AI-Driven Framework for vector-borne disease surveillance and early warning

Abstract

Vector-borne diseases, especially dengue fever, are a significant global public health issue. They account for 17% of infectious diseases and cause over 700,000 deaths each year. Current surveillance systems struggle with insufficient detail and slow response times, making it hard to predict outbreaks effectively. This project creates an AI-powered framework for predicting dengue outbreaks at a local level by combining epidemiological, meteorological, geospatial, and demographic data into a single predictive tool. The proposed framework integrates data management, disease forecasting, surveillance visualization, and AI-assisted decision support within a unified operational architecture. The data integration method aligns time-series cases from Punjab with multimodal data from Puerto Rico, such as NDVI and precipitation, along with diagnostic features. This setup allows for validation across different regions. The machine learning engine follows a three-phase approach. In Phase 1, it implements XGBoost, achieving a test R² score of 0.970 on the temporally separated Punjab forecasting dataset. In Phase 2, it develops an ensemble approach that combines XGBoost, LightGBM, and Random Forest, achieving an R² of 0.9779. Phase 3 explores LSTM-based forecasting architectures, while Transformer-based approaches are considered as potential extensions for future enhancement. An exploratory component of the framework is the integration of a Lang Chain-powered Large Language Model to provide contextual decision-support recommendations and improve interpretability of forecasting outputs. This module was functionally demonstrated and qualitatively evaluated rather than quantitatively validated. A web-based surveillance dashboard was implemented for visualization of predictions, environmental indicators, and geographic risk patterns. Validation shows strong performance, with the Phase 2 ensemble achieving the R² of 0.9779. Under the evaluated datasets and temporal validation strategy, the proposed ensemble demonstrated higher R² than several studies reported in the reviewed literature, including studies reporting R² values around 0.84 and Random Forest models reporting approximately 89% accuracy. The proposed framework has the potential to support district-level outbreak monitoring and public health decision-making. The system's design allows for future adjustments to address multiple vector-borne diseases by using cross-regional applicability. Further evaluation is required to assess large-scale deployment and operational effectiveness in real-world public health settings.