Abstract
Injury and disease affect neural processing and increase individual variations in patients when compared with healthy controls. Understanding this increased variability is critical for identifying the anatomical location of eloquent brain areas for pre-surgical planning. Here we show that precise and reliable language maps can be inferred in patient populations from resting scans of idle brain activity. We trained a predictive model on pairs of resting-state and task-evoked data and tested it to predict activation of unseen patients and healthy controls based on their resting-state data alone. A well-validated language task (category fluency) was used in acquiring the task-evoked fMRI data. Although patients showed greater variation in their actual language maps, our models successfully learned variations in both patient and control responses from the individual resting-connectivity features. Importantly, we further demonstrate that a model trained exclusively on the more-homogenous control group can be used to predict task activations in patients. These results are the first to show that resting connectivity robustly predicts individual differences in neural response in cases of pathological variability.