Temporal modeling of patient education engagement using health information access
Abstract
In the digital healthcare transformation era, patient engagement through access to health information is crucial for improving education outcomes, adherence, and long-term health management. Personalized, temporally aware educational interventions are essential for effective patient engagement, yet current approaches often fail to capture the temporal dynamics of patient interactions. Traditional models rely on static representations or autoregressive methods that struggle with long-range dependencies, noise, missing data, and fail to incorporate domain specific constraints like behavioral trends or causal influences. To address these challenges, we introduce an innovative framework called the Temporal Latent Gradient Network (TLGN), which models temporal dynamics using latent trajectories derived from gradients in a non-autoregressive manner. TLGN embeds latent variables into a continuous gradient field modeled by neural differential equations, effectively capturing long-term engagement patterns in patient behavior. To further enhance this model, we introduce the Causal-Aware Temporal Denoising (CATD) strategy, which integrates domain knowledge and causal priors through structured constraints. This strategy enforces trend continuity, causal alignment, and robust forecasting against noise. Experiments on both synthetic and real-world patient education datasets show that TLGN with CATD outperforms traditional methods in predictive accuracy and interpretability. This approach significantly improves the modeling of health information access sequences, preserves key behavioral trends, and enhances system responsiveness to irregular user interactions. The combined framework offers a powerful, generalizable solution for advancing digital patient education, aligning with the growing need for personalized, temporally adaptive healthcare technologies.