A Model-Agnostic Feature Attribution Approach to Magnetoencephalography Predictions Based on Shapley Value

Deep learning has greatly enhanced the predictive performance of magnetoencephalography (MEG) decoding. However, the lack of interpretability has become a major obstacle to the practical application of deep learning-based MEG decoding algorithms, which may lead to non-compliance with legal requirements and distrust among end-users. To address this issue, this article proposes a feature attribution approach, which can provide interpretative support for each individual MEG prediction for the first time. The approach first transforms a MEG sample into a feature set, then assigns contribution weights to each feature using modified Shapley values, which are optimized by filtering reference samples and generating antithetic sample pairs. Experimental results show that the Area Under the Deletion test Curve (AUDC) of the approach is as low as 0.005, which means a better attribution accuracy compared to typical computer vision algorithms. Visualization analysis reveals that the key features of the model decisions are consistent with neurophysiological theories. Based on these key features, the input signal can be compressed to one-sixteenth of its original size with only a 0.19% loss in classification performance. Another benefit of our approach is that it is model-agnostic, enabling its utilization for various decoding models and brain-computer interface (BCI) applications.