Rethinking causality and data complexity in brain lesion-behaviour inference and its implications for lesion-behaviour modelling. (May 2020)

• Record Type: Journal Article
• Authors: Sperber, Christoph
• Is Part Of: Cortex. Volume 126(2020)
• Page Start: 49
• Page End: 62
• Publication Date: 2020-05
• Subjects: Stroke -- Outcome prediction -- Machine learning -- Support vector machine -- Multivariate lesion behaviour mapping; VLBM Voxel-based lesion behaviour mapping -- MLBM Multivariate lesion behaviour mapping -- SVM Support-vector machine; Neuropsychology -- Periodicals; Nervous system -- Periodicals; Neurology -- Periodicals; Psychophysiology -- Periodicals; Behavior -- Periodicals; Neurology -- Periodicals; 612.825
• DOI: 10.1016/j.cortex.2020.01.004
• Languages: English
• ISSNs: 0010-9452

Abstract

Abstract: Modelling behavioural deficits based on structural lesion imaging is a popular approach to map functions in the human brain, and efforts to translationally apply lesion-behaviour modelling to predict post-stroke outcomes are on the rise. The high-dimensional complexity of lesion data, however, evokes challenges in both lesion behaviour mapping and post stroke outcome prediction. This paper aims to deepen the understanding of this complexity by reframing it from the perspective of causal and non-causal dependencies in the data, and by discussing what this complexity implies for different data modelling approaches. By means of theoretical discussion and empirical examination, several common strategies and views are challenged, and future research perspectives are outlined. A main conclusion is that lesion-behaviour inference is subject to a lesion-anatomical bias that cannot be overcome by using multivariate models or any other algorithm that is blind to causality behind relations in the data. This affects the validity of lesion behaviour mapping and might even wrongfully identify paradoxical effects of lesion-induced functional facilitation – but, as this paper argues, only to a minor degree. Thus, multivariate lesion-brain inference appears to be a valuable tool to deepen our understanding of the human brain, but only because it takes into account the functional relation between brain areas. The perspective of causality and inter-variable dependence is further used to point out challenges in improving lesion behaviour models. Firstly, the dependencies in the data open up different possible strategies of data reduction, and considering those might improve post-stroke outcome prediction. Secondly, the role of non-topographical causal predictors of post stroke behaviour is discussed. The present article argues that, given these predictors, different strategies are required in the evaluation of model quality in lesion behaviour mapping and post stroke outcome prediction.