Interpretable Patient Trajectories from Temporally Annotated Health Records

By a trajectory, in the medical world, we mean the sequence of clinical events that occur to a patient in some time frame, as implicitly stored in patients' Electronical Health Records. A set of trajectories can be summarized in a trajectory graph, whose paths contain the most common trajectories followed by patients. The graph contains events on its nodes and the edges indicate the temporal relations. Previous works on building trajectory graphs only allow for one event at each node, and conversely for an event type to appear in only one node, thus losing information and potentially mixing different groups of patients. Here we develop a procedure to extract the trajectory graphs that goes beyond both limitations, thus more accurately reflecting the original dataset. In addition, it is close to a notion of patient state which clinicians use intuitively, facilitating interpretation. We evaluate the procedure on two real-world datasets, one related to diagnostics at hospital admissions, and the other on prescriptions in intensive care units, with reasonable and potentially useful results. The method is described here in the medical context only, but it is of general applicability for sequences of events.