Patient Similarity Analysis
Since the early days of computing, healthcare professionals have dreamt of using the vast storage and processing powers of computers to sift through vast medical archives and automatically discover new facts and medical knowledge locked inside electronic health records (EHRs) and similar digital patient data archives. This is not unlike the activity of physicians that use their clinical experience to identify patterns (such as symptoms) that are common among a group of patients, leading to new disease classifications and eventually treatment strategies. However, rather than having to base judgement on few patients, digital medical information allows the inspection of data from thousands if not millions of patients at the same time. In order to compare patients at this scale, we need to teach computers to form a robust understanding of patient similarity, a non-trivial skill that relies on matching multi-dimensional and multi-modal patient features that are captured at different times and resolutions. As these features are captured over time, they form unique patient journeys that chronicle healthcare interventions and record patient outcomes. Measuring the similarities among those journeys enables the discovery of common health states that precede disease development. As patients traverse these states in distinct pathways, it is possible to inspect the trajectories for a better understanding of possible disease etiologies. Also, as pathways are linked to particular health outcomes (trajectories), they are useful for prognosis and medical decision making. In this project, our general aim is to investigate and apply computational methods for patient journey analysis. Our specific aims are to (1) establish algorithmic and visualization approaches for learning and assessing patient journey similarities and to (2) perform personalized outcome prediction based on these journeys using data from the Departments of Rheumatology and Intensive Care of the University Hospital of Zurich (USZ).