Air pollution and urban warming are increasingly intertwined, with heat-enhanced photochemistry making ozone an increasingly important concern for population health even where PM_2.5 has declined. Here we develop an interpretable machine-learning environmental health framework to characterize non-linear long-term exposure-response patterns for cold-season PM_2.5 and warm-season ozone and to compare pollutant-specific risk patterns over time. Using cancer panel data from 813 counties in China (2018-2019) and multi-source exposure and contextual indicators, we estimate model-based reference thresholds and construct threshold-anchored exceedance and normalized excess-hazard indices for cross-year spatiotemporal comparison across 2017-2023. The resulting indicators suggest a divergence in pollutant-related risk patterns, with lower PM_2.5-related pressure and more prominent ozone-related pressure over time, a pattern consistent with strengthening temperature-O₃ coupling. The spatial pattern of the ozone-dominance index shows clustered hotspots aligned with major urban agglomerations and densely populated corridors, with notable concentrations in the Sichuan-Chongqing urban cluster and the Yangtze River Delta. When combined with population ageing, the ozone-dominance pattern indicates compound vulnerability that is unevenly distributed and increasingly concentrated in higher-ageing urban areas over time, suggesting growing spatial inequality in ozone-related risk patterns. These findings provide effect-anchored indicators for spatiotemporal risk surveillance and environmental safety prioritization under evolving air-pollution regimes.