Skip to main content
. 2019 Mar 25;62(3):587–601. doi: 10.1044/2018_JSLHR-S-ASTM-18-0244

Figure 3.

Figure 3.

(A–C) Procedures to implement a decoding model using linear support vector machines (SVMs) to analyze EEG responses to continuous speech stimuli. (a) Extraction of EEG responses time-locked to phoneme onset (PRPs) with a predefined time window (e.g., 0–600 ms). (B) Examples of the PRPs at electrode FCz from one participant (from an unpublished data set in our lab). The corresponding phonemes are grouped into phonological categories plosive, fricative, nasal, and vowel. (C) Procedures to classify PRPs into one of the four phonological categories using linear SVM classifier (model). The SVM model is estimated with PRP data from 15 of the 16 participants and is validated to see how well it can generalize to PRP data in the held-out participant. Decoding accuracy reflects the percentage that the SVM model correctly identified the phonological categories in the held-out participant. (D) Results from one example participant (from an unpublished data set in our lab). This plot shows the topographic distribution of decoding accuracies across 62 electrodes. (E) The left (black) histogram represents the null distribution of decoding accuracies (n = 9,920) derived from permutation tests. The red vertical dashed line represents the actual decoding accuracy averaged across seven frontocentral electrodes (FC5, FC3, FC1, FCz, FC2, FC4, FC6). The black vertical dashed line indicates the 99th percentile in the null distribution. If the actual decoding accuracy is higher than the 99th percentile in the null distribution, it indicates that the actual decoding accuracy is statistically significant at an alpha value of .01. (F) A contingency table (also called confusion matrix) of the true versus predicted phonological categories from the decoding analysis. Each column corresponds to the true phonological category, and each row corresponds to the predicted phonological category. The shade and the numbers of a given cell denote the occurrence of a predicted phonological category in proportion to the total instances of a given true phonological category (i.e., probability).