Abstract
Using electronic health records (EHR) as the source of data for mining and analysis of different health conditions has become an increasingly common approach. However, due to irregular observation times and other uncertainties inherent in medical settings, the EHR data sets suffer from a large number of missing values. Most of the traditional data mining and machine learning approaches are designed to operate on complete data. In this paper, we propose a novel imputation method for missing data to facilitate using these approaches for the analysis of EHR data. The imputation is based on a set of interpatient, multivariate similarities among patients. For a missing data point in a patient's lab results during his/her intensive care unit stay, the method ranks other patients based on their similarities with the ego patient in terms of lab values, then the missing value is estimated as a weighted average of the known values of the same laboratory test from other patients, considering their similarities as weights. A comparison of the estimated values by the proposed method with values estimated by several common and state-of-the-are methods, such as MICE and 3D-MICE, shows that the proposed method outperforms them and produces promising results.