Slide Slam R16
Predictive Validity Comparison for Multivariate Lesion-Behavior Maps
Tatiana Schnur1, John Magnotti2, Jaclyn Patterson1; 1Baylor College of Medicine, 2University of Pennsylvania
Introduction. Multivariate lesion-behavior mapping (LBM) provides a statistical map of the association between patterns of brain damage and individual differences in behavior. To understand whether behaviors are mediated by distinct brain regions, researchers often compare LBM beta weights by either the subtraction method (emphasizing differences) or the correlation method (emphasizing similarity). However, both methods lack a principled way to determine LBM distinctness and are disconnected from a major goal of LBMs: predicting behavior from brain damage. Without such criteria, researchers may unwittingly draw conclusions from numeric differences between LBMs that are irrelevant to predicting behavior from brain damage. Here, we developed and validated a Predictive Validity Comparison (PVC) method that establishes such a criterion. Method. The PVC method conceptualizes the LBM comparison problem as a choice between two hypotheses: Under the null hypothesis (H0), individual differences across two behaviors are the result of a single lesion pattern (+ noise). Under the alternative hypothesis (HA), individual differences across behaviors are the result of distinct lesion patterns. The method fits multivariate LBMs under each hypothesis (one under H0; two under HA) using SCCAN (Pustina et al., 2018) and compares their predictive accuracy. If the predictions under HA are better than under H0 (using Akaike Information Criterion; AIC), we conclude that separate LBMs better fit the behaviors. Otherwise, only a single LBM is needed. To assess the practical utility of PVC, we compared it with common LBM comparison methods, the subtraction and correlation methods, using published lesion-behavior stroke datasets with either highly similar behaviors (r=0.89, Moss Rehabilitation Research Institute Data, MRRI) or highly dissimilar behaviors (r=-0.02; Ding et al., 2020). Second, we validated PVC using region-of-interest based simulations. Two simulated behaviors were derived from proportion damage in one (33 simulations) vs. two (666 simulations) Brodmann regions using the MRRI data set. Results were organized by the regional proportion damage across subjects and between-region damage correlations. Results. For similar behaviors (MRRI dataset), PVC concluded only a single LBM was necessary (AIC difference -382 in favor of H0); for dissimilar behaviors (Ding et al. dataset), PVC concluded separate LBMs were necessary (AIC difference 372 in favor of HA). In contrast, the subtraction and correlation methods provided indeterminate results, lacking clear decision thresholds and showing strong sensitivity to LBM-fitting parameters. On simulated data, PVC had 94% sensitivity (accurately determining behaviors had distinct neural bases) on high-powered regions (>=10% of participants had >=5% damage to both regions; n=300) and 96% sensitivity on low-powered regions (n=366). Sensitivity depended on the between-region damage correlation, near ceiling for low to moderate correlations, but near the floor for extreme correlations (>0.7), although these were rare (n=31). Specificity (accurately determining behaviors had a shared neural basis) was excellent for both high (100%; 25/25 correct) and low-powered regions (92%; 7/8 correct). Conclusion. The PVC method’s excellent performance with real and simulated data facilitates principled determination of LBM distinctness. With this advance, researchers can better determine whether individual differences across two behaviors arise from distinct lesion patterns.