Augmenting Mortality Prediction in Critically Ill Adults With Medication Data and Machine Learning Models.

BACKGROUND: Mortality prediction in ICU adults is only marginally improved when medication regimen complexity (MRC) data is incorporated into traditional regression models. Machine learning (ML) may improve this prediction. OBJECTIVE: To compare the performance of different ML approaches incorporating MRC data to both traditional and advanced regression approaches, with and without MRC data, to predict hospital mortality in ICU adults. DERIVATION COHORT: Nine hundred ninety-one ICU adults at the University of North Carolina (UNC) Health System. VALIDATION COHORT: A temporally distinct cohort of 4,878 ICU adults at UNC and an external cohort of 12,290 ICU adults at the Oregon Health and Science University. PREDICTION MODEL: Supervised, classification-based ML models (e.g., Random Forest, Support Vector Machine [SVM], and XGBoost) were developed. Twenty-seven variables at ICU baseline (age, sex, service, diagnosis) and 24 hours (illness severity, supportive care use, fluid balance, laboratory values, MRC-ICU, vasopressor use) associated with mortality, and 14 missingness indicator variables, were included in each ML model. Traditional and advanced (equipped with linear predictors, predictors in nature cubic splines, predictors in smoothing cubic splines, and local linear predictors) regression models were optimized using stepwise selection by Bayesian Information Criterion. Area under the receiver operating characteristic (AUROC) was compared among models. RESULTS: Random Forest, SVM, and XGBoost achieved AUROCs of 0.83, 0.85, and 0.82, respectively, on the test set. Traditional regression models based on Sequential Organ Failure Assessment, Acute Physiology and Chronic Health Evaluation (APACHE) II, MRC-ICU + Sequential Organ Failure Assessment + APACHE II with and without an interaction term, and a full model including all 27 available variables demonstrated AUROCs of 0.81, 0.72, 0.82, 0.83, and 0.86, respectively. Advanced regression models yielded AUROCs of 0.85, 0.86, 0.85, and 0.84, respectively. The MRC-ICU exhibited a moderate level of feature importance in both XGBoost and Random Forest models. Models demonstrated lower performance in the validation cohorts. CONCLUSIONS: Use of ML, compared with traditional and advanced regression methods, did not improve hospital mortality prediction despite medication data inclusion. The MRC-ICU demonstrates moderate feature importance in select ML models.

Duke Scholars

Author Rishi Kamaleswaran Trauma, Acute, and Critical Care Surgery