Fig. 3: Description of the risk prediction and temporal clustering outputs and trajectories generated.

A Kaplan Meier plot of the survival distributions and confidence intervals for the 4 temporal phenotypic clusters identified from the DDHL model showing the different rates of progression to Cambridge Prognostic Group 3 (CPG3) event over time. B Same temporal clusters represented in the space of the two principal components (PCA) of the latent embeddings (each point corresponds to a sampled patient). Trajectories of two illustrative example patients A and B also shown: A deteriorates as more observations are collected, moving from Cluster 3 to 4, while B improves, moving from Cluster 2 to 1. C A schematic illustrating the risk prediction and temporal clustering aspects of the DDHL model. (Right-hand side) Hypothetical Clusters 3 and 4 are illustrated to show the sets of past patients assigned to them in training. Cluster 4 is presented to contain higher risk patients; hence the histogram indicates an overall earlier time-to-event among these patients. Cluster 3 is presented to contain relatively lower risk patients, hence the histogram peaks at a later time-to-event. (Left-hand side) A hypothetical patient’s progression: at time t1 the patient has a comparatively steep risk prediction curve (red slope), and the model assigns him to Cluster 4; as more observations are made at time t2, the risk profile is lowered, patient is now in Cluster 3.