How much information is needed to predict outcomes after cardiac arrest?

Area of Science:

• Medical Informatics
• Neurology
• Critical Care Medicine

Background:

• Prognostication following cardiac arrest is complex, with potential for improvement using machine learning (ML).
• Assembling heterogeneous data from sources like registries, EHR, and EEG presents practical challenges for ML model development.
• This study investigates whether a subset of data sources is sufficient for accurate ML-based prognostication.

Purpose of the Study:

• To compare the performance of ML models using various combinations of registry, EHR, and EEG data for predicting poor outcomes after cardiac arrest.
• To evaluate sequential ML models at different time points post-arrest (presentation, 12, 24, 48, 72 hours) using available data.
• To determine the optimal data source combination for maximizing sensitivity at perfect specificity in predicting poor outcomes.

Main Methods:

• A cohort study included 1,106 adult patients who were unresponsive after cardiac arrest between January 2010 and February 2022.
• Machine learning models, primarily random forests, were developed to predict poor outcome (Cerebral Performance Category 4 or 5).
• Models utilized combinations of registry, EHR, and EEG data, assessed sequentially at multiple time points post-arrest with sensitivity at perfect specificity as the primary metric.

Main Results:

• The best performing ML models integrated both registry and EEG data across all time points.
• After 12 hours, models combining registry, EHR, and EEG data demonstrated superior performance.
• Peak median sensitivity at perfect specificity was 70% at 24 hours, with excellent discrimination (AUC 0.949).

Conclusions:

• Optimal sensitivity for predicting poor outcomes after cardiac arrest necessitates the integration of multiple data sources.
• Current single-source or limited-combination approaches are insufficient for robust prognostication.
• Development of large, comprehensive, multicenter datasets is crucial for advancing post-arrest prognostication using ML.