Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks
Critical-Region Lesion-Symptom Mapping
Poster C112 in Poster Session C, Wednesday, October 25, 10:15 am - 12:00 pm CEST, Espace Vieux-Port
Grant Walker1, Gregory Hickok1, Julius Fridriksson2; 1University of California, Irvine, 2University of South Carolina
Lesion-symptom mapping (LSM) continues to be an important tool for inferring the necessity of brain areas for healthy behavior and cognitive processes. LSM has seen rapid development of methodological tools for making structure-function inferences, moving from raw lesion overlap maps to mass univariate statistical tests to multivariate models of the relationship between brain damage and functional consequences. The introduction of machine learning tools like support vector regression (SVR-LSM; Zhang et al., 2014) has started to move the field toward a more predictive approach. We continued this trend by developing the critical-region lesion-symptom mapping (CRLSM) approach. CRLSM begins with a mass univariate region-of-interest analysis that uses permutation and bootstrap tests to identify candidate regions of an atlas (AICHA), followed by “brute force” combinatorial cross-validation to identify a set of regions that are linearly related to behavioral performance. Essentially, we divide the lesion volume into two parts: critical and non-critical volume. We compared the new approach against the SVR-LSM approach using real lesion data and both simulated and real behavioral data. For simulated data, we used the JHU atlas to form a ground truth based on 81 lesion masks, where a frontal region (15) and a posterior temporal region (186) were the target neural substrates. Behavioral scores were simulated by calculating the proportion of damage to the critical region and applying Gaussian noise. We attempted to recover the neural regions using the CRLSM and SVR-LSM approaches. We compared Dice overlap statistics. For real data, we compared maps of the WAB Fluency and WAB Comprehension subscores collected from 91 participants with left hemisphere stroke. We expected fluency to be associated with frontal regions and comprehension to be associated with posterior temporal regions. In the simulation, the target network included 25,904 voxels. CRLSM identified 26,120 voxels, recovering 12,817 target voxels, with approximately equal coverage of both regions (Dice overlap = 49.3%). SVR-LSM identified 16,238 voxels, recovering 9,756 target voxels, mostly concentrated in the posterior region (Dice overlap = 46.3%). While the CRLSM approach was less conservative, it paid off by identifying more of the non-contiguous network. Adjusting the SVR-LSM threshold to identify 26,120 voxels led to better results (Dice overlap = 53.3%) but required (typically unavailable) knowledge of the critical network size. In the real data, fluency was predicted by inferior precentral gyrus damage and comprehension was predicted by middle and superior temporal damage. The Fluency map accounted for 32% of the variance in Fluency scores (average prediction error = .98 on a 1-10 scale); the Comprehension map accounted for 43% of the variance in Comprehension scores (average prediction error = 1.75 on a 1-10 scale). More variance was explained by lesion volume in the SVR-LSM map (Fluency = 53% and Comprehension = 46%), but most of this map implicated white matter. Our preliminary investigations of the CRLSM approach were encouraging, with simulated recovery of non-contiguous cortical networks and converging results of real data analysis with previous results. CRLSM reduces modeler decisions based on neurobiological theory and elegantly handles volume confounds and parasitic associations.
Topic Areas: Methods, Disorders: Acquired