Identification of Significant Instants of Voxels for Cognitive State Classification Using Interpretable Machine Learning Models

Despite decades of significant research, task-based functional MRI cannot reliably predict individual differences in cognition. Furthermore, searching for methods with greater predictability alone is insufficient. We need to clarify how these techniques use brain input to create predictions in order to comprehend the links between cognition and the brain. In this study, we have applied the Interpretable Machine Learning (IML) framework to decode cognition from fMRI data and find the significant instants of the voxel time course. We compared the ability of three predictive models to decode cognitive states. The predictive IML models considered in the current study include an explainable boosting machine (EBM), a decision tree (DT) classifier, and linear regression (LR). Furthermore, the classification accuracy of Support Vector Machine (SVM) and Gaussian Naïve Bayes (GNB) classifiers is reported for cognitive state classification. The standard Star plus fMRI dataset with two cognitive tasks has been used in this study. Initially, a few voxels are selected using a clustering-based maximum margin feature engineering framework. Then, the IML models are built with selected voxels from fMRI data. The classification accuracy of 80%, 82%, 80%, 93.7%, and 82% is achieved using EBM, DT, LR, SVM, and GNB classifiers, respectively. Moreover, the IML classifiers EBM, LDT, and LR can identify the significant instants of voxels.