Effective Patient Similarity Computation for Clinical Decision Support Using Time Series and Static Data

This paper presents a technique for computing patient similarity using time series data effectively combined with static data. Time series data of inpatients, such as heart rate, blood pressure, Oxygen saturation, respiration are measured at regular intervals, especially for inpatients in intensive care unit (ICU). The static data are mainly patient background and demographic data, including age, weight, height and gender. The similarity computation is done in unsupervised way. It is therefore free from data labeling requirement. However, such patient similarity can be very useful in developing various clinical decision support systems including treatment, medication, hospital admission and diagnosis. Our proposed technique works in three main steps. First, patient similarity is computed for each individual time series. Second, patients are grouped by clustering the static data. Finally, similarities from individual time series are combined and effectively blended with the patient group information to create a nearest neighborhood model. This model consists of a collection of the nearest neighbors for a given patient. We encounter several challenges for this task, including dealing with multi-variate time series data, variable sampling quantities and rates, missing values, and combining time-series with static data. We evaluate the proposed technique on a real patient database on two target features, namely, 'diagnosis' and 'admission type'. Notable performance is recorded for both targets, achieving f1-score as high as 0.8. We believe this technique can effectively combine different types of clinical data and develop an efficient unsupervised framework for computing patient similarity to be utilized for clinical decision support systems.