection and classification in public health, advocating for a robust, data-driven approach to epidemic outbreak management and emphasizing the need for targeted interventions and efficient resource allocation.