Analysis of clinical anatomical correlates of motor deficits in stroke by multivariate lesion inference based on game theory

 NOTHING HERE gets survivors recovered! Useless!

Analysis of clinical anatomical correlates of motor deficits in stroke by multivariate lesion inference based on game theory

Monica N. Toba^1*†Caroline Malherbe^2,3*†Melissa Zavaglia^2,4†Audrey Arnoux¹Mélanie Barbay¹Claus C. Hilgetag^2,5‡Olivier Godefroy^1‡

• ¹Laboratory of Functional Neurosciences (EA 4559), University Hospital of Amiens, University of Picardie Jules Verne, Amiens, France
• ²Department of Computational Neuroscience, Hamburg Center of Neuroscience, University Medical Center Hamburg-Eppendorf, Hamburg, Germany
• ³Department of Neurology, Head and Neuro Center, University Medical Center Hamburg-Eppendorf, Hamburg, Germany
• ⁴MIRMI - Munich Institute of Robotics and Machine Intelligence, Technische Universität München, Munich, Germany
• ⁵Department of Health Sciences, Boston University, Boston, MA, United States

Introduction: The exploration of causal functional inferences on the basis of deficits observed after neurological impairments is often based on the separate study of gray matter regions or white matter tracts. Here, we aimed at jointly analysing contributions of gray matter and white matter by using the domain of motor function and the approach of iterative estimated Multi-perturbation Shapley Analysis (MSA), a multivariate game-theoretical lesion inference method.

Methods: We analyzed motor scores assessed by the National Institute of Health Stroke Scale (NIHSS) together with corresponding lesion patterns of 272 stroke patients using a finely parcellated map of 150 gray matter regions and white matter tracts of the brain.

Results: MSA revealed a small set of essential causal contributions to motor function from the internal capsule, the cortico-spinal tract, and the cortico-ponto-cerebellum tract.

Discussion: These findings emphasize the connectional anatomy of motor function and, on the methodological side, confirm that the advanced multivariate method of iterative estimated MSA provides a practical strategy for the characterization of brain functions on the basis of finely resolved maps of the brain.

1 Introduction

A wide variety of gray matter regions and white matter connections of the brain interact in order to produce complex movements. Specifically, the motor system includes gray matter structures, such as the premotor and motor cortices, basal ganglia, the cerebellum, areas of the association cortex and portions of the thalamus, as well as white matter bundles, such as the corticospinal tract (traditionally considered the principal mediator of voluntary movements), the vestibulospinal, reticulospinal, rubrospinal and tectospinal tracts (Nolte, 2009).

A frequently used technique for defining the anatomical correlates of functions in stroke patients is voxel-based lesion-symptom mapping (VLSM) (Bates et al., 2003). This method indicates the univariate association of damaged voxels with a particular deficit. Previous studies showed that VLSM maps obtained, for instance, for speech fluency and language comprehension (Bates et al., 2003), as well as such derived for the orienting of attention (Toba et al., 2017; Verdon et al., 2010) were in agreement with findings from functional brain imaging. However, several studies have emphasized that VLSM is sensitive to the distribution of lesions within vascular territories and the frequency of impaired voxels (Mah et al., 2014; Arnoux et al., 2018), emphasizing the need to use newly developed lesion inference approaches, particularly anchored in multivariate inferences (Karnath and Smith, 2014; Khalilian et al., 2024; Nachev, 2015). Specifically, such techniques are capable of defining and calculating the interrelated contributions of network elements from a dataset of multiple perturbations (or lesions) (Keinan et al., 2004a). Multivariate machine learning approaches, such as classification by support vector machines (SVMs), can be used to map brain functions onto cerebral structures (Corbetta et al., 2015; Forkert et al., 2015; Smith et al., 2013; Zavaglia et al., 2015; Zhang et al., 2014). As a further alternative, the Multi-perturbation Shapley value Analysis (MSA) represents a lesion inference approach based on game theory, designed to calculate the contribution of the network elements (specifically, brain regions) and the interactions existing between them, based on a dataset of multiple lesions. Brain regions are considered as “players” in a game who interact to achieve a behavioral outcome. This approach was validated in ground truth simulations as a better-performing option for lesion inference than VLSM (Zavaglia et al., 2024) and has already been applied to lesion inference in studying brain functions (Zavaglia et al., 2015; Malherbe et al., 2021) as well as specifically in the context of attentional functions (Kaufman et al., 2009; Zavaglia and Hilgetag, 2016; Malherbe et al., 2018; Toba et al., 2017; Toba et al., 2020c).

At the behavioral level, a very widely used measure of the motor function in stroke patients is the National Institute of Health Stroke Scale (NIHSS) (Brott et al., 1989). This scale is used to generally characterize the clinical or functional status of stroke patients and, to this end, regroups items testing different functions, such as the level of consciousness, horizontal eye movements, visual field, facial palsy, motor arm, motor leg, limb ataxia, sensory, language, dysarthria, extinction and inattention. NIHSS provides an efficient measure with strong clinical validity to quantify stroke severity. However, the detailed assessment of each function included in the NIHSS requires investigations of different functional elements and this aim cannot be accomplished with a global rating score such as the NIHSS. Because its validity in assessing deficits with known anatomy (such as motor function) has been demonstrated, the NIHSS has previously been used in new methods examining clinico-anatomical correlations (e.g., Arnoux et al., 2018; Menezes et al., 2007; Zavaglia et al., 2015; Malherbe et al., 2021). It has been shown that in the analyses conducted on behavioral results obtained with traditional scales, such as the NIHSS and the modified Rankin scale, a better prediction of stroke severity could be obtained only when considering both the volume and the lesion location obtained on structural magnetic resonance imaging (MRI) data (Forkert et al., 2015; Menezes et al., 2007). By analysing the global NIHSS and lesions of 148 acute stroke patients with a multivariate approach and focusing solely on gray matter structures, Zavaglia et al. (2015) inferred various locations underlying functions tested by the NIHSS, such as the bilateral caudate, left insula and bilateral parietal and frontal lobes. Of note in these results was the presence of bilateral frontal regions (also comprising primary and supplementary motor brain areas) and basal ganglia involved in the motor system, likely linked to the fact that motor symptoms result in high score values of the NIHSS (specifically, they can explain up to 18 of 42 possible score points). However, in order to more specifically explore motor functions, an individual sub-score of the NIHSS that focuses on motor tasks should be considered. Moreover, data available in Zavaglia et al. (2015) allowed only the analysis of functional inferences of gray matter structures while white matter connections should also be considered in order to completely characterize causal functional inferences in a given system (see also Toba et al., 2020a; Toba et al., 2020b; Godefroy et al., 1998). Malherbe et al. (2021) considered a high resolution parcellation of the brain into 294 white matter and gray matter regions in a large population of 394 acute stroke patients. These authors reduced the number of regions to only those that significantly inferred some brain functions (specifically the “left motor function,” “the right motor function” and the “language and consciousness function”) issued from NIHSS factors previously published by Lyden et al. (2004). Specifically, the VLSM approach was used to first reduce the regions to one hemisphere. Then, the study used an iterative loop performing MSA and discarding the region with the smallest contribution to a function. As a result, Malherbe et al. (2021) inferred for each function a base set of causally contributing regions. Concretely, the dorsolateral putamen and the posterior limb of the left and right internal capsule were related to the motor functions right and left, respectively. Whereas the left motor function was also associated with the superior corona radiata and the paracentral lobe of the right hemisphere as well as the right caudal area of the cingulate gyrus, the right motor function was related to the prefrontal gyrus, the external capsule and the sagittal stratum fasciculi of the left hemisphere.

In the present paper, by taking advantage of the well-characterized anatomical and functional model of the motor system and an improved estimated MSA algorithm, we aimed to explore causal functional contributions of both, gray matter structures as well as white matter connections, without any preselection of a subset of regions. To this aim, we used the newly developed approach of iterative estimated MSA in a large sample of 272 patients with motor impairments assessed by the NIHSS motor sub-score.

More at link.