Leveraging Deep Learning Approaches for Patient Case Similarity Evaluation

Abstract

One of the fundamental problems in Health Informatics is evaluating the clinical similarity between two patients for treatment recommendation. Retrieving clinical records of existing patients who are potentially similar to a newly arrived patient could help a physician in faster diagnosis and recommending informed treatment options, especially in the case of areas where specialist medical care is scarce. In Western countries, patient records are extensively stored in the form of Electronic Health Records (EHR), which are created manually by human experts, which can take a lot of time and is a cost-intensive process. In developing countries like India, patient records are increasingly being stored in digital formats and often contain diverse, heterogeneous, unstructured reports of patients. These can be potentially utilized for designing patient similarity assessment and recommendation systems. In this paper, we propose a patient similarity evaluation framework built on two supervised learning models—Convolutional Neural Networks (CNN) and Gated Recurrent Unit (GRU). Our method creates an optimal patient representation for existing patients by aggregating reports collected over the duration of treatment, to overcome the loss of temporal information, for which a cohort of 16,723 patients across 8 disease categories was used. Both the models (CNN and GRU) learn by passing through the records of a patient chronologically and achieve an accuracy of 97.60 and 93.62%, respectively, on standard EHR dataset like MIMIC-III.