Explainable machine learning for predicting ICU mortality in myocardial infarction patients using pseudo-dynamic data

Area of Science:

• Cardiology
• Artificial Intelligence
• Health Informatics

Background:

• Myocardial infarction (MI) is a leading cause of mortality, with intensive care unit (ICU) patients facing elevated risks.
• Accurate and timely mortality prediction in the ICU is crucial for effective patient management.

Purpose of the Study:

• To develop an explainable, pseudo-dynamic machine learning framework for predicting mortality in ICU patients with MI.
• To evaluate the performance of novel tabular deep learning approaches against standard machine learning algorithms.

Main Methods:

• Utilized two retrospective ICU cohorts (eICU and MIMIC-IV) for developing and validating a machine learning framework (XMI-ICU).
• Employed an integrated XGBoost model within an EHR time-series extraction framework.
• Leveraged time-resolved Shapley values for interpretability and feature identification.

Main Results:

• XMI-ICU achieved an AUROC of 92.0 and a balanced accuracy of 82.3 for 6-hour mortality prediction.
• Demonstrated reliable predictive performance across various prediction horizons (6-24 hours) and successful external validation on the MIMIC-IV cohort.
• Outperformed the standard APACHE IV system in clinical risk analysis.

Conclusions:

• The XMI-ICU framework effectively predicts mortality in ICU patients with MI using time-series physiological data.
• The explainable nature of the framework provides valuable clinical insights, enhancing patient care strategies.
• This approach offers a significant advancement in leveraging EHR data for critical care prognostication.