FIGURE 2.

Schematic diagram of the PathoGraph. PathoGraph is a clinically-informed neural architecture designed to model temporal dependencies in patient records through a combination of ontology-aware embedding initialization, disentangled temporal representation, and a knowledge-guided masking mechanism. The model integrates domain knowledge from medical ontologies to enrich event embeddings, disentangles latent clinical factors over time to enhance interpretability, and uses relational structures to filter implausible co-occurrences. A modular fusion of features via dynamic routing and MoLE (Mixture-of-Low-rank Experts) adapters within a large language model further supports participatory design and equity evaluation in medical decision-making.