CRISP: A causal relationships-guided deep learning framework for advanced ICU mortality prediction.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Biomedical Data Science

Background:

• Mortality prediction in intensive care units (ICUs) is crucial for clinical decision-making.
• Deep learning (DL) models show promise but often lack generalizability and struggle with imbalanced electronic health record (EHR) data.
• Existing methods face challenges in stable and reliable mortality prediction.

Purpose of the Study:

• To develop a causally-informed deep learning model for enhanced patient mortality prediction.
• To improve the generalizability and stability of mortality prediction models in ICUs.
• To address the class imbalance problem in EHR data using causal relationships.

Main Methods:

• Introduction of the CRISP (Causal Relationship Informed Superior Prediction) model.
• Leveraging native counterfactuals for minority class augmentation.
• Constructing patient representations using causal structures for improved prediction.

Main Results:

• The CRISP model was evaluated on 69,190 ICU cases from MIMIC-III, MIMIC-IV, and WCHSU datasets.
• Achieved high performance in mortality prediction with AUROC ranging from 0.9042-0.9480 and AUPRC from 0.4771-0.7611.
• CRISP's data augmentation module demonstrated comparable performance to traditional oversampling techniques.

Conclusions:

• CRISP demonstrates superior generalizability across diverse patient cohorts compared to baseline algorithms.
• The model enhances the practical applicability of deep learning in clinical decision support systems.
• Causal inference integration improves the robustness of mortality prediction models.