Post-Cardiac arrest outcome prediction using machine learning: A systematic review and meta-analysis

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Prognostication in post-cardiac arrest patients is challenging.
• Machine learning (ML) and deep learning (DL) models offer potential for improved predictions.
• This review assesses the effectiveness of AI in predicting clinical outcomes using structured data.

Purpose of the Study:

• To systematically review and meta-analyze the effectiveness of ML and DL models in predicting clinical outcomes after cardiac arrest.
• To evaluate prediction accuracy at different time points.
• To identify common predictive features and assess model bias.

Main Methods:

• Systematic review and meta-analysis following PRISMA guidelines.
• Searched PubMed, Scopus, and Web of Science databases until March 2024.
• Included studies predicting return of spontaneous circulation (ROSC), survival, or neurological outcomes using ML/DL on structured data.
• Extracted data using CHARMS checklist and assessed bias with PROBAST tool.
• Performed quantitative synthesis and meta-analysis of models reporting AUC with 95% confidence intervals.

Main Results:

• Included 41 studies with 97 ML and 16 DL models.
• Pooled AUC for favorable neurological outcomes at discharge: 0.871 (ML) and 0.877 (DL).
• Pooled AUC for survival prediction: 0.837.
• High heterogeneity and risk of bias observed, mainly due to missing data handling and lack of calibration plots.
• Commonly used features included age, sex, and initial arrest rhythm.

Conclusions:

• AI-based predictive models (ML/DL) are more effective than traditional regression algorithms.
• Significant heterogeneity and high risk of bias limit current model dependability.
• Further research on state-of-the-art DL models for tabular data and their clinical generalizability is needed to enhance outcome prediction.