Time-Resolved EEG Decoding of Semantic Processing Reveals Altered Neural Dynamics in Depression and Suicidality

Abstract

Depression and suicidality affect cognitive and emotional processes, yet objective, task-evoked neural readouts of mental health remain limited. We investigated the spatiotemporal dynamics of affective semantic processing using multivariate decoding of time-resolved, 64-channel electroencephalography (EEG). Participants (N=137) performed a sentence-evaluation task with emotionally salient, self-referential statements. We identified robust neural signatures of semantic processing, with peak decoding accuracy between 300-600 ms -- a window associated with rapid, stimulus-driven semantic evaluation and conflict monitoring. Relative to healthy controls, individuals with depression and suicidal ideation showed earlier onset, longer duration, and greater amplitude decoding responses, along with broader cross-temporal generalization and enhanced contributions from frontocentral and parietotemporal components. These findings suggest altered sensitivity and impaired disengagement from emotionally salient content in the clinical groups, advancing our understanding of the neurocognitive basis of mental health and establishing a compact and interpretable EEG-based index of semantic-evaluation dynamics with potential diagnostic relevance.

Figures (8)

• Figure 1: Sentence evaluation task and behavioral responses.A. Participants were presented with 320 sentences and prompted to indicate whether they agreed or disagreed with each sentence. Each trial began with a 300 ms fixation cross, followed by sequential presentation of each word in the sentence for 300 ms, with a 300 ms black screen inter-stimulus-interval (ISI) between words, except for the final word, which was presented for 600 ms. Participants then had a 2-second window to respond by indicating “agree” or “disagree” using a push-button box. B. Average response time across participants for different sentence types (congruent and incongruent) and responses (agree and disagree). Responses that were faster than 100 ms were discarded. The error bars denote standard error. Stars correspond to statistically significant differences between conditions (Welch's t-test, $***$$p<0.001$).
• Figure 2: Multivariate EEG pattern analysis pipelineA. Channel topography of 64-channel configuration. All peripheral channels were excluded from analysis to avoid noisy inputs. Black dots indicate the inner set of 47 channels used in the EEG analysis. B. Latent features shared across groups (C: control, D: depressed, and S: suicidal) were extracted using PCA on the grand-averaged EEG data across the training set (see Methods for details). Every trial in the training set of all participants within a group was averaged to obtain the grand-averaged EEG per group. The resulting grand-averaged EEGs were concatenated across time, and PCA was performed. The weight vectors for the top three PCs $\left(\bm{w}_{\bm{pc}_{1}},\ \bm{w}_{\bm{pc}_{2}},\ \bm{w}_{\bm{pc}_{3}}\right)$ were retained, forming the projection weight matrix $\bm{W}$. C. Within-subject EEG decoding using linear kernel SVM was applied. Each subject's EEG data $\bm{E}$ was projected onto a shared latent space using $\bm{W}$, yielding the latent representation of the neural data $\bm{\hat{\bm{E}}}$. Then, 250 bootstrapped trials were generated for each condition (200 for training and 50 for testing) by randomly sampling with replacement and sub-averaging 12 trials independently for each training/testing cohort. For each time point $t$, a 5-fold cross-validation using an SVM classifier was performed. Decodability (%) was reported by averaging the accuracies across all folds (see Methods for details).
• Figure 3: EEG decoding of sentence semantics and behavioral responsesA, B. Time-resolved decoding accuracy (%, decodability) of multivariate EEG patterns for (A) congruent vs. incongruent sentence evaluation and (B) agree vs. disagree responses, aligned to the onset of the final word of the sentence ($N=137$, solid blue lines). Red horizontal lines along the x-axis indicate time points, where decoding accuracy was significantly above chance, as determined by a two-sided cluster-based permutation test ($N=137$, cluster-defining threshold $p<0.05$, corrected significance level $p<0.05$, 5000 permutations). Shaded blue areas represent the standard error of the mean. The response window began at 900 ms (blue triangle), and the red triangle marks the mode of response time across all participants (1199 ms).
• Figure 4: Group sentence decoding from EEGA. Sentence decoding averaged across participants within each group (control, depressed, and suicidal) aligned to the onset of the final word (solid blue lines). Red-colored lines at the bottom of each plot indicate time points where the decoding accuracy (decodability, %) was significantly above chance level (50%, dashed black lines; two-sided cluster-based permutation test, cluster-defining threshold $p<0.05$, corrected significance level $p<0.05$, 5000 permutations). Blue triangle indicates the response window start (900 ms), and red triangle shows the mode of the response time for each group (control: 1154 ms, depressed: 1192 ms, and suicidal: 1154 ms). B. Onset latency, offset latency, peak latency, and peak amplitude for sentence decoding in each group (control: green, depressed: blue, and suicidal: red). Error bars denote bootstrapped 95% confidence intervals. Stars indicate significant difference between groups (one-sample two-sided bootstrap test, $***$$p<0.001$, $**$$p<0.01$, FDR-corrected).
• Figure 5: Spatial contribution estimates of the sentence decoding neural featuresA. Topographical distributions of the top 3 PCs estimated from the multivariate analysis (Fig. \ref{['fig2']}B). Red-highlighted regions indicate the enhanced channel contribution, and the blue-highlighted regions indicate the suppressed channel contribution. B. Mean estimation of the weight assigned to each of the 3 PCs by the linear SVM across participants at each time point relative to the final word onset during sentence decoding. The colored lines at the bottom of the plot denote the significant time points (two-sided cluster-based permutation test, cluster-defining threshold $p<0.05$, corrected significance level $p<0.05$, 5000 permutations). C, D. Relationship between sentence decoding accuracy and spatial feature contributions across groups (control: green, depressed: blue, and suicidal: red). C. Empirical distribution of the Pearson correlation between sentence decoding accuracy (Fig. \ref{['fig4']} A) and time-resolved absolute SVM feature weights (Fig. \ref{['fig5']} B) for each group. The x-axis represents the Pearson correlation with $PC_1$ weights, and the y-axis denotes the Pearson correlation with $PC_2$. D. Group-level comparison of correlation coefficients for $PC_1$ (left) and $PC_2$ (right). Error bars denote bootstrapped 95% confidence intervals. Stars indicate significant difference between groups (one-sample two-sided bootstrap test, $***$$p<0.001$, FDR-corrected).