Fig. 4: Clinical variable overlays on PaPaGei PPG embedding space.

We visualised the projections of patient-level PaPaGei embeddings using PCA, UMAP, and t-SNE, overlaid with clinically relevant variables. Dot outlines indicate binary cardiovascular disease (CVD) status (black = CVD, white = no CVD), while colour represents values of selected clinical predictors. (Top row) TDF_3TC_DTG usage is associated with a lower density of CVD-positive patients, consistent with protective antiretroviral therapy effects. (Bottom row) Changes in total cholesterol (TOTAL_CHOL_delta) across visits are higher in the CVD-positive clusters, supporting its predictive relevance. These overlays confirm that clinical variables identified as important by supervised models align with separable physiological patterns learned by PaPaGei without encoder fine-tuning.