Explainable Machine Learning Approach to Predict and Explain the Relationship between Task-based fMRI and Individual Differences in Cognition

Despite decades of costly research, we still cannot accurately predict individual differences in cognition from task-based fMRI. Moreover, aiming for methods with higher prediction is not sufficient. To understand brain-cognition relationships, we need to explain how these methods draw brain information to make the prediction. Here we applied an explainable machine-learning (ML) framework to predict cognition from task-based fMRI during the n-back working-memory task, using data from the Adolescent Brain Cognitive Development (n=3,989). We compared nine predictive algorithms in their ability to predict 12 cognitive abilities. We found better out-of-sample prediction from ML algorithms over the mass-univariate and OLS multiple regression. Among ML algorithms, Elastic Net, a linear and additive algorithm, performed either similar to or better than non-linear and interactive algorithms. We explained how these algorithms drew information, using SHapley Additive explanation, eNetXplorer, Accumulated Local Effects and Friedman’s H-statistic. These explainers demonstrated benefits of ML over the OLS multiple regression. For example, ML provided some consistency in variable importance with a previous study (Sripada et al. 2020) and consistency with the mass-univariate approach in the directionality of brain-cognition relationships at different regions. Accordingly, our explainable-ML framework predicted cognition from task-based fMRI with boosted prediction and explainability over standard methodologies.