A Context-Aware End-to-End Predictive Analytics Architecture for Cholera Early Warning and Medical Resource Allocation Using Machine Learning

Abstract

This paper details the architecture, development, and assessment of a comprehensive predictive analytics platform designed to forecast cholera outbreaks and optimize the allocation of medical resources within Zambia's public health system. Conventional cholera management in the region is predominantly reactive, resulting in operational delays and suboptimal resource deployment. This research confronts this issue by creating a localized, proactive decision-support tool. The system utilizes a hybrid modeling approach, combining supervised learning algorithms (Logistic Regression, Random Forest, XGBoost) with a linear programming (LP) model for resource optimization. A comparative analysis was performed using a synthetic dataset from 2017-2024 that mirrors Zambia's epidemiological trends. The XGBoost model yielded the most effective performance for an early warning system, attaining an accuracy of 84.69%, a flawless recall of 1.0, and an AUC-ROC score of 0.9361. Conversely, the Random Forest model provided perfect precision (1.0) but with a minimal recall of 0.125, underscoring significant performance trade offs. The resulting prototype, which includes an interactive Streamlit dashboard, effectively transforms predictive outputs into actionable resource allocation strategies, offering a scalable and data driven solution for epidemic preparedness in resource limited settings.