Process mining of disease trajectories in MIMIC-III: a case study

A temporal disease trajectory describes the sequence of diseases that a patient has experienced over time. Electronic health records (EHRs) that contain coded disease diagnoses can be mined to find common and unusual disease trajectories that have the potential to generate clinically valuable insights into the relationship between diseases. Disease trajectories are typically identified by a sequence of timestamped diagnostic codes very similar to the event logs of timestamped activities used in process mining, and we believe disease trajectory models can be produced using process mining tools and techniques. We explored this through a case study using sequences of timestamped diagnostic codes from the publicly available MIMIC-III database of de-identified EHR data. In this paper, we present an approach that recognised the unique nature of disease trajectory models based on sequenced pairs of diagnostic codes tested for directionality. To promote reuse, we developed a set of event log transformations that mine disease trajectories from an EHR using standard process mining tools. Our method was able to produce effective and clinically relevant disease trajectory models from MIMIC-III, and the method demonstrates the feasibility of applying process mining to disease trajectory modelling.