From R² to W²: rethinking predictive metrics for AI models in environmental health

Artificial intelligence (AI) has been increasingly used in environmental health research to predict pollutant concentrations, human exposure levels, and disease risks. However, the evaluation of predictive models depends largely on traditional metrics such as the coefficient of determination (R²) or predicted R², which is frequently misinterpreted in predictive modeling contexts. In many studies, R² is used interchangeably with predictive performance metrics, potentially leading to overestimation of model reliability. Moreover, conventional evaluation indicators primarily quantify correlation or error magnitude and may fail to detect systematic prediction bias across exposure gradients. As a result, models with high apparent accuracy can still exhibit discriminatory prediction patterns, such as consistent overestimation or underestimation across the prediction range. Here, we clarify the distinction between the R² and predictive squared correlation coefficient (Q²), and highlight the overlooked issue of prediction discrimination in AI-based environmental modeling. To address this limitation, we propose an integrated evaluation metric, W², which combines predictive accuracy with a penalty for systematic deviation from the 1:1 prediction line, as demonstrated with real environmental health data. This provides an auxiliary tool for assessing predictive reliability and improving the methodological rigor of AI-driven environmental health studies.