Knee-xRAI: An Explainable AI Framework for Automatic Kellgren-Lawrence Grading of Knee Osteoarthritis

Explainability improves clinical trust: The Knee-xRAI framework breaks down osteoarthritis grading into interpretable components (joint space narrowing, osteophytes, and sclerosis) rather than presenting a "black box" diagnosis. SHAP analysis revealed that joint space narrowing is the dominant factor in grading decisions, with osteophytes adding consistent value—transparency that helps clinicians understand why the AI made its recommendation.

High-stakes medical decisions require accountability: The paper emphasizes that even a single-grade disagreement in osteoarthritis severity can fundamentally change patient treatment (from conservative therapy to surgery or injections). This real-world consequence justifies the investment in explainable AI, as clinicians need to verify the model's reasoning before trusting it with such consequential decisions.

Modular AI design enables auditing: The pipeline's two-path approach—Path A for audit/explanation using XGBoost-SHAP and Path B for deployment using ConvNeXt—demonstrates how explainability and performance can coexist. The audit path achieved a quadratic weighted kappa of 0.6294, confirming that the structured feature vector contains genuine diagnostic signal that can be independently verified.

AI can outperform human consistency: The framework achieved a quadratic weighted kappa of 0.8436 on deployment, addressing the paper's opening problem that radiologist grading is "poorly reproducible across readers." This suggests explainable AI systems could reduce harmful variability in clinical decision-making when properly validated.

Feature importance reveals model limitations: SHAP analysis showed sclerosis contributes only marginally to grading decisions, suggesting either the feature is genuinely less clinically important or the model hasn't learned to use it effectively—a distinction that explainability makes visible