(349) Augmenting Mortality Prediction With Medication Data and Machine Learning Models

Introduction: Mortality prediction remains an important question in the intensive care unit (ICU). Medications are important factors in both beneficial outcomes and adverse drug events (ADEs). The purpose of this evaluation was to test the performance of different machine learning (ML) approaches using medication data to predict hospital mortality in critically ill patients.

Methods: This retrospective cohort study included a random sample of 5,869 adult ICU patients taken at two different time points to create a training (n=991) and validation cohort (n=4,878). A total of 27 variables [e.g., age, severity of illness, medication regimen complexity-intensive care unit (MRC-ICU)] potentially associated with mortality were assessed. Advanced logistic regression models using nature splines, smoothing splines, and local linearity were created. Supervised, classification-based ML models [e.g., Random Forest, Support Vector Machine (SVM), and XGBoost] were developed. Area under the receiver operating characteristic (AUROC), positive predictive value (PPV), and negative predictive value (NPV) were compared among models.

Results: The optimized traditional regression models had an AUROC of 0.84-0.86, and AUROCs were 0.83, 0.85, and 0.82 for Random Forest, SVM, and XGBoost ML models, respectively. MRC-ICU had a moderate level of feature importance in both XGBoost and Random Forest.

Conclusions: ML methods did not substantially improve performance relative to traditional regression. While medication data were not included as a significant predictor in the advanced regression models, MRC-ICU did show moderate feature importance in certain ML models.