Multimorbidity profiles and stochastic block modeling improve ICU patient clustering

Identifying groups of patients with similar morbid-ity profiles can help us understand the relationships between their pre-existing conditions and the risks of adverse events in the ICU. To find such groups, common approaches apply clustering algorithms such as k-means and latent class analysis. However, these techniques present drawbacks such as the lack of principled methods for choosing the number of clusters, the need for assumptions about the relationships between variables, and outputs which are hard to explain. To overcome these limitations, we map the problem of patient clustering to that of community detection in complex networks. We construct a bipartite network in which nodes represent patients and their features, including morbidities and demographics. Then, we find homogeneous groups of patients using stochastic block modeling (SBM), an unsupervised probabilistic approach to find structure in networks. We show that this approach has several advantages over traditional clustering methods, and enables us to retrieve more fine-grained clusters that are commonly missed by existing approaches. We also show that these clusters have a stronger relationship with mortality and sepsis rates of patients in the ICU.