NEXIM: A Nash Equilibrium-Based Framework for Stable Explainable AI in Medical Applications

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning Explainability
• Medical Data Science

Background:

• Attribution-based explanations in medical AI can lack reliability due to model randomization and minor analytical changes.
• Ensuring trustworthy medical AI requires robust and stable explanations for model predictions.

Purpose of the Study:

• To introduce NEXIM (Nash Equilibrium-based Explainability and Interpretability Model), a novel framework for selecting AI models in medical applications.
• To enhance the reliability and reproducibility of AI-driven medical explanations by evaluating prediction error, explanation stability, and model connectivity.

Main Methods:

• NEXIM is an accuracy-constrained, equilibrium-inspired model-selection framework.
• It jointly evaluates prediction error (RMSE), explanation stability (Spearman rank correlation), and cross-model connectivity.
• Ten GradientBoostingRegressor models were evaluated using longitudinal Parkinson's Progression Markers Initiative data.

Main Results:

• NEXIM selected the RMSE-optimal model at one- and three-year prediction horizons.
• At the two-year horizon, NEXIM selected a model with improved stability (0.8847 vs. 0.8757) and connectivity (1.000 vs. 0.889) over the RMSE-optimal model, with a minimal RMSE increase.
• The selected models maintained similar top feature sets, indicating NEXIM acts as a reproducibility screen.

Conclusions:

• NEXIM enhances AI model selection in medicine by prioritizing explanation stability and reproducibility.
• It serves as a governance checkpoint for AI model refresh and documentation, but external validation and clinical evaluation are necessary.
• Stability and consensus are crucial reproducibility criteria for trustworthy medical AI, not direct measures of clinical utility.