Figure 3

\({\textsc {TransMED}}\) architecture. Patient context encoders are shown in (a) for static attributes and (b) for multi-modal temporal attributes. The proposed hierarchical transfer learning model is shown in (c). The transfer learning components take as input the patient’s multi-modal encoded state and produce a contextualized vector. The vectors for all time steps are combined along with static attributes to model patient’s (task-specific) evolution over time.