Integrated Parameter-State Estimation and Deep Learning Forecasting for Measles Dynamics Using Hybrid Evolutionary Filtering

Abstract

Measles transmission in Indonesia remains difficult to predict due to its high transmissibility and irregular outbreaks. This paper proposes a hybrid GA-EKF-LSTM framework to estimate and forecast measles dynamics through the $\mathrm{S V}_{1} \mathrm{V}_{2} \mathrm{E I H R}$ compartmental model, calibrated to monthly cumulative case data from January 2022 to November 2025. A Genetic Algorithm (GA) estimates the model parameters, the Extended Kalman Filter (EKF) reconstructs the unobserved epidemiological states, which then train a Long Short-Term Memory (LSTM) network for forecasting. The calibrated model achieved a fitting Mean Absolute Percentage Error (MAPE) of 12.34%. Although the infected compartment alone does not fully determine the system, EKF reconstructs all states with low MAPE. For prediction, EKF outperformed EKF-LSTM for all compartments, while a pure LSTM trained directly on the observed infected series achieved the highest accuracy. These findings support Sustainable Development Goal 3 (Good Health and Well-being) by providing a decision support tool for measles outbreak management in Indonesia.