Learning the Representation of Medical Features for Clinical Pathway Analysis

Clinical Pathway (CP) represents the best practice of treatment process management for inpatients with specific diagnosis, and a treatment process can be divided into several stages, usually in units of days. With the explosion of medical data, CP analysis is receiving increasing attention, which provides important support for CP design and optimization. However, these data-driven researches often suffer from the high complexity of medical data, so that a proper representation of medical features is necessary. Most of existing representation learning methods in healthcare domain focus on outpatient data, which get weak performance and interpretability when adopted for CP analysis. In this paper, we propose a new representation, RoMCP, which can capture both diagnosis information and temporal relations between days. The learned diagnosis embedding grasps the key factors of the disease, and each day embedding is determined by the diagnosis together with the preorder days. We evaluate RoMCP on real-world dataset with 538K inpatient visits for several typical CP analysis tasks. Our method demonstrates significant improvement on performance and interpretation.