On the interpretation of weight vectors of linear models in multivariate neuroimaging

Abstract

The increase in spatiotemporal resolution of neuroimaging devices is accompanied by a trend towards more powerful multivariate analysis methods. Often it is desired to interpret the outcome of these methods with respect to the cognitive processes under study. Here we discuss which methods allow for such interpretations, and provide guidelines for choosing an appropriate analysis for a given experimental goal: For a surgeon who needs to decide where to remove brain tissue it is most important to determine the origin of cognitive functions and associated neural processes. In contrast, when communicating with paralyzed or comatose patients via brain-computer interfaces, it is most important to accurately extract the neural processes specific to a certain mental state. These equally important but complementary objectives require different analysis methods. Determining the origin of neural processes in time or space from the parameters of a data-driven model requires what we call a forward model of the data; such a model explains how the measured data was generated from the neural sources. Examples are general linear models (GLMs). Methods for the extraction of neural information from data can be considered as backward models, as they attempt to reverse the data generating process. Examples are multivariate classifiers. Here we demonstrate that the parameters of forward models are neurophysiologically interpretable in the sense that significant nonzero weights are only observed at channels the activity of which is related to the brain process under study. In contrast, the interpretation of backward model parameters can lead to wrong conclusions regarding the spatial or temporal origin of the neural signals of interest, since significant nonzero weights may also be observed at channels the activity of which is statistically independent of the brain process under study. As a remedy for the linear case, we propose a procedure for transforming backward models into forward models. This procedure enables the neurophysiological interpretation of the parameters of linear backward models. We hope that this work raises awareness for an often encountered problem and provides a theoretical basis for conducting better interpretable multivariate neuroimaging analyses.

Congruent Visual Speech Enhances Cortical Entrainment to Continuous Auditory Speech in Noise-Free Conditions

“…Unlike the stimulus reconstruction approach, it is not a multivariate regression but represents multiple univariate mappings between stimulus and EEG. TRF model parameters are neurophysiologically interpretable, i.e., nonzero weights are only observed at channels where cortical activity is related to stimulus encoding (Haufe et al, 2014). This allows for examination of the amplitude, latency, and scalp topography of the stimulus-EEG relationship, complementing the stimulus reconstruction approach.…”

Spatiotemporal dynamics in human visual cortex rapidly encode the emotional content of faces

“…The weight vector for a linear SVM is based on the Lagrange multipliers assigned to each data point. To achieve interpretable spatial patterns (Haynes, 2015), feature weights were transformed into relevance patterns through multiplication by the data covariance matrix (Haufe et al, 2014). This allowed us to dynamically and directly assess the relative importance of all virtual electrodes used in source‐space decoding, as each ROI was represented by one feature and each decoding iteration was run on the whole brain.…”

Neuroadaptive technology enables implicit cursor control based on medial prefrontal cortex activity

“…Following Haufe et al (29), linear discriminant analysis (LDA) patterns A = ða j Þ j were generated for each participant from the LDA filters M = ðm j Þ j originally used for online classification by conjugation with the features' covariance matrix C: A = CMC −1 . Spatial interpretation of these patterns for each time window reflects a mixture of scalp activations related to discriminative source activity = ðâ j Þ j and classinvariant noise representation N, with A = + N. The latter was filtered out by weighting each pattern entry a j with the correlation of its associated feature activity vector over trials F j to the binary vector of true class labels…”