Red Flag/Blue Flag visualization of a common CNN for text classification

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Medical Informatics

Background:

• TextCNN models are widely used for clinical text classification.
• Assessing model interpretability, especially for smaller datasets, remains a challenge.
• Overfitting can lead to misclassification by activating irrelevant features.

Purpose of the Study:

• Introduce Red Flag/Blue Flag (RFBF), a novel explainable AI (X-AI) software for TextCNN binary classification.
• Visualize convolutional filter discriminative capabilities for enhanced model interpretability.
• Provide tools for model diagnosis, feature verification, and overfit prevention.

Main Methods:

• Developed RFBF software for TextCNN models.
• Visualized convolutional filter discriminative capabilities.
• Conducted experiments on physician-authored operative notes for surgical misadventure prediction.

Main Results:

• RFBF offers a more informative approach than direct logit contribution assessment.
• Demonstrated filter consistency assessment, predictive performance improvement, and information leakage estimation.
• Utilized RFBF to analyze TextCNN performance in predicting surgical misadventure outcomes.

Conclusions:

• RFBF enhances the interpretability of TextCNN models in clinical text analysis.
• The X-AI approach aids in diagnosing, verifying, and preventing overfitting in medical text classification.
• This tool can benefit clinical text research and contribute to improved patient outcomes.