Jan 12, 2018 | 14:00—15:30
Abstract: The availability of large scale healthcare data, promises to revolutionise medical care, with treatments tailored individually to each patient. In the last decade, the analysis of medical records, often recorded for billing purposes, was centered around binary relations, such as the comorbidity, between two diseases. Comorbidity quantifies the probability that a patient with disease A also has disease B. While this measure is often sufficient for practical purposes, in other cases it may be beneficial to consider the long term history of a patient, that is the interplay of more than two diseases, and their temporal order. For example, the simultaneous presence of diseases A,B and C could indicate the presence of a particular genetic defect, increasing the likelihood for disease D. To our knowledge, the time sensitive analysis of interactions of more than two diseases has not yet been pursued systematically in the literature. The aim of our work is to bridge this gap by developing a method to predict future diagnoses of a patient which takes into account the entire history of the patient’s diagnoses. To this end, we analyse a data set containing for each inpatient hospital stay in Austria the ID of the patient, the main and side diagnoses and the admission and release dates for the years 2006 and 2007, and the same information for Lower Austria for the years between 2008 and 2011. The data thus contains a “disease trajectory” for each of the around 8M people insured via an Austrian social security provider. This data can be combined with data for health care provider (e.g. GP, dermatologist, …) visits and data on medical prescriptions, which is available to us for the same patient cohort. In this talk I will explain our current ideas on how to tackle the problem of diagnosis prediction beyond pairwise statistics.