Climate change and environmental degradation increasingly threaten urban public health, yet monitoring infrastructure is often unavailable in resource-constrained settings. This study develops and critically evaluates a machine learning framework that predicts environmental health-related search behavior , used as a low-cost digital proxy rather than a measure of clinical burden, in Morocco using exclusively open-access global datasets. We integrated five years of weekly data (April 2020 to April 2025) from ERA5 meteorological reanalysis, Copernicus Atmosphere Monitoring Service products, and Google Trends to model binary high-interest states for four search categories: air quality, respiratory conditions, heat-related illness, and allergies. Ensemble models with 71 engineered features achieved strong classification performance (ROC-AUC 0.92 to 0.96) with temporal stability over 52-week horizons. An ablation study showed that this discrimination is carried primarily by the autocorrelated search signal; the environmental reanalysis block added negligible incremental ROC-AUC at national scale, yet retained meaningful standalone signal for the respiratory and air-quality categories while collapsing to near-random for allergy and heat-related illness. This contrast tracks biological plausibility, which varied substantially across categories: respiratory predictions aligned with established epidemiological patterns, whereas heat-related predictions showed weak temperature correspondence and implausible winter peaks. We therefore frame the system as a complementary screening layer pending external validation against clinical endpoints, not as a validated disease-surveillance or operational warning system. Because it relies solely on freely available data, the framework offers a replicable national-scale prototype for environmental health monitoring across North Africa and the Global South, contributing to SDG 3, 11, and 13.