DISCRET: Synthesizing Faithful Explanations For Treatment Effect Estimation

Area of Science:

• Artificial Intelligence
• Machine Learning
• Causal Inference

Background:

• Accurate and interpretable AI models are crucial for individual treatment effect estimation (ITE), especially in healthcare.
• Current methods struggle to balance accuracy with faithful explanations, with black-box models lacking interpretability and self-interpretable models sacrificing accuracy.

Purpose of the Study:

• To develop a novel AI framework, DISCRET, that generates accurate ITE predictions with inherently faithful, rule-based explanations.
• To enable explanations to function as database queries for identifying similar patient subgroups.

Main Methods:

• Proposed DISCRET, a self-interpretable framework for ITE.
• Developed a novel reinforcement learning (RL) algorithm for efficient synthesis of rule-based explanations from a large search space.
• Utilized explanations as database queries to identify relevant subgroups.

Main Results:

• DISCRET achieves accuracy comparable to state-of-the-art black-box models.
• Provides faithful, rule-based explanations for each prediction.
• Outperforms existing self-interpretable models on diverse datasets (tabular, image, text).

Conclusions:

• DISCRET offers a robust solution for accurate and interpretable ITE.
• Demonstrates the dual utility of explanations as both interpretable outputs and data querying tools.
• Presents a significant advancement in developing trustworthy AI for critical applications.