Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Critical Network Lesion-Symptom Mapping
Poster Session D, Thursday, October 1, 4:30 - 6:30 pm, Wangari Maathai
Grant M. Walker1, G. Lynn Kurteff2, Akbar Hussain1, Leonardo Bonilha3, Chris Rorden3, Julius Fridriksson3, Gregory Hickok1; 1University of California, Irvine, 2University of California San Francisco, 3University of South Carolina
Lesion-symptom mapping (LSM) is a powerful tool for linking brain damage to behavioral impairment, yet existing methods carry significant limitations. Mass-univariate approaches such as voxel-based LSM (VLSM) rely on null hypothesis statistical testing, treat each voxel independently, and are vulnerable to lesion-volume confounds and inflated false positive rates. Multivariate approaches like support vector regression LSM (SVR-LSM) address some concerns but introduce a high-dimensional feature space that is difficult to interpret, requires extensive hyperparameter tuning, and focuses on a single optimal predictive model rather than acknowledging uncertainty in complex neuroimaging data. To address these limitations, we introduce Critical Network Lesion-Symptom Mapping (CN-LSM), a neuroanatomical atlas-based method that uses low-dimensional regression models and out-of-sample prediction to identify brain regions critically associated with behavioral performance, while explicitly accommodating multiple plausible explanations for observed lesion-behavior relationships. CN-LSM operates through a two-step procedure. First, candidate regions from a pre-specified atlas are identified using permutation tests and bootstrap confidence intervals, selecting regions where lesion damage significantly and reliably predicts behavioral scores. Second, all possible combinations of candidate regions are evaluated via leave-one-out cross-validation, comparing each network model's prediction accuracy against a lesion-volume baseline. Models that significantly outperform this control and are statistically indistinguishable in predictive accuracy form an ensemble network. The method was validated using ground-truth simulation and real clinical data. Simulations used lesion data from 109 participants with chronic left hemisphere stroke from the Predicting Outcomes of Language Rehabilitation (POLAR) clinical trial, with behavioral scores simulated to reflect known relationships to target regions in the Inferior Frontal Gyrus and Posterior Middle Temporal Gyrus. Real data analyses examined Western Aphasia Battery (WAB) Fluency and Comprehension scores from the same cohort. CN-LSM was applied using three neuroanatomical atlases (AALCAT, AICHA, JHU, distributed with NiiStat software) and compared to SVR-LSM. In the no-noise simulation condition, CN-LSM outperformed SVR-LSM in ground-truth recovery (F1: AALCAT = .47, AICHA = .40, SVR = .31). In the realistic noise condition, ensemble models from CN-LSM outperformed SVR-LSM in at least one model across 8–10 of 10 simulated datasets, and the model yielding optimal prediction accuracy was never the model that best recovered the ground truth. In the real data experiment, CN-LSM identified an expected anterior-posterior dissociation between fluency and comprehension networks, with consistent overlap in posterior insula, arcuate fasciculus, and superior longitudinal fasciculus across all three atlases. SVR-LSM produced more diffuse, less interpretable maps with minimal overlap between the two behavioral measures. CN-LSM offers a theoretically grounded, methodologically transparent approach to LSM that reduces researcher degrees of freedom, minimizes false positives, and provides holistic characterization of uncertainty across the statistical map. By analyzing an ensemble of plausible models rather than a single optimal solution, CN-LSM reveals converging evidence that is more robust to noise and analytical variability. These results support CN-LSM as a complementary technique to existing approaches, with particular value for advancing reproducible and interpretable research on the neural substrates of language.
Topic Areas: Methods, Disorders: Acquired