Abstract Similar patient retrieval has become increasingly important with the explosive growth of electronic health records (EHRs). A similar patient cohort identified from all patients can provide data-driven insights for personalized healthcare. However, the high dimensionality and heterogeneity of EHRs increase the difficulty of measuring patient similarity. How to accurately and efficiently retrieve similar patients from a large number of EHRs remains challenging. In this paper, we propose a novel similar patient retrieval method based on interactive patient labeling and automatic model updating. Combined with the knowledge and experience of physicians, it can be adaptively modified for different patients and diseases. We also develop a visual analytics system to assist patient labeling through pairwise comparisons and support outcome analysis of similar patients. The case studies on two real-world datasets in collaboration with physicians demonstrate the effectiveness and usefulness of our method.Graphic Abstract
展开▼
