Predictive Modelling of Lassa Fever Outbreaks: A Comparison of Deep Learning and Ensemble Approaches

A Comparison of Deep Learning and Ensemble Approaches

Abstract

Lassa fever remains a critical and highly perilous public health threat across West Africa, with Ondo State, Nigeria, consistently experiencing severe annual outbreaks. While accurate forecasting is essential for timely medical intervention, conventional predictive models often fail to capture the complex spatiotemporal transmission dynamics and the delayed ecological triggers of the Lassa virus. To address this, this study introduces a comprehensive Health Information Pattern Discovery framework to forecast outbreaks across four high-risk localized hotspots: Akure South, Akure North, Akoko Southwest, and Owo. Advancing beyond standard machine learning benchmarking, this research integrates spatial-temporal graph theoretic modelling to map the relational transmission velocity between regions , coupled with a Genetic Algorithm (GA) to optimize complex lagged meteorological variables, including temperature, relative humidity, and precipitation. The optimized spatiotemporal feature space was utilized to train and evaluate four distinct architectures: Random Forest, Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Long Short-Term Memory (LSTM) networks. Empirical evaluation demonstrated that the GA-optimized Random Forest ensemble outperformed the other models, achieving a Root Mean Squared Error (RMSE) of 5.627, a Mean Absolute Error (MAE) of 3.747, and an of 0.87. Beyond baseline predictive accuracy, this hybrid graph-theoretic and evolutionary approach successfully isolates the exact environmental triggers preceding an outbreak, providing interpretable, actionable intelligence to strengthen proactive public health planning and targeted interventions in endemic regions.