Similarity Measure for Patients via A Siamese CNN Network

In the medical health field, assessing the similarities between patients is a basic task. A suitable patient similarity measurement has a very wide range of applications. For example, patient group identification, comparative study of treatment methods, etc. The electronic health records (EHRs) contain rich personal information of patient, which is hierarchical, longitudinal, and sparse. Although there have been some studies aimed at learning the similarities of patients from EHRs to solve real medical problems, there still exist some problems. Many works lack of effective patient representation. In addition, most of the research works are limited to one or more specific diseases. However, in fact, many diseases accompany with other diseases. In this case, the similarity of patients with multiple diseases are ignored. In this paper, we designed a siamese CNN network structure to learn patient expression while effectively measure the similarity between patient pairs. The experimental results show the effectiveness of this method.