Interpreting predictions of cognition from simulated versus empirical resting state functional connectivity

The relation between structure and function of the brain, and how behavior arises from it, is a central topic of interest in neuroscience. This problem can be formulated in terms of Structural Connectivity (SC) and Functional Connectivity (FC), respectively representing anatomical connections and functional interactions between regions in the brain. Recently, a study by Sarwar and colleagues has demonstrated individualized prediction of FC from SC using machine learning, additionally showing that variation in cognitive performance is explained by simulated FC (sFC) almost as well as by empirical FC (eFC). We investigated how decisions made to predict cognition differ between the models based on eFC and sFC. We predicted cognitive performance with Lasso regression in 100 cross-validation loops from both eFC and sFC separately, using FC between each of the 2278 pairs of regions in the 68-region Desikan-Killiany parcellation as features. We identified relevant predictors of cognition by inspecting permutation importance scores and keeping only features whose importance scores were consistently high across validation loops. 13 eFC features and 21 sFC features survived this procedure. Of these, only one feature overlapped between eFC and sFC. Analyzing overlap between regions corresponding to important features and functional systems known to support cognition revealed no patterns for either eFC or sFC features. In conclusion, we found that while cognition can be predicted from sFC almost as well as from eFC, different features are used in the models, and these features were not found to follow any structure in terms of functional systems. This shows that while machine learning models provide a theoretical upper bound on how accurately function can be predicted from structure, they do not necessarily produce output that can be interpreted in the same way as the data the models were trained on.