Machine Learning-Based Prediction of Prolonged Duration of Mechanical Ventilation Using Medication Data.

Area of Science:

• Critical Care Medicine
• Health Informatics
• Pharmacology

Background:

• Prediction algorithms for prolonged mechanical ventilation (PMV) in intensive care units (ICUs) often overlook detailed medication data.
• Medications significantly influence patient outcomes and PMV risk.

Purpose of the Study:

• To develop and validate PMV prediction models.
• To assess the contribution of medication-related variables alongside traditional physiologic predictors.

Main Methods:

• Retrospective cohort study using data from 318 ICU patients receiving mechanical ventilation.
• Trained logistic regression and machine learning (ML) models (XGBoost, Random Forest, SVM) on 30 variables including demographics, clinical data, labs, and medications.
• Validated models in two external datasets.

Main Results:

• The best logistic regression model achieved an AUROC of 0.75 with added medication regimen complexity and severity of illness data.
• Random Forest and SVM ML models achieved AUROCs of 0.78.
• Explainability analyses identified medication metrics, including the MRC-ICU score, as key predictors. Medication data offered modest improvements in discrimination and negative predictive value.

Conclusions:

• Medication-related variables provide incremental value in predicting PMV.
• Machine learning methods offer marginal improvements over regression models for PMV prediction.
• Medication data holds potential for improving patient outcome prediction, though simpler models may suffice.