Interpretable machine learning model for imaging-based outcome prediction after cardiac arrest.

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Neurology and critical care

Background:

• Early identification of brain injury post-cardiac arrest is vital for prognostication.
• Lack of interpretability in current machine learning models hinders clinical adoption.

Purpose of the Study:

• To develop an interpretable machine learning method for identifying CT imaging patterns linked to cardiac arrest prognostication.
• To predict patient survival and awakening status using these interpretable imaging patterns.

Main Methods:

• Retrospective analysis of 1284 adult patients undergoing brain CT within 24 hours of cardiac arrest.
• Decomposition of CT images into subspaces to identify interpretable injury patterns.
• Development of machine learning models to predict outcomes (survival, awakening) using identified patterns.

Main Results:

• Machine learning models achieved an AUC of 0.710 for predicting survival and 0.702 for predicting awakening.
• Expert physicians confirmed the clinical relevance of identified imaging patterns.
• 35% of patients awakened from coma, and 34% survived hospital discharge.

Conclusions:

• An interpretable method was developed to identify early brain injury patterns on CT scans.
• These identified imaging patterns are predictive of patient outcomes after cardiac arrest.
• The approach enhances the trustworthiness and potential clinical translation of machine learning in prognostication.