Unsupervised Learning Approaches for Identifying ICU Patient Subgroups: Do Results Generalise?

TL;DR

It is found that there is significant variation between ICUs; thus, a standardised restructuring approach is unlikely to be appropriate and potential efficiency gains might be greater when the number and nature of the subunits are tailored to each ICU individually.

Abstract

The use of unsupervised learning to identify patient subgroups has emerged as a potentially promising direction to improve the efficiency of Intensive Care Units (ICUs). By identifying subgroups of patients with similar levels of medical resource need, ICUs could be restructured into a collection of smaller subunits, each catering to a specific group. However, it is unclear whether common patient subgroups exist across different ICUs, which would determine whether ICU restructuring could be operationalised in a standardised manner. In this paper, we tested the hypothesis that common ICU patient subgroups exist by examining whether the results from one existing study generalise to a different dataset. We extracted 16 features representing medical resource need and used consensus clustering to derive patient subgroups, replicating the previous study. We found limited similarities between our results and those of the previous study, providing evidence against the hypothesis. Our findings imply that there is significant variation between ICUs; thus, a standardised restructuring approach is unlikely to be appropriate. Instead, potential efficiency gains might be greater when the number and nature of the subunits are tailored to each ICU individually.

Figures (3)

• Figure 1: Features used to derive the clusters. The clustering features can be separated into five domains: patient details, hospital admission, ICU, hospital discharge and post-discharge. The clustering features span the duration of patients' hospitalisation to ensure that the resulting clusters represent medical need throughout ICU rather than at a specific point in time. Notably, this means that direct patient triage at ICU admission would not be possible and this issue is discussed further in Section \ref{['sec:limitations_future_research']}.
• Figure 2: Consensus clustering results.A: The ordered consensus matrices show the stability of the clustering solutions from $K=2$ to $K=7$. Darker shading shows that a pair of examples were more frequently clustered together across the iterations. Therefore, cleaner, sharper matrices represent more stable clustering solutions. The colour bars above the plots show the partitioning of the data into clusters. The $K=8$ and $K=9$ matrices are significantly less clean and not shown. B: The CDFs show the proportion of ICU stays with unstable cluster memberships. Each CDF plots the cumulative distribution of indices in the corresponding ordered consensus matrix. A more stable clustering solution would have a higher proportion of consensus indices near 0 and 1 (white and dark blue on the ordered consensus matrices). This corresponds to a more step-like CDF. C: The tracking plot shows how cluster membership changes as $K$ increases from $2$ to $9$ (read from the top downwards). Unstable partitions may be visible at one level of the plot and then disappear as the granularity of clustering increases. For instance, a cluster at one level might be amalgamated into a larger cluster at a more granular level. Further details about interpreting these plots can be found in monti_consensus_2003senbabaoglu_critical_2014wilkerson_consensusclusterplus_2010.
• Figure 3: The relationship between COPS II and CCI. It appears that CCI is a relatively good substitute for COPS II because they have an approximately linear relationship. A limitation is that this trend breaks down for higher CCI scores. Despite differences, both should capture similar underlying dynamics. This figure is taken from the supplementary methods of Escobar et al. escobar_risk-adjusting_2013.