Explanation Beyond Individual Features: Instance-wise Feature Grouping for EHR Predictive Analytics.

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Machine Learning

Background:

• Clinical prediction models require explainability for trust and personalization.
• Instance-wise feature selection (IWFS) offers personalized explainability but suffers from instability.
• Existing feature grouping methods improve stability but often reduce prediction accuracy.

Purpose of the Study:

• To introduce FlexGPC, a novel instance-wise feature grouping method for robust and stable feature selection.
• To enhance the accuracy and stability of clinical prediction models.
• To enable computational phenotyping through feature group identification.

Main Methods:

• Developed FlexGPC using neural networks to learn flexible representations and combinations of feature groups.
• Explored various feature group combination schemes.
• Conducted extensive experiments on real-world electronic health records (EHR) data.

Main Results:

• FlexGPC significantly outperformed state-of-the-art (SOTA) baselines in accuracy and feature selection stability.
• Demonstrated effectiveness in both mortality prediction and next-admission diagnosis prediction tasks.
• Showcased the ability of identified feature groups to serve as potential phenotypes for computational phenotyping.

Conclusions:

• FlexGPC provides a robust and stable approach to instance-wise feature selection for clinical prediction models.
• The method enhances model explainability and prediction accuracy while enabling computational phenotyping.
• FlexGPC represents a significant advancement in personalized explainability for healthcare AI.