Robust lesion network mapping reveals symptom-specific circuits

Lesion network mapping (LNM) is a powerful tool for linking distributed brain lesions to symptom-relevant networks, but similar LNM maps across unrelated symptoms have raised concerns about its statistical validity and specificity. Here we introduce robust LNM (rLNM), a statistically principled framework that combines two complementary controls: an anatomy-preserving spatial null model to account for connectome-derived hub bias shared across symptoms, and hypothesis-driven contrasts to resolve selective specificity among related symptoms. rLNM effectively controls both false positives and negatives. Across 10 lesion datasets (384 lesions) and two task-fMRI meta-analyses (339 experiments), rLNM recovers distinct symptom-specific networks (mean r = 0.044), which show significantly reduced similarity to the connectome hub pattern (r = 0.12 vs. 0.33 in conventional LNM). In 56 patients with neurosurgical lesions, rLNM predicts postoperative language decline at the single-patient level. rLNM thus resolves these concerns and lays a unified statistical foundation for network-symptom inferenc.