PathoCast: An Explainable AI Framework for Predicting Daily Dengue Risk in Low-Resource Urban Settings

Introduction:
Dengue poses a recurring public health challenge in urban centers in low- and middle-income countries, where limited surveillance capacity delays outbreak detection and response. Early-warning tools are especially needed in regions like Lahore, Pakistan, where rapid transmission and climate variability make reactive strategies insufficient. This study introduces PathoCast, an explainable artificial intelligence framework designed to classify daily dengue risk using multimodal environmental, epidemiological, and behavioral indicators.

Methods:
A daily dataset for the year 2022 was constructed using dengue case counts, meteorological variables, and Google Trends search behavior. Temporal and environmental predictors were derived through feature engineering, and synthetic geographic identifiers were added to enable spatial visualization. An XGBoost classifier served as the primary model, supported by an interpretability layer based on grouped SHAP (SHapley Additive exPlanations) values to quantify the contribution of predictor categories.

Results:
Grouped SHAP analysis demonstrated that temporal transmission indicators exerted the strongest influence on risk classification (mean |SHAP| = 0.13), followed by environmental variability features (mean |SHAP| = 0.07). Behavioral trend proxies contributed modestly (mean |SHAP| = 0.04) but provided complementary signal value. Rainfall-related indicators showed minimal short-term predictive effects. Spatial heatmaps generated from synthetic district coordinates revealed distinct clusters of elevated predicted risk, illustrating how the system could guide targeted vector control once linked to real administrative boundaries.

Conclusions:
This study on PathoCast demonstrates the feasibility of an interpretable, low-cost AI framework capable of supporting localized dengue early-warning efforts in resource-constrained settings. Future work will include multi-season model validation, the integration of real-time meteorological and behavioral data APIs, and the development of a mobile interface for community health workers.