Understanding various subpopulations in chronic kidney disease can improve patient care and aid in developing treatments targeted to patients' needs. Due to the general slow disease progression, electronic health records, which comprise a rich source of longitudinal real-world patient-level information, offer an approach for generating insights into disease. Here we apply the open-source ConvAE framework to train an unsupervised deep learning network using a real-world kidney disease cohort consisting of 2.2 million US patients from the OPTUM EHR database. Numerical patient representations derived from ConvAE are used to derive disease subtypes, inform comorbidities and understand rare disease populations. To identify patients at high risk to develop end-stage kidney disease, we extend a validated algorithm classifying disease severity to hypothesize subpopulations of rapid chronic kidney disease progressors. We demonstrate that using a combination of data-driven methods offers a powerful exploratory approach to understand disease heterogeneity and identify high-risk patients who could be targeted for early therapeutic intervention to prevent end-stage kidney disease.