Early Prediction of Cardiac Arrest in the Intensive Care Unit Using Explainable Machine Learning: Retrospective...

Area of Science:

• Critical Care Medicine
• Health Informatics
• Machine Learning in Healthcare

Background:

• Cardiac arrest (CA) is a significant cause of mortality in intensive care units (ICUs).
• Existing CA prediction models often struggle with generalization and validation across diverse patient populations and ICU subtypes.
• Patient heterogeneity presents a challenge for accurate and timely CA prediction.

Purpose of the Study:

• To develop a clinically interpretable ensemble approach for predicting CA within 24 hours.
• To ensure the model's accuracy and generalizability across varied patient populations and ICU subtypes.
• To provide interpretable results for real-time clinical adoption and patient-independent evaluation.

Main Methods:

• Retrospective analysis of data from MIMIC-IV and eICU-CRD databases.
• Feature extraction using vital signs, multiresolution statistical analysis, and Gini index within a 12-hour window.
• Development of a TabNet model with feature screening and cost-sensitive learning, validated via cross-validation and cross-dataset methods.

Main Results:

• The proposed ensemble method demonstrated superior performance over conventional approaches across different populations and ICU subtypes.
• The model achieved higher accuracy than baseline models in both MIMIC-IV and eICU-CRD datasets.
• External validation confirmed the model's strong generalization ability, outperforming baseline models.

Conclusions:

• The novel framework offers stable predictive power for CA across diverse ICU environments.
• Interpretable results highlight statistical differences between CA and non-CA groups, aiding clinical decision-making.
• The validated CA prediction system supports early clinical intervention and has potential for digital health clinical trials.