Finding Similar Patient Subpopulations in the ICU Using Laboratory Test Ordering Patterns
Anis Sharafoddini, Joel A. Dubin, Joon Lee · 2018
In this paper, we focus on phenotyping critically ill patients in intensive care units (ICUs). Various data types have been used to cluster patients. We introduce laboratory-test-ordering patterns as a source of information for finding clinically similar patients. We employed Density-Based Spatial Clustering of Applications with Noise (DBSCAN) clustering method to find patient subpopulations based on the first 24 hours of laboratory test ordered. The DBSCAN identified 25 clusters, and we utilized t-Distributed Stochastic Neighbor Embedding (t-SNE) to visualize the subpopulations. Then, we evaluated the clinical interpretability of the clusters by using cluster characteristics and two outcomes: in-hospital mortality and 30-days post-discharge mortality. Our results demonstrate that laboratory-test-ordering patterns are informative and can be used to identify patient cohorts.