Forecasting tuberculosis through mechanistic learning of transmission dynamics: Insights from a case study in India.

Abstract

Tuberculosis continues to pose a significant global health issue, with India facing the most considerable burden. We develop and calibrate an SEIR-type model to understand disease dynamics and perform sensitivity analysis to pinpoint crucial parameters. The basic reproduction number acts as the epidemic threshold, while backward bifurcation highlights the dangers of reinfection and superinfection. To enhance forecasting, we merge mechanistic modeling with deep learning techniques, utilizing feedforward, recurrent, and memory-based networks. Evaluated on TB case data from India, these hybrid models outperform the SEIR baseline, with the gated recurrent unit best capturing residual trends and the feedforward network demonstrating robust generalization. This integrated framework enhances predictive accuracy and interpretability, providing a valuable tool for TB forecasting and aiding in targeted interventions.