Explainable Model Differencing for MEG Decoding via Predict-Probability Differences and Boundary-Optimized Rules.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks advance magnetoencephalography (MEG) decoding.
• Explainable AI interprets individual models, but lacks methods for comparing model decision logic (model differencing).
• Existing model differencing methods suffer from low accuracy and poor localization of decision boundaries.

Purpose of the Study:

• To develop a novel, accurate, and interpretable model differencing approach for MEG.
• To address limitations in current methods, particularly in high-dimensional, low-sample data.
• To facilitate model selection, optimization, and error analysis in MEG decoding.

Main Methods:

• Proposed boundary-optimized rules via predict-probability difference (BO-RPPD), a rule-based model differencing technique.
• Introduced a novel measurement using predict-probability differences between models for direct rule learning.
• Integrated counterfactual generation and feature reduction to optimize decision boundaries in high-dimensional MEG data.

Main Results:

• BO-RPPD significantly outperformed benchmarks, improving F1-score by up to 24%.
• The approach demonstrated effective coverage across broader samples and generated an optimal number of rules for explainability.
• Achieved practical value in error pattern analysis and decision fusion.

Conclusions:

• BO-RPPD offers a superior, model-agnostic approach for MEG model differencing, enhancing interpretability and performance.
• The method generalizes effectively to electroencephalography (EEG) and other structured datasets.
• This work bridges the gap in understanding differences between AI models in neuroscience applications.