An Uncertainty-Aware Framework for Integrating Clinical AI and Environmental Exposure Prediction

Abstract

Ground-level ozone has been associated with adverse respiratory outcomes and lung cancer risk. However, existing approaches rarely integrate clinical risk modeling with environmental exposure prediction into a unified AI-based decision-support framework. Using 1,795,575 SEER records (2010 - 2023, multiple cancer types), we developed an XGBoost baseline predicting vital status (AUC-ROC = 0.848) and a Cox proportional hazards model as our primary clinical model, accounting for variable patient follow-up time (concordance index = 0.805). As validation, both recovered the established stage-survival relationship (localized disease: HR = 0.24; liver cancer: HR = 3.25 vs. bladder cancer). Because SEER lacks patient-level geographic identifiers, patient-level clinical data cannot be directly linked to environmental exposure data. Instead, independently trained models are combined at inference using user-provided clinical and location information through deterministic score-level fusion. The environmental model predicts ground-level ozone concentrations from meteorological, atmospheric, temporal, and location features, estimates probabilities of exceeding 55 and 70 ppb with conformal prediction intervals, and combines these outputs with a patient-derived clinical-context score to generate four-tier contextual alerts. This framework was deployed within the Aegis AI application. The integrated framework was evaluated through algorithmic validation, robustness testing, and ablation studies, demonstrating stable, monotonic, and uncertainty-aware fusion behavior across representative patient scenarios. These results demonstrate that uncertainty-aware score-level fusion can provide contextual environmental decision support, forming the computational foundation of Aegis.