Multivariate pattern analysis for MEG: a comprehensive comparison of dissimilarity measures

Abstract

1Multivariate pattern analysis (MVPA) methods such as decoding and representational similarity 2 analysis (RSA) are growing rapidly in popularity for the analysis of magnetoencephalography (MEG) 3 data. However, little is known about the relative performance and characteristics of the specific 4 dissimilarity measures used to describe differences between evoked activation patterns. Here we used 5 a multisession MEG dataset to qualitatively characterize a range of dissimilarity measures and to 6 quantitatively compare them with respect to classification accuracy (for decoding) and between-7

Decoding diagnosis and lifetime consumption in alcohol dependence from grey‐matter pattern information

“…We have shown previously that WeiRD, despite its simplicity, is competitive with other classifiers (SVM, Random Forest) across a range of real‐world and simulated classification tasks similar to the present . In more recent work , we have compared WeiRD to a larger set of classifiers (including support vector machine, Gaussian Naïve Bayes, linear discriminant analysis) in the MEG modality and could likewise confirm the strong performance of WeiRD.…”

“…On the other hand, while simple classification strategies such as decision trees, which implement a hierarchical chain of decision rules, can expose the contribution and role of each feature in a human‐readable format, they suffer from overfitting and thus low generalization performance . By contrast, the voting‐based classification scheme used by WeiRD combines high performance, as shown in this and previous studies , with high transparency, by quantifying the contribution of each feature to classification in the most direct way possible (i.e. votes).…”

“…). In addition, all analysis schemes were submitted to two different classifiers: weighted robust distance (WeiRD) and support vector machine (SVM). Figure a shows that parcellation‐based schemes (parcelled voxel patterns and averages) were generally superior to non‐parcelled schemes (whole‐brain averages and voxel patterns).…”

“…support vector machine classification) provide only limited information about which parts of the data are crucial for diagnostic classification. Here, we unpacked the black box of machine learning by means of the recently proposed weighted robust distance (WeiRD) classifier, which combines high transparency with competitive performance . Disregard of graded disease measures in binary classification schemes. A drawback of binary classification between case and control groups is the fact that classification may often capture between‐group differences that are unrelated to the disease mechanisms.…”

Consistent pre-stimulus influences on auditory perception across the lifespan

“…Since Y(t) is indicative of the extent of separability between frequency levels, we exploit this as a measure representing the amount of encoded sensory evidence about the task relevant tones (Grootswagers et al, 2017;Guggenmos et al, 2017). We computed each components' time course by applying the respective weight to all trials and time points, resulting in a one-dimensional projection of single-trial task-related activity which we then analysed further.…”