An Unsupervised Homogenization Pipeline for Clustering Similar Patients using Electronic Health Record Data

Abstract

Electronic health records (EHR) contain a large variety of information on the clinical history of patients such as vital signs, demographics, diagnostic codes and imaging data. The enormous potential for discovery in this rich dataset is hampered by its complexity and heterogeneity. We present the first study to assess unsupervised homogenization pipelines designed for EHR clustering. To identify the optimal pipeline, we tested accuracy on simulated data with varying amounts of redundancy, heterogeneity, and missingness. We identified two optimal pipelines: 1) Multiple Imputation by Chained Equations (MICE) combined with Local Linear Embedding; and 2) MICE, Z-scoring, and Deep Autoencoders.