Fig 1.
A Confusion matrix showing the proportion of correctly classified trials (diagonal) and misclassified trials. B Classifier accuracy. The mid blue bars represent the test accuracy in the 20 folds, where every fold is a subject the classifier was not trained on. The dark blue bar shows the overall classifier accuracy, which was tested against chance level.
Fig 2.
The 2% most important predictors are shown across time, frequency and space. A higher variable importance score implies that this predictor had a higher informative value in the random forest model to partition the data into trials with auditory and visual perception. The orthogonal views are centered on the voxel showing the highest variable importance.
Fig 3.
Variable importances in visual and auditory peak voxels.
A Peak voxel locations for auditory and visual cortex (compare peak voxels from Fig 2) B Time-frequency representation of variable importances in those peak voxels. Black boxes indicate those variables which were among the 2% most informative predictors.
Fig 4.
This figure shows the difference in averaged gamma power between visual and auditory word presentation trials. A Spatial representation of gamma power for two frequency bands (75-95 Hz, top, and 105-125 Hz, bottom). Red hues represent a higher gamma power in the average of visual trials, the blue colors depict higher gamma power in the average of the auditory condition. Black boxes indicate the 2% most informative predictors as shown in Fig 2. B Gamma power in visual and auditory peak voxels. Shown is the difference between the visual and auditory condition, black boxes again indicate the most informative predictors for the classifier model.
Fig 5.
Performance comparison of the random forest algorithm to SVM models. A Results from the linear SVM compared to the results from the random forest algorithm across cross-validation folds. Random forest performed significantly better than the linear SVM (p = 0.002, 0.004 with Bonferroni correction). B Results from a non-linear SVM with RBF kernel compared to the random forest results. The non-linear SVM yielded a higher classification accuracy than the random forest model (p = 0.047), applying a Bonferroni-correction for multiple comparisons renders this effect insignificant (pcorrected = 0.094).