DECAF: An interpretable deep cascading framework for ICU mortality prediction.

Area of Science:

• Medical Informatics
• Computational Biology
• Artificial Intelligence in Medicine

Background:

• Medical risk detection in Intensive Care Units (ICUs) is critical for improving patient outcomes.
• Existing deep learning and bio-statistical models for patient mortality prediction lack interpretability.
• Understanding the reasoning behind predictions is essential for clinical trust and application.

Purpose of the Study:

• To develop an interpretable framework for dynamic patient risk prediction in ICUs.
• To introduce the DEep CAscading Framework (DECAF) for simulating physiological deterioration.
• To enable multi-task prediction of risks across all physiological functions at each clinical stage.

Main Methods:

• Utilized cascading theory to model the physiological domino effect.
• Developed the DEep CAscading Framework (DECAF) for dynamic patient condition simulation.
• Trained and evaluated the model on the MIMIC-III dataset comprising 21,828 ICU patients.

Main Results:

• DECAF demonstrated high performance in mortality prediction, achieving an AUROC of 89.30%.
• The framework provides interpretable predictions, unlike traditional feature-based or score-based models.
• DECAF showed applicability for multi-prediction tasks and learning from clinical knowledge.

Conclusions:

• The DEep CAscading Framework (DECAF) offers a novel, interpretable approach to medical risk detection in ICUs.
• DECAF effectively simulates patient deterioration and surpasses existing methods in predictive accuracy.
• The framework's interpretability and multi-task capability enhance its clinical utility and potential for integrating medical common sense.