A network-based approach to assess task-compliance in fMRI

Task-based functional MRI (fMRI) is widely used to map language networks and guide presurgical planning. However, the validity of activation maps depends critically on patients' engagement and task performance, which are rarely assessed directly during scanning. Failure to differentiate true atypical language organization from poor task execution may reduce the reliability of clinical fMRI. Complementary use of both task-positive activations (language network, LN) and task-negative activations (default mode network, DMN) may provide an objective means to approximate task compliance.

We retrospectively analysed task-based fMRI data from 43 patients with temporal lobe epilepsy (TLE) and 25 healthy controls. Each participant performed two language paradigms (sentence comprehension, verb generation). Engagement indices were defined based on suprathreshold activation within LN and DMN masks. For each task, participants were classified as fully compliant, partially compliant, or non-compliant. These indices were further combined into task specific compliance measures. Neuropsychological attention performance was compared across engagement categories to test for linear associations between functional engagement and cognitive capacity.

Healthy controls showed significantly higher compliance than TLE patients for the cognitively more demanding sentence comprehension task (p = 0.008), but not for verb generation. LN engagement proved to be the most robust predictor of attentional performance, with higher engagement associated with better outcomes in the d2 test, TMT-A, TMT-B, and the overall attention score (all p < 0.05). Crucially, the association with the overall attention score remained statistically significant after Bonferroni correction (p = 0.002). A significant linear trend was observed (p = 0.032), where attention scores decreased progressively from fully compliant to non-compliant groups. DMN engagement showed weaker, exploratory trends, suggesting it serves as a complementary "safety net" rather than a primary predictor.

Our findings demonstrate that network-based engagement indices derived from both LN activation and DMN suppression provide a practical approximation of task engagement during clinical fMRI. While LN engagement emerged as the strongest predictor of attention, DMN suppression might provide a qualitative safeguard against false negatives, helping to distinguish remediable "state" lapses in engagement from "trait" functional reorganization. Incorporating both task-positive and task-negative markers into clinical workflows may improve the reliability of language mapping, enabling real-time identification of unsuccessful runs and ultimately supporting presurgical planning.