Supervised Machine Learning Evaluation of ICU Pharmacist-to-Patient Ratio: An OPTIM Analysis

Introduction: High workload in the ICU is associated with reduced patient safety. Unlike physicians and nurses, critical care pharmacists (CCPs) do not have an established safe CCP-to-patient ratio. Given that CCPs are essential members of the interprofessional team, this unknown ratio is a critical gap. The purpose of this multi-center study was to generate a rich, externally generalizable dataset to support robust machine learning (ML) analyses capable of capturing complex, non-linear relationships between workload and mortality. The hypothesis was that CCP-to-patient ratio is an important feature for ML performance.


Methods: CCPs prospectively collected data on CCP workload and ICU patients for 100 consecutive days, then retrospectively collected patient characteristics (e.g., sequential organ failure assessment (SOFA) score, medication regimen complexity (MRC-ICU) score) and outcomes. Adult patients admitted to an ICU without restrictions to care for at least 24 hours were included. A Random Forest (RF) supervised ML model was developed for the primary outcome (mortality). Model performance was measured by area under the receiver operating characteristic (AUROC). A feature importance graph was plotted to characterize the independent variables that were most important for model performance. Partial dependence plots (PDPs) were plotted to determine the directionality of relationships between independent variables and mortality.


Results: A total of 213 CCPs from 64 sites collected data on 25,069 ICU patients. A RF model using 39 variables related to CCP workload, ICU team composition, and patient characteristics was developed (AUROC 0.79). Following SOFA and MRC-ICU score, the top 10 most important features were CCP workload measures, including CCP-to-patient ratio. The PDP graph for CCP-to-patient ratio revealed the lowest mortality for a ratio between 1:13 and 1:15, with a near-linear increase in mortality for ratios exceeding 1:15.


Conclusions: OPTIM represents the largest evaluation of CCP workload in the context of the ICU team. Results indicate that CCP workload is an important feature for ML performance, with CPP-to-patient ratios greater than 1:15 conferring the greatest mortality risk. Results suggest that quality improvement focused on optimizing CCP-to-patient ratio may improve patient outcomes.