Fig. 2: Variability of brain patterns learned by unimodal and multimodal models.
From: Multimodal fusion of brain signals for robust prediction of psychosis transition
A Linear classifiers use feature weight vectors for making a classification, representing the multivariate patterns of feature importance and variability across training samples and folds. B–D We used principal component analysis (PCA) to visualize this weight vector variability across train sets (100 total) for each modality in unimodal SVM and MKL when predicting conversion (CHR-converters vs. CHR-nonconverters). Relative to SVM models which formed clusters based on training samples, MKL benefitted from other modalities to stabilize the feature weights. Please refer to the supplement for a similar analysis that identifies CHR status.
