Introduction

Urban health inequalities remain a pressing challenge for city planners and public health policymakers, particularly in rapidly growing metropolitan areas of developing countries where uneven infrastructure distribution, environmental pollution, and socioeconomic disparities intersect in complex ways. Ahvaz, a major industrial and demographic hub in southwestern Iran, exemplifies these challenges, facing persistent environmental stressors such as air pollution, degraded water quality, and inequitable access to health services. Much of the existing literature on urban health inequality, however, tends to rely on coarse administrative units—such as districts or municipal zones—that obscure fine-grained spatial patterns, or on composite indices constructed through subjective, expert-driven weighting schemes that limit reproducibility and generalizability. This study addresses both limitations by introducing a novel, data-driven metric termed the Urban Health Needs Index (UHNI), designed to identify vulnerable areas requiring prioritized health interventions at a micro-spatial scale.

Methodology

To move beyond conventional administrative boundaries, we adopted a hexagonal grid system with a side length of approximately 58 meters as the spatial unit of analysis, ensuring uniform coverage across the city and eliminating boundary-related distortions common in irregularly shaped administrative zones. The empirical foundation of this research comprises over 60,000 clinical records covering cardiovascular, respiratory, ophthalmic, dermatological, and oncological conditions, alongside approximately 9,650 mortality records—including infant mortality and cardiovascular mortality—spanning the years 2015 to 2019. These health outcome data were complemented by 25 environmental, social, economic, demographic, and infrastructural variables, including multiple water quality indicators (total dissolved solids, coliform bacteria, turbidity, electrical conductivity, and pH), municipal and industrial wastewater discharge, pesticide residues, proximity to sports facilities and healthcare clinics, and social indicators such as child labor prevalence.

Given that raw disease rates in small spatial units are prone to statistical instability arising from sparse population counts, we applied spatial empirical Bayes smoothing to stabilize the rate estimates prior to modeling. A Random Forest Regressor was then trained to model the relationship between the smoothed health outcome rates and the 25 predictor variables, with hyperparameters optimized using the Optuna framework to maximize predictive performance while minimizing overfitting. To move beyond a “black-box” application of machine learning and to derive interpretable, data-driven variable weights—rather than relying on arbitrary or expert-assigned weighting—we employed SHAP (SHapley Additive exPlanations) values to quantify each variable’s relative contribution to the model’s predictions. This approach represents a methodological departure from conventional composite health indices, replacing subjective weighting with an empirically grounded, transparent alternative. Spatial clustering of the resulting UHNI values was examined using the Getis-Ord Gi∗ statistic to detect statistically significant hot and cold spots, and external validity was assessed through Spearman correlation between UHNI scores and independent mortality outcomes.

Results and Discussion

The final Random Forest model achieved a coefficient of determination of R2=0.687 and a root mean square error of RMSE=0.015, indicating that the model captured a substantial portion of the variance in urban health needs based on environmental and social predictors. SHAP-based feature importance analysis revealed that the most influential determinants of urban health needs were, in descending order: municipal wastewater (13.7%), total dissolved solids or TDS (10.7%), coliform contamination (8.5%), water turbidity (7.9%), electrical conductivity or EC (7.1%), industrial wastewater (5.9%), pesticide residues (5.7%), distance to sports facilities (5.5%), water pH (5.2%), child labor (4.9%), distance to healthcare clinics (4.6%), and inflow wastewater (4.1%). Collectively, these findings underscore that environmental pollution—particularly degraded water quality and wastewater contamination—constitutes the dominant driver of health inequality in Ahvaz, outweighing purely socioeconomic factors in explanatory power.

Spatially, hot-spot analysis identified persistently elevated UHNI values concentrated in the central and northwestern districts of Ahvaz, particularly along the corridor adjacent to the Karun River and areas receiving concentrated municipal wastewater discharge. Neighborhoods including Kuy-e Alavi (Shalangabad), Khashayar, Kamplo, and portions of Ahanafshar consistently emerged as health-need hot spots across nearly all five years of analysis, while cold spots stabilized predominantly in the eastern and southern parts of the city. The persistence of this spatial pattern over a five-year period strongly supports the interpretation that health inequality in Ahvaz is structurally embedded rather than a product of year-to-year statistical noise.

External validation demonstrated that UHNI scores were significantly and positively correlated with infant mortality (p=0.003) and cardiovascular mortality (p=0.036), lending empirical credibility to the index’s capacity to reflect real-world health disparities. Nevertheless, the authors acknowledge that actual health outcomes are shaped by numerous additional factors beyond the scope of this model. To further assess the robustness of the findings, spatial uncertainty was quantified based on the variability of the three most influential predictors (municipal wastewater, TDS, and coliform levels) combined with the overall model error term (1−R2=0.3131 ), highlighting the importance of jointly considering predicted need levels and their associated confidence when informing policy decisions.

Conclusion

The principal methodological contribution of this study lies in its integration of interpretable machine learning—Random Forest combined with SHAP—with fine-grained, hexagonal-grid spatial analysis and systematic uncertainty quantification, offering a replicable framework for constructing transparent, data-driven urban health indices. Based on these findings, we recommend that health policy in Ahvaz shift away from uniform resource allocation and instead adopt a spatially targeted approach guided by UHNI mapping, prioritizing interventions in persistently high-need central and northwestern neighborhoods. Concrete policy priorities include stricter control of municipal and industrial wastewater discharge, continuous monitoring and improvement of water quality parameters, expanded access to healthcare facilities, and the development of dynamic, periodically updated versions of the index. More broadly, this study demonstrates that combining clinical and mortality data with environmental and social variables through interpretable machine learning can provide an effective, evidence-based tool for equitably prioritizing urban health needs in rapidly developing metropolitan contexts.