In 228 stroke patients, lesion network maps from a normative connectome predicted aphasia no better than a distance-only null connectome, and their mean map matched that of random lesions

Question: lesion network mapping (LNM) projects a lesion onto a normative functional connectome. A 2026 critique showed that LNM maps of different lesion sets converge on one map; defenders said that specificity tests recover symptom networks. We tested, out of sample and in one cohort with a graded deficit, whether the network content of the connectome predicts aphasia better than lesion size, lesion location, and a null connectome that keeps only the distance decay of functional connectivity. Approach: 228 left-hemisphere stroke lesions with WAB-R scores (Aphasia Recovery Cohort, registered masks from sci:sha256:d67c3e3fb31dc3ba3369dbb957ba24e7b0d71ad8ec67654326b0a30070ea8485) and a normative connectome from 138 healthy adults (AOMIC PIOP2 resting fMRI, two independent halves). Outcomes, covariates and two-stage ridge cross-validation (10-fold, 100 repeats, subject bootstrap) were the same as in the earlier structural-disconnection study of these patients. The plan was frozen before analysis. Findings: without global signal regression, the mean LNM map of the ARC lesions and of the same shapes at random positions correlated at r = 0.95; a median 74% of each map was on one generic map (global connectivity), whose loading followed lesion volume and carried no aphasia information beyond volume (partial r = 0.02). With GSR the convergence was weaker (r = 0.74). LNM features added +0.003 to cross-validated R2 for aphasia severity beyond volume (baseline 0.366); the distance-only null added +0.037; real minus null was -0.034 (95% interval -0.062 to -0.006), with the same sign in both connectome halves, without GSR and with voxel features. For fluency type real and null were equal; for conduction versus anomic type LNM was below plain parcel lesion load. A planted network outcome gave real minus null +0.037 to +0.042 (positive in 95-100% of draws), so the test had power. The voxel-wise LNM-severity map had 7 significant voxels and split-half r 0.17, while the network-free null gave 2,188 significant voxels and r 0.55. Normative tract disconnection added +0.086 R2 and LNM added nothing to it. LNM maps were reliable between connectome halves (median r 0.94). Limitations: one syndrome in one arterial territory, where location is informative; LNM was designed to join lesions in distant places, which we did not test; a 138-subject connectome of young adults; AQ and WAB type only.

lesion network mapping · normative connectome · functional connectivity · stroke · aphasia · null model · cross-validation · degeneracy · Aphasia Recovery Cohort · AOMIC · structure-function · global signal regression