Towards modeling evolving longitudinal health trajectories with a transformer-based deep learning model

Purpose: Health registers provide valuable insights into individuals’ health trajectories. This study explores the use of deep learning to model and analyze these trajectories using a nationwide longitudinal dataset containing coded features such as clinical codes, procedures, and drug purchases. Methods: We introduce Evolve, a transformer-based deep learning model designed to provide continuous multi-label predictions over time. The model predicts disease onsets at each time step conditioned on the health history up to that time step and the time until a given 5-year forecast window. Evolve is evaluated against several baseline models for basic prediction performance. Additionally, we analyze health trajectories by tracking changes in prediction probabilities and in the latent embedding neighborhood to identify important events. Results: Evolve performed comparably to baseline models in disease onset prediction while offering unique trajectory modeling capabilities. The model identified early predictive events and demonstrated that changes in embedding space could indicate shifts in health trajectories. Visualization of evolving health trajectories showed how individuals may become most similar to others with similar profiles and outcomes over time. Conclusions: The Evolve model seems promising at enabling continuous health monitoring, early disease detection, and retrospective analysis, making it a promising tool for personalized healthcare interventions. Code available at: https://github.com/hansmoen/evolvehealth.