Technical considerations of a game-theoretical approach for lesion symptom mapping

Abstract

BackgroundVarious strategies have been used for inferring brain functions from stroke lesions. We explored a new mathematical approach based on game theory, the so-called multi-perturbation Shapley value analysis (MSA), to assess causal function localizations and interactions from multiple perturbation data. We applied MSA to a dataset composed of lesion patterns of 148 acute stroke patients and their National Institutes of Health Stroke Scale (NIHSS) scores, to systematically investigate the influence of different parameter settings on the outcomes of the approach. Specifically, we investigated aspects of MSA methodology including the choice of the predictor algorithm (typology and kernel functions), training dataset (original versus binary), as well as the influence of lesion thresholds. We assessed the suitability of MSA for processing real clinical lesion data and established the central parameters for this analysis.ResultsWe derived general recommendations for the analysis of clinical datasets by MSA and showed that, for the studied dataset, the best approach was to use a linear-kernel support vector machine predictor, trained with a binary training dataset, where the binarization was implemented through a median threshold of lesion size for each region. We demonstrated that the results obtained with different MSA variants lead to almost identical results as the basic MSA.ConclusionsMSA is a feasible approach for the multivariate lesion analysis of clinical stroke data. Informed choices need to be made to set parameters that may affect the analysis outcome.

Game theoretical mapping of causal interactions underlying visuo‐spatial attention in the human brain based on stroke lesions

“…It is, however, important to mention that the SVM results could be biased by the fact that we selected the kernels that produced the best classification accuracies for each task. Accordingly, better results could be obtained from a MSA analysis with a more generalizable SVM, trained on a higher number of patients who exhibit lesions across more of the examined regions [Zavaglia et al, ]. We should also note that in the present patients cohort, several regions were impacted by very small lesions, which occurred only in a few patients (e.g., BA19 and BA7).…”

“…At difference with Kaufman et al [], we used an improved model of prediction to analyze the degree of damage of a set of regions of interest, which were also combined with a larger number of well‐standardized behavioral tasks characterizing visuo‐spatial and attention domains. More specifically, to characterize our dataset, we employed a predictive model built by a support vector machine where the model parameters were chosen to obtain the most accurate and reliable predictions (for details, see Zavaglia et al []). By using the MSA approach in this study, we were able to explore brain–behavior relationships at different levels of complexity (including individual positive and negative causal contributions of brain regions, as well as combined contributions reflecting interactions between areas), and quantitatively characterize the distributed cortical network underlying a specific cognitive function, such as spatial attention.…”

Game-theoretical mapping of fundamental brain functions based on lesion deficits in acute stroke

“…All possible combinations of these parameters were tuned for the graded dataset but also with a binarized dataset, which was thresholded by the median value of the percentage of lesioned voxels for each region of interest, as described previously. 18 With each combination of parameters and database, for each NIHSS factor, we applied a ‘leave-one-out’ cross-validation, using in turn each patient from the database as the validation data and all the remaining patients as the training data . To ensure the quality of the prediction, we compared the true set of performance values of an NIHSS factor with each set of estimated performance values (one for each combination of parameters), by computing the associated F1 score: with tp: true positive, fn: false negative and fp: false positive.…”

Systematic perturbation of an artificial neural network: A step towards quantifying causal contributions in the brain

“…A knowledge of how much the element is contributing to a function allows us to rank those elements accordingly. The Shapley value as a measure of causal contribution is powerful and intuitive, however, it is important to emphasize what this value does not reflect (see [ 36 ] for a technical perspective). For example, the Shapley value by default does not reveal mechanisms, in the sense that it does not show what computations were done by individual elements.…”