Using ICU Regimen Complexity to Forecast Pharmacist Workload

• Support Vector Machine (SVM) achieved the lowest mean absolute error (MAE 4.71) while the mean documented pharmacist interventions per ICU stay were 4.7, so prediction error approximates average workload.

• Random Forest had the lowest root mean square error (RMSE 9.26); all machine learning models and stepwise regression outperformed a parsimonious model using severity-of-illness plus medication-regimen complexity.

• Practical implication: model performance is sufficient for aggregate workload forecasting and staffing discussions but is not reliable for patient-level triage.

• ICU medication regimens are complex and high-risk; critical care pharmacist (CCP) interventions reduce preventable adverse drug events but pharmacist staffing is constrained and patient surges are unpredictable.

• Hospital pharmacy lacks a standardized, objective workload metric; site-specific, self-reported intervention logs ("i-Vents") produce heterogeneous data that undermines triage, staffing justification, and cross-institutional comparisons.

• A reproducible, automatable metric of medication regimen complexity (Medication Regimen Complexity—ICU, MRC-ICU) that correlates with documented interventions offers a practical signal for aggregate workload forecasting and prioritization, and can be paired with stewardship, clinical decision support, or resource-allocation planning.

• Retrospective cohort of 13,373 adult ICU patients at a single academic center (OHSU), including first ICU stays from June 1, 2020 to June 7, 2023; excluded stays <24 hours or early comfort care.

• Predictors from the first 24 hours (MRC-ICU, Sequential Organ Failure Assessment [SOFA], APACHE II, medication variables, mechanical ventilation/continuous renal replacement therapy status, and demographics) were used to predict total documented pharmacist interventions and an intervention-intensity score.

• Six prediction approaches were compared: a simple SOFA+MRC-ICU model, full and stepwise negative binomial regression, and three supervised machine learning models (Random Forest, Support Vector Machine, XGBoost).

• Missing data handled with multiple imputation (10 imputations); models trained on an 80/20 train/test split, ML hyperparameters tuned with 5-fold cross-validation; performance evaluated by RMSE, MAE, and sMAPE and interpreted with SHAP; subgroup and 95%-quantile analyses examined influence of outliers.

• Key design note: the primary outcome was self-reported pharmacist interventions logged as i-Vents in the local Epic EHR, so results reflect local documentation practice and completeness.

• Observed workload: mean documented interventions per ICU stay 4.7 (SD 7.1); mean intervention intensity 24.0 (SD 40.3); high-intervention subgroup mean 9.0 vs low 1.5.

• Model performance: Random Forest had lowest RMSE (9.26) and SVM lowest MAE (4.71); all ML models and stepwise regression outperformed the SOFA+MRC-ICU baseline (SOFA+MRC RMSE 10.19, MAE 5.47). 95%-quantile analyses reduced influence of extreme errors; SVM and XGBoost performed best across the central distribution.

• Predictors: multivariable regression found each 1-point increase in MRC-ICU associated with ~4% higher intervention rate (rate ratio 1.04, 95% CI 1.03–1.04); SOFA associated with ~2% per point (rate ratio 1.02). Continuous renal replacement therapy and mechanical ventilation were strong independent predictors (rate ratios ~1.91 and 1.75).

• Interpretation for practice: because MAE (~4.71) approximates the mean documented interventions (4.7), prediction error is on the order of average workload—useful for high-level staffing forecasts but not for reliable patient-level triage.

• Early medication-regimen complexity (MRC-ICU) is strongly associated with documented pharmacist intervention volume and can anchor an objective, automatable ICU workload signal to inform aggregate staffing and coverage discussions.

• Because the models were trained on self-reported documentation and retain large errors for individual patients, similar tools are better suited to department-level workload forecasting than to automated, patient-level prioritization lists.

• Operationally, treat model outputs as a rough 'workload weather' forecast to guide where pharmacists may focus effort; clinicians should continue to use clinical judgment for patient-level decisions.

Strengths:

• Large, curated ICU dataset with detailed medication, complexity (MRC-ICU), and severity (SOFA/APACHE II) data linked to granular documented pharmacist activity.

• Robust modeling pipeline: multiple imputation for missing data, negative binomial regression plus three supervised ML algorithms with cross-validation, pooled testing per Rubin’s rules, and SHAP-based interpretability to explain feature contributions.

Limitations:

• Single-center, retrospective design using locally self-reported i-Vent documentation with unknown adherence limits causal inference and external generalizability.

• Highly skewed distribution of documented interventions produced reduced accuracy for extreme high- and low-end outliers; real-time, patient-level triage performance remains uncertain because the MAE approximates mean workload.

• Outcome measures reflect documented activity (i-Vents) not an objective measure of true pharmacist cognitive workload or unmet intervention need, introducing likely under-reporting bias especially for busiest patients.

MRC-ICU and early severity data can predict documented pharmacist interventions well enough for aggregate workload forecasting and to inform staffing discussions, but model error and reliance on self-reported documentation preclude safe use for patient-level triage or cross-site deployment without standardized documentation and external validation.