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

Area of Science:

• Critical Care Medicine
• Health Informatics
• Machine Learning in Healthcare

Background:

• Traditional regression models show limited improvement in predicting ICU adult mortality with medication regimen complexity (MRC) data.
• Machine learning (ML) presents a potential avenue for enhancing mortality prediction accuracy.

Purpose of the Study:

• To compare ML approaches with traditional and advanced regression methods for predicting hospital mortality in ICU adults.
• To evaluate the impact of incorporating MRC data into various prediction models.

Main Methods:

• Supervised classification ML models (Random Forest, SVM, XGBoost) were developed using baseline and 24-hour ICU variables, including MRC-ICU.
• Traditional and advanced regression models were optimized using stepwise selection.
• Model performance was assessed using Area Under the Receiver Operating Characteristic (AUROC) curves.

Main Results:

• ML models achieved AUROCs between 0.82-0.85; advanced regression models yielded AUROCs of 0.84-0.86.
• Traditional regression models showed AUROCs ranging from 0.72-0.86.
• MRC-ICU data had moderate feature importance in XGBoost and Random Forest models.
• Model performance decreased in external validation cohorts.

Conclusions:

• ML and advanced regression methods did not significantly improve hospital mortality prediction compared to traditional methods, despite the inclusion of MRC data.
• MRC data contributes moderately to prediction in specific ML models.
• Further research may be needed to optimize ML for ICU mortality prediction.