Abstract
Resting-state fMRI has shown the ability to predict task activation on an individual basis by using a general linear model (GLM) to map resting-state network features to activation z-scores. The question remains whether the relatively simplistic GLM is the best approach to accomplish this prediction. In this study, several regression-based machine-learning approaches were compared, including GLMs, feed-forward neural networks, and random forest bootstrap aggregation (bagging). Resting-state and task data from 350 Human Connectome Project subjects were analyzed. First, the effect of the number of training subjects on the prediction accuracy was evaluated. In addition, the prediction accuracy and Dice coefficient were compared across models. Prediction accuracy increased with the training number up to 200 subjects; however, an elbow in the prediction curve occurred around 30-40 training subjects. All models performed well with correlation matrices, which displayed correlation between actual and predicted task activation for all subjects, exhibiting a strong diagonal trend for all tasks. Overall, the neural network and random forest bagging techniques outperformed the GLM. These approaches, however, require additional computing power and processing time. These results show that, while the GLM performs well, resting-state fMRI prediction of task activation could benefit from more complex machine learning approaches.