Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks
Using predictive validity to compare associations between brain damage and behavior
There is a Poster PDF for this presentation, but you must be a current member or registered to attend SNL 2023 to view it. Please go to your Account Home page to register.
Poster C111 in Poster Session C, Wednesday, October 25, 10:15 am - 12:00 pm CEST, Espace Vieux-Port
Tatiana Schnur1, John Magnotti1,2; 1Baylor College of Medicine, Houston, USA, 2University of Pennsylvania, Philadelphia, USA
Introduction. Multivariate lesion-behavior mapping (LBM) estimates the association between patterns of brain damage and individual behavior. The LBM algorithm assigns weights to voxels in patients’ lesion volumes which are then used to predict their behavior. A common problem is determining whether two behaviors are mediated by damage to the same vs. different brain regions (e.g., Western Aphasia Battery scores (WAB-AQ) vs. naming accuracy). Researchers often use an “Overlap Method” to compare LBM weights via intersection (finding areas common to both LBMs) or subtraction (finding areas unique to each LBM). Alternatively, researchers correlate LBM weights, extrapolating that high correlations suggest similar neural bases, whereas low/negative correlations suggest distinct neural bases. However, both methods lack validated decision rules for determining LBM distinctiveness and are disconnected from the goal of accurate prediction of behavior. We developed the Predictive Validity Comparison method (PVC) to determine multivariate LBM uniqueness based on predictive accuracy: two behaviors have distinct neural bases if and only if separately-fit LBMs provide unique predictive power. If a single LBM accurately predicts both behaviors, the behaviors cannot be assumed to have separable neural bases, regardless of the similarity/difference of the voxel weights. Methods. PVC Algorithm. PVC requires two behavioral scores and one lesion volume per patient. Two sets of multivariate LBMs are built according to a null hypothesis (behaviors assumed to have the same neural basis) vs. the alternative hypothesis (behaviors have separable neural bases). Under the null hypothesis a single LBM is fitted to a combined behavioral score. Under the alternative hypothesis separate LBMs are fitted to each behavior. Fitted LBMs are then used to predict behavior. Prediction accuracy is compared using Akaike Information Criterion (AIC): if the alternative hypothesis has lower AIC the data are consistent with the two behaviors having separable neural bases. Approach. We compared PVC with the Overlap and Correlation methods using two stroke lesion-behavior datasets (n = 52 from Ding et al., 2020, connected speech measures of words per minute vs. proportion pronouns to nouns; n = 131 MRRI data from Pustina et al., 2018, WAB-AQ vs. Philadelphia Naming Test accuracy) across a range of LBM hyperparameters. Simulations based on the MRRI data compared the methods with known ground truth. Results. Across two real datasets and extensive simulations, PVC was the most accurate method (average sensitivity = 99% vs. 100% specificity) at determining same vs. different neural bases for behaviors, followed by the Overlap method (72% vs. 100%) and the Correlation Method (71% vs. 97%). PVC judgements were unaffected by LBM sparseness (number of voxels in each LBM), whereas the Overlap and Correlation methods were strongly affected. Conclusions. The goal of the PVC method is to help researchers decide if two behaviors have distinct neural bases. Without clear, validated decision rules, researchers may draw conclusions from numeric similarities/differences in voxel-level weights that are irrelevant to predicting behavior. With PVC, researchers can make stronger conclusions about the neural separability of behaviors. An open-source, GUI-driven app implementing PVC is available for download from our website: https://sites.google.com/site/ttschnur/researchprojects/predictive-validity-comparison-for-lesion-behavior-mapping.
Topic Areas: Methods, Disorders: Acquired