Transparent Machine Learning Algorithms for Explainable AI on Motor fMRI Data

Abstract

With the emergence of explainable artificial intelligence (xAI), two main approaches for tackling model explainability have been put forward. Firstly, the use of inherently simple and transparent models that with easily understandable inner-workings (interpretability) and can readily provide useful knowledge about the model’s decision making process (explainability). The second approach is the development of interpretation and explanation algorithms that may shed light upon black-box models. This is particularly interesting to apply on fMRI data as either approach can provide pertinent information about the brain’s underlying processes. This study aims to explore the capability of transparent machine learning algorithms to correctly classify motor fMRI data, if more complex models inherently lead to a better prediction of the motor stimulus, and the capability of the Integrated Gradients method to explain a fully connected artificial neural network (FCANN) used to model motor fMRI data. The transparent machine learning models tested are Linear Regression, Logistic Regression, Naive Bayes, K-Neighbors, Support Vector Machine, and Decision Tree, while the Integrated Gradients method is tested on a FCANN with 3 hidden layers. It is concluded that the transparent models may accurately classify the motor fMRI data, with accuracies ranging from 66.75% to 85.0%. The best transparent model, multinomial logistic regression, outperformed the most complex model, FCANN. Lastly, it is possible to extract pertinent information about the underlying brain processes via the Integrated Gradients method applied to the FCANN by analyzing the spatial expression of the most relevant Independent Components for the FCANN’s decisions.