A Machine Learning Model to Predict Treatment Effect Associated with Targeted Temperature Management After Cardiac...

Area of Science:

• Intensive care medicine
• Neurology
• Machine learning in healthcare

Background:

• Targeted temperature management (TTM) improves neurological recovery in comatose cardiac arrest survivors.
• Predicting outcomes in the hyperacute phase post-cardiac arrest is crucial for patient management.
• Multimodal data analysis offers potential for early prognostication.

Purpose of the Study:

• To evaluate machine learning models for predicting short-term outcomes after TTM.
• To determine if hyperacute phase data can predict neurological recovery and survival.
• To identify key clinical and laboratory predictors of outcome.

Main Methods:

• Analysis of clinical, physiological, and laboratory data from the hyperacute phase (first 12 hours post-ICU admission).
• Application of three machine learning algorithms: generalized linear models, random forest, and gradient boosting.
• Tenfold cross-validation and resampling for model performance assessment.

Main Results:

• The generalized linear model demonstrated strong predictive performance (AUC 0.86 ± 0.04 for survival, 0.85 ± 0.03 for neurological outcome).
• Key predictors included lower serum chloride, higher serum pH, and increased neutrophil counts.
• Machine learning models accurately predicted survival and favorable neurological outcome.

Conclusions:

• Short-term outcomes in TTM patients post-cardiac arrest can be accurately predicted using early, routinely collected data.
• Hyperacute phase prediction using machine learning facilitates personalized decision-making in postcardiac arrest care.
• Further validation of these models could enhance clinical management strategies.