Twin GNN AERIS: An Intelligent Physics-Guided Graph Neural Digital Twin Framework for Proactive Asthma Exacerbation Forecasting and Intervention Optimizations

Abstract

The primary reason asthma exacerbations continue to be a significant source of preventable morbidity worldwide is due to the dynamic interplay among environmental exposures, individual physiological vulnerabilities and patient behaviors. Predictive systems currently available are primarily based on population level statistical modeling or single variable analysis; both approaches do not address the need for understanding the evolving patient-specific causal mechanisms underlying the relationship between multiple scales of exposure and how patients will respond differently to varying environmental conditions. As such, predictive systems have short warning timescales, offer little to no personalized decision making options and provide minimal actionable information to clinicians as they attempt to navigate the rapid changes in environmental pollutants and weather patterns associated with their local environment. The proposed Graph Neural Digital Twin anticipates future asthma exacerbations using an integrated approach that utilizes combined environmental, spirometry and wearable sensor data samples to forecast asthma exacerbations. The Graph Neural Digital Twin system is comprised of five tightly-coupled analytical modules. The first module is referred to as the Multi-Scale Respiratory Exposure Graph Constructor (MR-EGC). The MR-EGC constructs a temporal supra-graph from heterogeneous physiological and environmental variables that encode cross-scale trigger relationships. The second module is referred to as the Causal Pulmonary Response Simulator using Physics-Guided GNN (CPRS-PGNN). CPRS-PGNN embeds airway mechanics into graph propagation to generate latent respiratory trajectories that are physiologically plausible. The third module is referred to as the Adaptive Environment-Behavior Interaction Encoder (AEBIE). AEBIE models context-dependent susceptibility by learning nonlinear interactions between activity patterns and exposure states. The fourth module is referred to as the Counterfactual Exacerbation Forecasting via Twin Divergence Modeling (CEFT-DM). CEFT-DM generates early risk estimates by calculating the divergence between the observed and optimized twin trajectories. The fifth and final module is referred to as the Personalized Intervention Optimization via Reinforced Twin Control (PIOR-TC). PIOR-TC develops actionable mitigation strategies through reinforcement learning within the digital twin environment. Together, the five modules comprise the Twin-GNN-AERIS system. The proposed Twin-GNN-AERIS system provides patient-specific asthma exacerbation forecasts over long warning periods, enhanced causal interpretability and preventative advice; all of which demonstrate substantial potential for reducing asthma exacerbation incidence, optimizing medication use and supporting precision respiratory care in real world environments subject to environmental volatility in process.