Concept-Level Local Explanations of Kidney Transplant Survival Predictions by Black-Box ML Models

Abstract

Existing explainable AI (XAI) methods, when applied to kidney transplant outcome prediction models, typically provide feature importance as opposed to a clinical concept-level description of how a prediction was generated from the input data. In this paper, we propose a novel XAI framework that provides explanations at the clinical concept level. Our framework first generates local explanations in the form of feature-level decision paths—i.e. sequences of conditions, represented as feature-value pairs, leading to a prediction that explain the models’ decision-making process based on input features. These decision paths are then translated into higher-level clinical concepts relevant to kidney transplantation. We use large language models (LLMs) enhanced with nephrology-specific knowledge and authoritative clinical guidelines (e.g., KDIGO standards) to generate clinically actionable concepts from low-level input features. The concepts are cross-validated across multiple LLM instances to enhance the clinical validity of the feature-to-concept mappings, and a systematic mapping approach using propositional logic and threshold-based rules is employed to balance expressiveness and simplicity of the mappings. The approach represents a step forward in integrating advanced AI systems with real-world clinical practice.

Ähnliche Arbeiten