Poster D34, Wednesday, August 21, 2019, 5:15 – 7:00 pm, Restaurant Hall
Navigating the turbulent seas of lesion symptom mapping: Comparative analyses of univariate and multivariate lesion symptom mapping methods
Maria V. Ivanova1,2,3, Timothy Herron2, Brian Curran2, Nina F. Dronkers1,2,4, Juliana V. Baldo2;1University of California, Berkeley, 2Center for Aphasia and Related Disorders, VA Northern California Health Care System, 3National Research University Higher School of Economics, Moscow, 4University of California, Davis
Lesion symptom mapping (LSM) tools are used to identify brain regions critical for a given behavior. Univariate lesion-symptom mapping (ULSM) methods provide statistical comparisons of behavioral test scores in patients with and without a lesion on a voxel by voxel basis. Multivariate lesion-symptom mapping (MLSM) methods consider the effects of all lesioned voxels in one model simultaneously and analyze their contribution to behavior. Advantages and disadvantages of both techniques have been extensively discussed in the literature, but very little systematic work has been done to empirically test these claims. In the current study we conducted a comprehensive comparison between ULSM and MLSM methods by analyzing their performance under varying conditions with artificial behavioral scores. We used lesion masks from 404 left hemisphere chronic stroke patients. Thirteen LSM methods were compared: 5 ULSM (voxel-based linear LSM methods with different permutation-based FWER thresholds for multiple comparisons) and 8 MLSM (6 data reduction (DR) and 2 regression methods). Using artificial behavioral scores (based on lesion load to atlas-based anatomical ROIs), we investigated mapping power and accuracy for single and dual (network type) anatomical target simulations. We tested spatial precision of mapping across anatomical target location, sample size, noise level, and lesion smoothing by using different distance- and overlap-based metrics as indices of spatial accuracy. Additionally, we performed a false positive simulation, where the behavioral target variable consisted of pure Gaussian white noise and thus should not lead to detection of any anatomical areas as significantly related to behavior. Here, we evaluated the size and number of the false positive clusters. Single anatomical target simulations revealed: a) good spatial accuracy for ULSM methods with conservative FWER thresholds and some of the simpler DR (e.g., SVD-based) and regression-based (e.g., SVR) MLSM methods; b) variable accuracy across spatial locations, with especially poor performance in cortical locations on the edge of the lesion masks (areas of lower power); c) more accurate localization with lesion mask smoothing for all LSM methods; d) the importance of having a sample with ≥ 64 patients (with the majority of MLSM methods requiring on average 10-20 more patients to achieve a ULSM level of spatial accuracy); e) robustness of the weighted centroid as a measure of LSM statistical map location. Simulations with a dual anatomical target showed: a) more accurate localization with some of the DR MLSM techniques (e.g., LPCA) as well as ULSM methods with relatively liberal cluster-based FWER thresholds; b) the importance of having a sample with ≥ 100 patients. False positive cluster sizes were generally the lowest for ULSM methods with conservative FWER thresholds and regression-based MLSM methods. In summary, our simulations show no clear superiority of MLSM techniques over the ULSM methods. Depending on the design of a particular LSM study and specific hypotheses regarding the expected brain-behavior relationship, different LSM methods are indicated. In general, it is advantageous to implement both ULSM and MLSM methods in tandem to enhance confidence in the results, as significant matching foci identified with both methods are unlikely to be spurious.
Themes: Methods, Disorders: Acquired
Method: Other