Contrastive Learning on Temporal Graphs Models Disease Trajectories

A new framework for contrastive representation learning has been introduced by researchers, which depicts multivariate disease trajectories as temporal graphs and utilizes contrastive graph neural networks to learn embeddings. In this model, nodes symbolize patient observations across time, while edges illustrate both temporal continuity and structural resemblance. By employing structure-aware random walks, the contrastive learning process maintains the temporal context and trajectory topology. This approach facilitates effective clustering of patients exhibiting similar disease progression patterns and uncovers hidden structures within longitudinal data. The study tackles the difficulties associated with interpreting disease trajectories from longitudinal clinical data, given the intricate temporal dynamics and diverse patient groups.