An Unsupervised Machine Learning Analysis of OPTIM to Identify Workload Factors Related to Mortality

Introduction: High workloads for healthcare workers undermine team-based care and result in reduced patient safety and healthcare worker burnout rates exceeding 70%. The primary purpose of this large, multi-center trial was the generation of a rich, externally valid dataset capable of supporting robust unsupervised machine learning (ML) analyses to capture the complex, non-linear relationship between ICU workload and mortality. Unsupervised clustering methods are an important means of novel pattern identification and hypothesis generation. We hypothesized that distinct clusters of ICU workload factors with varying mortality rates would be observed.


Methods: In this 64-center observational study, 30,147 adult patients admitted to an ICU for greater than 24 hours, patient-level workload data about the ICU care team and patient characteristics (e.g., sequential organ failure assessment (SOFA) score, medication regimen complexity (MRC-ICU) score) were collected. Following principal component analysis (PCA), K-means clustering was applied. Uniform Manifold Approximation and Projection (UMAP) was applied to visualize the distinctness of the generated clusters. Clusters were stratified by mortality to explore the association of ICU workload characteristics with patient mortality.


Results: Five distinct clusters were identified. While all clusters had similar SOFA scores (range: 4.83-5.81), there were wide ranges in both mortality rates (12% to 31%) and a variety of workload factors. For example, Clusters 3 and 4 had equal mortality rates (14.0%) despite Cluster 4 having higher severity of illness as measured by SOFA and MRC-ICU scores. Yet, Cluster 4 had lower pharmacist-to-ICU-patient ratios (19.05 vs. 39.67) and higher rates of patient days with a pharmacist on rounds (10.92 vs. 7.75).


Conclusions: Mortality rates ranged widely given similar severity of illness, suggesting the potential of other causal factors for patient outcomes. The patterns between Cluster 3 and 4 suggest that improved workload factors may mitigate higher severity of illness and decrease mortality rates. While unsupervised methods are largely hypothesis-generating, these findings have important ramifications suggesting staffing optimization as a modifiable mortality risk factor and a source of quality improvement efforts in the ICU.