Neural Responses to Affective Sentences Reveal Signatures of Depression Authors: Aditya Kommineni, Woojae Jeong, Kleanthis Avramidis, Colin McDaniel, Myzelle Hughes, Thomas McGee, Elsi Kaiser, Kristina Lerman, Idan A. Blank, Dani Byrd, Assal Habibi, B. Rael Cahn, Sudarsana Kadiri, Takfarinas Medani, Richard M. Leahy, Shrikanth Narayanan Published: 2026 ABSTRACT Major Depressive Disorder (MDD) is a highly prevalent mental health condition, and a deeper understanding of its neurocognitive foundations is essential for identifying how core functions such as emotional and self-referential processing are affected. We investigate how depression symptoms alters the temporal dynamics of emotional processing by measuring neural responses to self-referential affective sentences using surface electroencephalography (EEG) in healthy and depressed individuals. Our results reveal significant group-level differences in neural activity during sentence viewing, suggesting disrupted integration of emotional and self-referential information in depression. Deep learning model trained on these responses achieves an area under the receiver operating curve (AUC) of 0.72 in distinguishing healthy from depressed participants, and 0.65 in differentiating depressed subgroups with and without suicidal ideation symptoms. Spatial ablations highlight anterior electrodes associated with semantic and affective processing as key contributors. These findings suggest stable, stimulus-driven neural signatures of depression symptoms that may inform future screening tools. Introduction Depression is one of the most prevalent mental health disorders worldwide, with estimates indicating that around 5% of the worlds’ adult population [1, 2] suffers from this condition. The primary methods for screening and monitoring depression rely on self-reported questionnaires, such as the Patient Health Questionnaire (PHQ-9) [3], Beck’s Depression Inventory (BDI) [4] and Hamilton Depression Ratings Scale (HDRS) [5]. While these questionnaires are effective to varying degrees at screening patients for depression, they provide only limited information about the affected underlying neuro-cognitive processes in individuals, limiting the ability to personalize treatments. Given the heterogeneity of depressive symptomatology across patient populations [6, 7], it is crucial to elucidate the underlying neurophysiological mechanisms to support the development of more effective and individualized procedures for screening, monitoring, and treatment. Prior functional imaging studies have identified increased activity in anterior cingulate cortex (especially the subgenual anterior cingulate) during presentation of emotional stimuli, altered connectivity in prefrontal cortical areas, and default mode network as potential differentiating markers in depressed participants [8–13]. Additionally, data from resting-state functional Magnetic Resonance Imaging (fMRI) have enabled successful classification of depressed participants [14–16], underscoring the promise for identifying biomarkers for depression. The scalability, low cost, and high temporal resolution of electroencephalography (EEG) make it a more practical modality for studying neural characteristics in depression and positioning it as a promising tool for large-scale screening in clinical settings [17, 18]. While there has been considerable interest in exploring resting-state differences between healthy and depressed populations [19, 20], task-based responses offer potentially greater specificity [21–25] to decipher specific disruptions (e.g., working memory, reward processing, emotional processing) in clinical populations with mental health disorders. Emotional processing has been observed to be a core component impacted in major depressive disorder (MDD) [26–29]. While some studies have found negative processing bias in a scrambled sentence task to be predictive of subsequent depressive symptoms [30], others report reduced neural activity to processing positive emotional information in a working memory task [31]. Subsequent studies have shown impaired semantic sentence processing in depressed individuals [32] but offer conflicting insights related to sentence processing effects, with some studies [33] noting no significant differences in N400 Event-related Potentials (ERP) response during a sentence processing task in healthy vs. depressed participants, and others [34] observing reduced N400 amplitudes in depressed participants for a self-referential phrase ending in negative adjectives. This warrants further investigation into whether healthy and depressed individuals exhibit distinct neural responses during the reading of affectively salient sentences, and whether these differences can serve as a reliable tool for screening. The present study employs a self-referential sentence response paradigm to investigate temporal dynamics of affective self-referential sentences in healthy and depressed populations, along with quantifying the ability to distinguish mental health groups through the recorded neural signals. While prior studies using self-referential sentence response tasks have identified differences in ERP [35–37], they have not assessed the feasibility of using this experimental design to distinguish healthy and depressed individuals in a classification paradigm. Our objectives are twofold: (1) to examine clinical versus healthy group-level differences in the temporal profile of neural responses to the sentence response task through Multivariate Pattern Analysis; and (2) to evaluate the feasibility of using these neural responses to classify individuals with depressive symptoms using a deep learning framework. The temporal dynamics reveal baseline differences between healthy and depressed participants, agnostic to the semantic nature of the sentence. This indicates a difference in cognitive processing during sentence viewing, presented in a one-word-at-a time reading task. Additionally both the temporal dynamics and deep learning experiments exhibit better distinguishability between healthy and depressed groups for positive sentiment sentences compared to negative sentiment sentences. A deep learning model to classify healthy and depressed participants achieves an AUC (Area Under the Receiver Operating Characteristic curve) of 0.72, highlighting the potential of the experimental paradigm and modeling approach. Additionally, attempts to classify subgroups within depressed participants –depressed non-suicidal versus depressed suicidal –yield an AUC of 0.65, while promising, indicate open challenges when differentiating subgroups within depressed participants. Methods Ethics Approval Our study was approved by the Institutional Review Board (IRB) of the University of Southern California (UP-23-00071). Personally identifiable information (PII) of the participants was de-identified for processing. Informed consent was obtained from all participants. All methods were performed in accordance with relevant institutional and national guidelines and regulations. Participants To study the differences in neural responses between mental health groups to sentence response task, 160 participants between 18 and 25 years old were recruited in the city of Los Angeles, California, for this study. The eligibility of participants was determined through Patient Health Questionnaire (PHQ-9) [3], and the Suicidal Ideation Scale (SIS) [38] scores at the time of recruitment. Participants were then grouped based on the questionnaire scores (recording during screening) as follows into Healthy control (C, PHQ-9 < 5), Depressed non-suicidal (D, PHQ-9 > 9; SIS ≤ 16) or Depressed suicidal (S, PHQ-9 > 9; SIS > 16). Additionally, participants who met the inclusion criteria completed GAD-7 [39] and TIPI [40] questionnaires on the day of the EEG recording. Of the 160 participants, 13 participants with large variations in PHQ-9 scores and one participant who was unable to complete the study were excluded. The demographics of the cohort are as shown in Table 1. For additional details about recruitment, refer to supplementary material.Table 1Demographics of the three mental health groups in the study.CDSGender (M/F/O)17/30/213/28/615/33/2Total N494750Age (Mean ± SD)20.59 ± 1.9221.15 ± 2.1920.16 ± 1.83Dominant Hand (R/L)45/439/749/1PHQ-9 Criteria≤4≥10≥10SIS Criteria≤16≤16> 16For Gender, M - Male, F - Female, O - Other. PHQ-9 and SIS thresholds were used to group participants into Healthy controls (C), Depressed non-suicidal (D), and Depressed suicidal (S). Participants with PHQ-9 scores between 5-9 were excluded. Average age (years) with standard deviation is reported. Dominant hand indicates R = right, L = left. Table 1Demographics of the three mental health groups in the study.CDSGender (M/F/O)17/30/213/28/615/33/2Total N494750Age (Mean ± SD)20.59 ± 1.9221.15 ± 2.1920.16 ± 1.83Dominant Hand (R/L)45/439/749/1PHQ-9 Criteria≤4≥10≥10SIS Criteria≤16≤16> 16For Gender, M - Male, F - Female, O - Other. PHQ-9 and SIS thresholds were used to group participants into Healthy controls (C), Depressed non-suicidal (D), and Depressed suicidal (S). Participants with PHQ-9 scores between 5-9 were excluded. Average age (years) with standard deviation is reported. Dominant hand indicates R = right, L = left. Demographics of the three mental health groups in the study. For Gender, M - Male, F - Female, O - Other. PHQ-9 and SIS thresholds were used to group participants into Healthy controls (C), Depressed non-suicidal (D), and Depressed suicidal (S). Participants with PHQ-9 scores between 5-9 were excluded. Average age (years) with standard deviation is reported. Dominant hand indicates R = right, L = left. Sentence Response Task In the study, participants took part in a sentence response task with multimodal physiological signal recording. For this analysis, we will focus on the EEG data. During the task, sentences presented to the participants were designed to contain 160 self-referential sentences with 80 minimal pairs such that each sentence in a pair differed in last word only. Within each pair, one sentence was designed such that its critical final word is likely congruent to the beliefs and/or experiences of healthy participants (My mental health is not problematic - neutral or positive sentiment sentences) and the other incongruent for healthy participants, and congruent for participants having compromised mental health (My mental health is not sound - negative sentiment sentences). Each word of the sentence was displayed sequentially at the center of a screen for 300 ms followed by an inter-word-interval (IWI) of 300 ms (blank screen) with the last critical word of the sentence being displayed for 600 ms followed by a prompt to whether they agree or disagree with the sentence. Participants were recommended to avoid responding until the end of the sentence in order to reduce impact of motor responses during sentence reading. Prior to presentation of the next sentence, a fixation cross at the center of the screen was displayed. Figure 1A provides a detailed pictorial representation of the recording setup and task design.Fig. 1Overview of task design, ERP computation and group decoding.A Sentences are displayed one-word-at-a-time on a computer screen. As depicted in the image, each word is shown for 300 ms followed by an Inter Word Interval (IWI) of 300 ms blank screen. The last word in the sentence is displayed for 600 ms followed by 300 ms IWI following which the subject is prompted to respond with agree or disagree by pressing the respective button. Each sentence is shown twice in a pseudo random order amounting to 320 trials per subject. Parts B, C provide details of group decoding. B ERP of a specific sentence grouping (for example positive trials) was computed per subject. C After ERP computation, the participants data at every timestep are passed to a classification pipeline wherein Leave One Subject Out Cross Validation is performed with a logistic regression classifier training per timestep per train test split. Overview of task design, ERP computation and group decoding. A Sentences are displayed one-word-at-a-time on a computer screen. As depicted in the image, each word is shown for 300 ms followed by an Inter Word Interval (IWI) of 300 ms blank screen. The last word in the sentence is displayed for 600 ms followed by 300 ms IWI following which the subject is prompted to respond with agree or disagree by pressing the respective button. Each sentence is shown twice in a pseudo random order amounting to 320 trials per subject. Parts B, C provide details of group decoding. B ERP of a specific sentence grouping (for example positive trials) was computed per subject. C After ERP computation, the participants data at every timestep are passed to a classification pipeline wherein Leave One Subject Out Cross Validation is performed with a logistic regression classifier training per timestep per train test split. Trial Grouping The set of 160 sentences displayed as a part of the sentence task vary across a variety of categories – (1) stimuli characteristics: sentiment of the sentence, valence of the last word and (2) Response Characteristics: response time of the participants, agree/disagree to the sentence. Based on the sentiment of the sentence, trials were grouped into positive, neutral and negative sentiment sentences. Similarly, based on the valence of the last word, trials were grouped into positive last word valence and negative last word valence trials. Response type grouping referred to grouping the sentences into agree and disagree sentences based on the response of the participants. Based on the response time to trials, they were grouped into high response time (top 25% reaction time) or low response time (bottom 25% reaction time) groups. Lastly, in order to leverage the combined information in the positive and negative groupings within a category, the trials were subtracted to form a contrasting grouping i.e., in the case of sentiment sentence contrasting group was defined by subtracting the positive and negative sentiment trials. In order to study the groups of sentences that provided the most discriminability to classify the participants into the underlying groups, individual trials per subject were grouped based on the aforementioned characteristics. These grouped trials were then subsequently used in Multivariate Pattern Analysis (MVPA) and deep learning pipelines. EEG Group Decoding Multi Variate Pattern Analysis (MVPA) [41, 42], has been widely used in EEG and MEG studies to understand the temporal dynamics of visual processing [43, 44]. In this work, we utilize MVPA to study the temporal regions within the neural responses that produce distinguishable signatures in healthy and depressed participants. ERPs of participants based on the aforementioned trial groupings were computed (positive sentiment sentences, agree sentences, or low reaction time sentences). The ERPs were time averaged with a window of 10ms and these inputs were passed to a classification pipeline for C vs D,S classification in a Leave One Subject Out Cross Validation (LOSOCV) setting across 10 random seeds (refer to Fig. 1B, C for visualization of decoding pipeline). The classification pipeline is composed of logistic regression model with L1 regularization. Area Under Receiver Operating Characteristics (AUC) is reported as evaluation metric. All experiments contain 2 classes in classification, and the chance level of AUC is 0.5. Deep Learning Experiments Deep learning models were trained to distinguish healthy controls from depressed participants (C vs D, S) and sub-group classification within depressed participants (D vs S). For each category of trial grouping, three types of inputs positive, negative and contrasting inputs (subtracting the positive and negative time series trials) were defined and distinct models were trained for each input type. The inputs to the models composed of bootstrapped trials generated from aforementioned trial grouping using Algorithm 1. The inputs were bootstrapped in order to reduce the intra-subject variability and enhance the common characteristics of the particular trial grouping. For each subject, B = 20 single-trial EEG responses were sampled with replacement and averaged across the time dimension to generate a single bootstrapped trial. This procedure was repeated N = 200 times, resulting in 200 bootstrapped trials per subject. These bootstrapped trials from all training participants were used to train a classifier to predict the mental health group (e.g., C vs DS or D vs S). During testing, the classifier produced one prediction per bootstrapped trial, yielding N predictions per test subject. The subject-level probability of being classified as depressed was computed as the ratio of trials predicted as “depressed” to the total number of bootstrapped trials (i.e., 200). A higher probability indicates a greater likelihood of the subject belonging to the depressed group. The final classification performance was measured using Area Under the Receiver Operating Characteristic curve (AUC). All models were trained in a k-fold (k = 5) between subject cross-validation setting with 5 randomized weight initializations. For additional details regarding the hyperparameters of the model and pre-processing pipeline, refer to the supplementary material. Algorithm 1 Bootstrapped Subject-Level EEG Classification Require: EEG trials per subject \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{T}}}_{s}\in {{\mathbb{R}}}^{{N}_{t}\times T\times C}$$\end{document}Ts∈RNt×T×C where Nt is number of trials, T timepoints, C channels 1: Set number of bootstrap samples N ← 200 2: Set trials per sample B ← 20 3: for each training subject sdo 4: fori = 1 to Ndo 5: Sample B trials with replacement from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{T}}}_{s}$$\end{document}Ts: {x1, …, xB} 6: Compute averaged trial: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{{\bf{x}}}}_{i}\leftarrow \frac{1}{B}{\sum }_{j=1}^{B}{{\bf{x}}}_{j}$$\end{document}x¯i←1B∑j=1Bxj 7: Store training pair \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$({\overline{{\bf{x}}}}_{i},{y}_{s})$$\end{document}(x¯i,ys) 8: end for 9: end for 10: Train classifier \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{M}}$$\end{document}M on all bootstrapped training pairs 11: Select best model via validation split from training data 12: for each test subject \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s{\prime}$$\end{document}s′do 13: fori = 1 to Ndo 14: Generate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{{\bf{x}}}}_{i}$$\end{document}x¯i via bootstrapping as above 15: Predict: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{y}}_{i}\leftarrow {\mathcal{M}}({\overline{{\bf{x}}}}_{i})$$\end{document}y^i←M(x¯i) 16: end for 17: Compute subject probability: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{s{\prime} }\leftarrow \frac{1}{N}{\sum }_{i=1}^{N}{\mathbb{I}}[{\widehat{y}}_{i}=\,depressed\,]$$\end{document}Ps′←1N∑i=1NI[y^i=depressed] 18: end for 19: Compute AUC over all test participants using \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{s{\prime} }$$\end{document}Ps′ and ground-truth labels Results Temporal Dynamics of Sentence Response Task For this analysis, within-subject Multivariate Pattern Analysis (MVPA) [41, 42] was adapted to a between-group framework. This enabled investigating the temporal windows during which neural responses of healthy controls (C) differ from those of depressed non-suicidal (D) and depressed suicidal (S) participants. To identify the types of affective stimuli that reveal larger differences in neural responses, group-level decoding was performed for the trials grouped based on their affective sentiment: (1) all sentences, (2) positive sentiment sentences, (3) neutral sentiment sentences, (4) negative sentiment sentences and (5) contrasting sentences (positive sentiment sentences - negative sentiment sentences). The trials corresponding to sentence presentation were epoched from 200ms before to 900ms after onset of the last word in the sentence, which excludes the time window corresponding to participant response for the displayed sentence. For each aforementioned sentence group, classification pipeline as mentioned in Fig. 1C was applied. Group-level decoding is performed between healthy (C) and depressed participants (D, S), with AUC being reported as the evaluation metric. All experiments are conducted in a between-subjects manner, with the average AUC being reported for a Leave-One-Subject-Out Cross-Validation (LOSOCV) setting across 10 random seeds. Temporal clusters with decoding metrics significantly better than chance were computed through a non parametric cluster-based permutation test (Refer to Supplementary material Statistical Analysis Section for additional details.) While prior studies have reported group differences in processing positive- or negative-sentiment sentences [30–34], here we observe significant group differences between healthy controls and depressed participants for all sentences group (positive + neutral + negative sentiment sentences) between (544 ms, 900 ms) from onset of the last word presentation in the sentence as seen in Fig. 2A. This indicates a baseline difference in neural signals between healthy and depressed participants, potentially arising from differences in underlying cognitive processing during integration of meanings over time and context, and decision making [45–48], as the region of significance (544 ms to 900 ms) corresponds to the interval immediately preceding response prompt in which participants indicate whether they agree or disagree with the sentence. Channel importance maps in the statistically significant regions indicate group differences arising from posterior regions of the scalp with bias towards left hemisphere.Fig. 2Temporal plots for per-timestep average AUC for Control versus Depressed non-suicidal & Depressed suicidal [C vs D,S] classification in a LOSOCV setting across 10 different seeds for.(A) All sentences (B) Positive sentiment sentences (C) Neutral sentiment sentences and (D) Negative sentiment sentences. Shaded regions correspond to clusters wherein the performance were found to be statistically significantly above chance from a cluster based permutation test (p ≤ 0.05). The histogram plot and channel importance map of the largest cluster in each setting are reported. Channel Importance Map: Each value represents the proportion of significant timepoints during which the corresponding electrode weight belonged to top 10% percentile of weights in L1-regularized logistic regression, indicating its contribution to group classification. Darker regions in the channel importance map indicate the specific region to have higher contribution to C vs D,S classification. Temporal plots for per-timestep average AUC for Control versus Depressed non-suicidal & Depressed suicidal [C vs D,S] classification in a LOSOCV setting across 10 different seeds for. (A) All sentences (B) Positive sentiment sentences (C) Neutral sentiment sentences and (D) Negative sentiment sentences. Shaded regions correspond to clusters wherein the performance were found to be statistically significantly above chance from a cluster based permutation test (p ≤ 0.05). The histogram plot and channel importance map of the largest cluster in each setting are reported. Channel Importance Map: Each value represents the proportion of significant timepoints during which the corresponding electrode weight belonged to top 10% percentile of weights in L1-regularized logistic regression, indicating its contribution to group classification. Darker regions in the channel importance map indicate the specific region to have higher contribution to C vs D,S classification. When performing group decoding with sentences grouped based on sentiment, a significant region of decodability is observed around 500ms after onset of the final word for positive (Fig. 2B), neutral (Fig. 2C), and negative (Fig. 2D) sentiment sentence groupings. In addition to the late component, neutral and negative sentences show an early component of significant decodability –from 100 ms to 183 ms and 226 ms to 290 ms respectively–as shown in Fig. 2C, D. For neutral and positive sentences, anterior and posterior channels show a greater proportion of importance during classification. This could indicate differences in late positive potential, which has been found in prior studies to originate as a result of differences in predictability of words, with distinct components in the anterior and posterior regions [49, 50]. Such anterior and posterior bias in channel importance map is not observed for negative sentences. Additionally, peak AUC value and the statistically significant cluster sizes for positive and neutral sentences are noticeably higher compared to negative sentences, suggesting that group differences in ERPs are less pronounced for negative sentence processing. When contrasting (subtracting positive and negative sentiment sentence ERPs), midline electrodes in anterior channels which are primarily responsible for emotional processing provide the highest distinguishability (Fig. 3). The temporal cluster with statistically significant decodability is from (256 ms to 657 ms), revealing earlier distinct responses compared to individual groups of sentences.Fig. 3Decoding plot for Control versus Depressed non-suicidal & Depressed suicidal [C vs D,S] classification with contrasting trials (positive sentiment - negative sentiment). Timesteps with statistically significant performance above chance are shaded. Channel Importance map indicates anterior electrodes to provide differentiating responses between healthy and depressed participants. Decoding plot for Control versus Depressed non-suicidal & Depressed suicidal [C vs D,S] classification with contrasting trials (positive sentiment - negative sentiment). Timesteps with statistically significant performance above chance are shaded. Channel Importance map indicates anterior electrodes to provide differentiating responses between healthy and depressed participants. Trial Grouping Based Depression Classifier While group decoding analyses primarily emphasize the temporal dynamics of neural signals, they do not allow for the models to leverage the spatiotemporal information of EEG responses. Hence, deep learning models were trained to classify mental health groups (Healthy (C), Depressed non-suicidal (D) and Depressed suicidal (S)). The EEG signals from 200 ms before to 900 ms after the onset of the final word of the sentence were used as input to the model, which corresponds to time window prior to participant response. These time regions for trials were chosen so as to minimize influence of response and capture neural signals corresponding to sentence reading. Classification results using the aforementioned model are presented in Table 2, across 4 broad categories of sentence grouping: (1) Sentence sentiment, (2) last word valence, (3) response type (agreement versus disagreement) and (4) response time of participant. The baseline category contains two conditions, with all sentences grouping the sentence set into a single group, while random contrasting divides the sentences into two groups at random and then generates inputs by subtracting the responses from the two groups. Models are trained for two classification tasks, (1) C vs (D, S) for distinguishing healthy and depressed participants, and (2) D vs S, aiming to disentangle depressed non-suicidal (D) from depressed suicidal (S) participants.Table 2AUC and 95% Confidence Intervals (CI) for EEG classification across sentence grouping conditions and trial types.ConditionTrial TypeC vs DSD vs SSentence SentimentPositive0.700 [0.60, 0.80]**0.536 [0.41, 0.66]Negative0.557 [0.45, 0.66]0.607 [0.49, 0.73]*Contrasting0.720 [0.61, 0.82]**0.622 [0.50, 0.73]*Last Word ValencePositive0.599 [0.50, 0.70]*0.589 [0.47, 0.71]Negative0.633 [0.53, 0.73]**0.570 [0.45, 0.69]Contrasting0.624 [0.53, 0.71]*0.542 [0.42, 0.67]Response TypeAgree0.657 [0.55, 0.75]**0.550 [0.42, 0.68]Disagree0.615 [0.51, 0.71]*0.552 [0.43, 0.67]Contrasting0.699 [0.60, 0.79]**0.628 [0.51, 0.75]*Response TimeSlow0.656 [0.55, 0.75]*0.535 [0.41, 0.66]Fast0.622 [0.52, 0.72]*0.579 [0.45, 0.70]Contrasting0.672 [0.58, 0.77]**0.650 [0.53, 0.75]*BaselineAll Sentences0.577 [0.47, 0.68]0.538 [0.41, 0.67]Random Contrasting0.643 [0.55, 0.73]*0.600 [0.43, 0.72]Statistical significance was determined using permutation tests against the null hypothesis of AUC = 0.5, with * indicating p < 0.05 and ** indicating p ≤ 0.001. For the baseline rows, All Sentences corresponds to grouping all sentences together, while Random Contrasting splits the sentences randomly into two groups.Numbers in bold indicate the best performing models for C vs DS and D vs S. Table 2AUC and 95% Confidence Intervals (CI) for EEG classification across sentence grouping conditions and trial types.ConditionTrial TypeC vs DSD vs SSentence SentimentPositive0.700 [0.60, 0.80]**0.536 [0.41, 0.66]Negative0.557 [0.45, 0.66]0.607 [0.49, 0.73]*Contrasting0.720 [0.61, 0.82]**0.622 [0.50, 0.73]*Last Word ValencePositive0.599 [0.50, 0.70]*0.589 [0.47, 0.71]Negative0.633 [0.53, 0.73]**0.570 [0.45, 0.69]Contrasting0.624 [0.53, 0.71]*0.542 [0.42, 0.67]Response TypeAgree0.657 [0.55, 0.75]**0.550 [0.42, 0.68]Disagree0.615 [0.51, 0.71]*0.552 [0.43, 0.67]Contrasting0.699 [0.60, 0.79]**0.628 [0.51, 0.75]*Response TimeSlow0.656 [0.55, 0.75]*0.535 [0.41, 0.66]Fast0.622 [0.52, 0.72]*0.579 [0.45, 0.70]Contrasting0.672 [0.58, 0.77]**0.650 [0.53, 0.75]*BaselineAll Sentences0.577 [0.47, 0.68]0.538 [0.41, 0.67]Random Contrasting0.643 [0.55, 0.73]*0.600 [0.43, 0.72]Statistical significance was determined using permutation tests against the null hypothesis of AUC = 0.5, with * indicating p < 0.05 and ** indicating p ≤ 0.001. For the baseline rows, All Sentences corresponds to grouping all sentences together, while Random Contrasting splits the sentences randomly into two groups.Numbers in bold indicate the best performing models for C vs DS and D vs S. AUC and 95% Confidence Intervals (CI) for EEG classification across sentence grouping conditions and trial types. Statistical significance was determined using permutation tests against the null hypothesis of AUC = 0.5, with * indicating p < 0.05 and ** indicating p ≤ 0.001. For the baseline rows, All Sentences corresponds to grouping all sentences together, while Random Contrasting splits the sentences randomly into two groups. Numbers in bold indicate the best performing models for C vs DS and D vs S. Majority of sentence grouping conditions result in classifiers with statistically greater than chance performance for C vs D,S classification, with sentence-sentiment contrasting models reaching an AUC of 0.720 [95% CI = (0.61, 0.82)]. Depressed sub-group classification models (D vs S) resulted in more modest performance with response time contrasting models providing an AUC of 0.650 [95% CI = (0.53, 0.75)]). When compared to baseline model (AUC=0.577), training on positive sentiment sentences (AUC=0.700) or response type agreement sentences (AUC=0.657) conditions provide performance gains for C vs D,S classification, while for D vs S classification negative sentiment sentences (AUC=0.607) provide marginally better performance over baseline (AUC=0.538) indicating the utility of grouping sentences based on the linguistic and response characteristics of the stimuli. It is worth noting that, unlike a random-chance or majority-voting baseline, the “all sentences" baseline model captures neural signal differences between mental health groups in the sentence response task without relying on prior knowledge of stimulus type. Hence, any statistically significant performance over this method indicates the utility of trial grouping. Contrasting involves subtracting the trial types within a category of sentence grouping, e.g., for sentence sentiment the positive and negative sentiment trials were subtracted. For contrasting models of sentence sentiment, response type and response-time grouping, performance gains are observed relative to their respective individual trial types in both for C vs D,S and D vs S classification tasks. This could be due to complementary information when subtracting opposite categories of a condition (positive sentiment vs negative sentiment or agree vs disagree). Within the individual trial type of sentence grouping, positive sentiment sentences provide the best distinguishability for C vs D,S classification. This is consistent with similar findings in prior work wherein reduced activity in a depressed population was observed when processing positive sentiment words during a working memory task, but no such differences were observed for negative sentences [31]. Additionally, this observation is in accordance with the group decoding results wherein positive sentiment sentences provided higher AUC and larger statistically significant cluster sizes over negative sentiment sentences. Interestingly, while positive sentiment sentences provide better distinguishability for C vs D,S classification, negative sentiment sentences show better performance for D vs S indicating the utility of both positive and negative sentiment sentences. With regard to comparison between sentence sentiment-based and last-word valence-based groupings, a consistent trend of greater discriminability is observed for sentence sentiment. This could be owing to task design emphasizing on participants requiring to respond to the displayed sentence, resulting in the sentence sentiment being more effective at eliciting consistent differences. When comparing response type based grouping and sentence sentiment based, no consistent performance gains are observed for one condition type. However the response type grouping requires explicit feedback from participants whereas the sentence sentiment is based entirely on the semantic characteristics of the sentence stimuli. This could be promising for future experimental approaches that do not necessitate an explicit feedback from the participants. Interpretability Analysis Temporal & Spatial Regions of Discriminability provide interpretable information highlighting the time regions and electrodes of significance for distinction between the groups. To this end, ablations on the two best performing models for C vs D,S classification models (sentence sentiment contrasting, response type contrasting) are plotted in Fig. 4. For temporal ablation, the input to the deep learning model was constrained to (-200 ms to X ms) with X ∈ {0, 150, 300, 450, 600, 750, 900}. The 900 ms input signal contains the entire duration of last word presentation and the inter-word interval of 300 ms before the subject responds to the sentence; whereas 0 ms input signal includes 200 ms prior to the presentation of the last word, which corresponds to the inter-word interval of the penultimate in the sentence. Both models show an expected increase in AUC magnitude as the length of the input signal increases. For the response type model, a saturation in performance is observed around 600 ms, whereas an increasing performance trend is seen in the sentence sentiment model up to 750 ms. In spatial localization ablation, the sentence sentiment contrasting model shows similarities to decoding experiments (sentence sentiment contrasting decoding experiment wherein channel importance map [refer to Fig. 3] indicates differences arising from anterior channels) with anterior channels providing better classification. For the response type contrasting model, central and anterior channels provide better differentiation. These spatial ablation results provide promising directions for considering the reduction in the number of electrodes used in data collection for future studies.Fig. 4Performance when varying the temporal and channel inputs to the model.LEFT: Temporal performance with input signal as (-200 ms to X ms); X ∈ 0, 150, 300, 450, 600, 750, 900. Performance increases as the input size increases and saturates around 600 ms. RIGHT: Performance variation when varying the input channels across different electrode regions. Anterior channels show better performance across both conditions, while central channels show importance in the response type model. Performance when varying the temporal and channel inputs to the model. LEFT: Temporal performance with input signal as (-200 ms to X ms); X ∈ 0, 150, 300, 450, 600, 750, 900. Performance increases as the input size increases and saturates around 600 ms. RIGHT: Performance variation when varying the input channels across different electrode regions. Anterior channels show better performance across both conditions, while central channels show importance in the response type model. Generalizability to Unseen Subgroup is an essential requirement for clinical diagnostic models that require translation across heterogeneous patient populations. In the context of depression, which encompasses multiple subtypes and symptom profiles this generalizability is critical for developing tools that are robust and widely applicable. Such models support the development of scalable, front-line screening tools that do not require fine-tuning for each clinical subtype, thereby enhancing real-world applicability. To evaluate this, we assessed how well models trained on one clinical subgroup generalized to another. Specifically, we trained separate models for (C vs D) and (C vs S) classifications, where either depressed suicidal or depressed non-suicidal individuals were excluded from the training set, respectively. The performance of these models was then tested on the held-out subgroup by examining the average probability that the model assigned to depressed class labels for the unseen group. Figure 5 reports the average depressed probabilities predicted by the model for test participants in the three groups, with the plot on the left trained for (C vs D) and the model on right is trained on (C vs S). The (C vs S) model predicted an average depressed probability of 0.72 for depressed non-suicidal participants (D), while the (C vs D) model assigned an average depressed probability of 0.61 to depressed suicidal participants (S). Importantly, for both models, the predicted probabilities for the unseen depressed subgroup were significantly higher than those for healthy participants, yet no significant differences were found between the two depressed subgroups, suggesting consistency in the neural signatures of depression across subtypes. Further, the (C vs D) and (C vs S) models achieved AUCs of 0.72 [95% CI: 0.68-0.76] and 0.69 [95% CI: 0.64-0.73], respectively, when evaluated on the full (C vs D,S) task. These results support the robustness of the models to missing clinical subgroups during training, underscoring their potential for broader deployment in diverse patient populations where detailed subtype labeling may not be available.Fig. 5Boxplot of prediction probabilities of models across the three groups of participants.Left: Model trained for C vs D; Right: Model trained for C vs S. At testing time, for both models all the three groups (C, D and S) are predicted. Statistical tests indicate the model probabilities to be statistically different (p < 0.05 after bonferroni correction for multiple t-tests) between C and D, C and S for both the models indicating generalization of predictions for D and S groups. Boxplot of prediction probabilities of models across the three groups of participants. Left: Model trained for C vs D; Right: Model trained for C vs S. At testing time, for both models all the three groups (C, D and S) are predicted. Statistical tests indicate the model probabilities to be statistically different (p < 0.05 after bonferroni correction for multiple t-tests) between C and D, C and S for both the models indicating generalization of predictions for D and S groups. Correlations of Predictions to Questionnaires: PHQ-9 scores of the participants are recorded twice, first during the recruitment screening phase and second on day of EEG recording. A higher correlation between model depression probabilities and PHQ-9 score on the day of the EEG recording (0.327) was observed as compared to the PHQ-9 from the earlier screening (0.243), and this difference is found to be statistically significantly based on a permutation test (p = 0.003). While a positive correlation with PHQ-9 score across groups is expected owing to the groups being defined based on the score, a higher score with the PHQ-9 on the day of the screening could indicate that the model depression probabilities are legitimately capturing within group effects. In addition to the PHQ-9 questionnaire, SIS [38] and GAD-7 [39] questionnaires are completed by the participants during screening and on the day of recording respectively. Plots for within-group correlations for SIS and GAD-7 scores can be seen in Fig. 6. Positive correlation between model depression probability and questionnaire score is observed for controls and suicidal participants in GAD-7 score and for suicidal participants in SIS score. While the trends show positive correlation, these effects are not found to be statistically significant. However, the ability of the models to assign higher depressed probability to participants with higher questionnaire scores without explicitly providing this information could indicate that the models are implicitly able to model symptom severity.Fig. 6Correlations of contrasting sentence sentiment model probabilities for C vs D, S task with Left: GAD-7 and Right: SIS questionnaires. Spearman correlation is reported per group and the p-values test against a null distribution of zero correlation. Correlations of contrasting sentence sentiment model probabilities for C vs D, S task with Left: GAD-7 and Right: SIS questionnaires. Spearman correlation is reported per group and the p-values test against a null distribution of zero correlation. Discussion In this study, we investigate whether neural responses to emotional self-referential stimuli can serve as reliable indicators for distinguishing healthy individuals from individuals with depressive symptoms. While prior research has demonstrated group-level differences in event-related potentials (ERPs) to emotional stimuli in depression [30, 31, 33], these findings have not yet been systematically evaluated for their utility in developing objective, data-driven screening tools. Our study makes a twofold contribution: first, by examining the temporal dynamics of neural responses to affective self-referential sentences using multivariate pattern analysis (MVPA); and second, by evaluating the feasibility of deep learning-based classification approaches for potential use in practical screening applications. MVPA results reveal that group differences in neural responses likely reflect a combination of emotional processing and decision-related cognitive mechanisms. We observe baseline group differences in response to self-referential sentences irrespective of the sentiment of the sentence. These differences predominantly emerge at latencies ≥ 500 ms, coinciding with the conclusion of the sentence presentation and immediately preceding the behavioral prompt (agree/disagree). This timing suggests that the observed group differences may be driven by decision-making processes and associated higher-order cognitive functions engaged during evaluative judgment. On the other hand, for group decoding with contrasting (positive - negative sentiment sentence) condition, we identify earlier neural differences between groups in the 250-650 ms window (congruent with ERP findings in [51] that indicate reduce N400 amplitude to self-referential negative sentiment sentences). These differences are localized in anterior scalp regions, consistent with known correlates of emotional and affective processing [52]. Crucially, the absence of significant group differences beyond 650 ms in this condition implies that later-stage neural differences are not modulated by sentence sentiment. Instead, they appear to reflect general decision making processes which are common across stimuli. The early group differences, however, likely reflect altered attentional or emotional responses triggered by the presentation of the affect-critical final word in the sentence, aligning with prior work on affective sentence processing [53]. Together, these findings suggest temporally and functionally dissociable neural markers of depression symptoms: early activity linked to emotion-specific stimulus processing and later activity related to decision-making. The results of our deep learning models underscore the potential of EEG as a viable modality for identifying neurophysiological biomarkers of depression. Achieving an AUC of 0.72 in distinguishing healthy individuals from depressed participants demonstrates that temporally dynamic neural responses to self-referential emotional stimuli encode discriminative patterns relevant to reported depression symptoms. While the sensitivity of the model is relatively high (77%), the lower specificity (59%) suggests that a subset of healthy individuals are misclassified as depressed. This misattribution may in part stem from the class imbalance in the dataset, where the proportion of depressed participants was larger. In addition to the overall classification performance, the spatial and temporal ablation analyses offer important practical insights. Specifically, they suggest the possibility of reducing the number of electrodes required without substantially compromising model accuracy. This opens the door to more portable and clinically viable EEG configurations that minimize setup time and participant burden. A major consideration for the translational application of our approach, however, is the experimental duration. With a task length averaging 55 minutes, the current protocol may be too lengthy for routine clinical deployment. To evaluate whether data collection could be made more efficient, we examined the effect of reducing the number of trials per subject. Two ablation strategies were tested: one reducing trials during both training and testing, and another simulating a realistic deployment scenario in which a new subject is evaluated with a reduced trial set while training is conducted on full-length data. The results reveal that while reducing trial numbers across both training and testing leads to performance degradation (Fig. 7Left), limiting the number of trials only at test time (Fig. 7Middle) preserves model robustness even with up to a 75% reduction. These findings suggest that broad stimulus variability during training is critical, but efficient screening at inference time may still be possible with fewer trials, offering a path toward shortened and more scalable EEG-based assessments. Given the practical challenges in acquiring large-scale clinical EEG datasets especially from depressed populations, we also evaluated model performance under low-resource conditions. Remarkably, the sentence sentiment model maintained stable classification accuracy even when the number of depressed participants in training was reduced by up to 75%, while the response type model tolerated reductions of up to 50% with minimal impact as seen in Fig. 7Right. These findings indicate that the models can extract meaningful discriminative features from relatively small samples of individuals depression symptoms, which has important implications for the feasibility of training such models in real-world contexts where participant recruitment is often limited.Fig. 7Ablations on number of trials presented per subject and the number of depressed participants in training data for C vs D,S task.Left: Reduction in trials in both train and test data. models show limited ability to learn with reduced trials in training data. Middle: When reducing the trials at testing time only, robustness to reduction in trials up to 75% is observed for response type model. Right: Ablation reducing the percentage of depressed participants used. Sentence sentiment models indicate the ability to learn discriminating features under low depressed subject count. Ablations on number of trials presented per subject and the number of depressed participants in training data for C vs D,S task. Left: Reduction in trials in both train and test data. models show limited ability to learn with reduced trials in training data. Middle: When reducing the trials at testing time only, robustness to reduction in trials up to 75% is observed for response type model. Right: Ablation reducing the percentage of depressed participants used. Sentence sentiment models indicate the ability to learn discriminating features under low depressed subject count. To benchmark neural performance against a simpler behavioral baseline, we trained a logistic regression classifier using each participant’s response profile–namely, the number of agree versus disagree decisions to positive and negative sentences. This behavioral model achieved near-ceiling AUCs (0.94 for C vs D,S and 0.88 for D vs S), suggesting strong group-level differences in explicit evaluative responses. However, this approach may be less robust in practice, particularly in settings where participants are unwilling or unable to respond transparently due to stigma, discomfort, or motivational deficits. An EEG-based model trained to predict the subject’s behavioral response achieved only chance-level performance (AUC = 0.552 ± 0.005), indicating that the neural classifier is not merely capturing response patterns but is instead leveraging distinct neural signatures associated with depression. Together, these results highlight the promise of EEG-based models for depression screening and shed light on practical considerations regarding trial duration, data efficiency, and model robustness under real-world constraints. Limitations The primary limitation of the current work is its limited immediate clinical applicability, as it achieves an AUC of 0.72 for distinguishing between healthy and depressed individuals. In order for establishing a practical screening tool, the discriminative power needs to be substantially improved through better experimental design and modelling approaches. Additionally, models that are able to provide consistent performance across participants agnostic to training splits should be explored in future works. The current study employs PHQ-9 scores in order to group individuals into healthy, depressed and suicidal groups which could mean that the identified neural signals could be measuring correlates of symptoms experienced by participants. Hence, future work should include formal clinical diagnosis of participants through a structured psychiatric interview. Validation on larger and more diverse cohorts spanning different age groups and co-morbidities is also necessary to establish clinical generalizability. Additionally, the sentences presented to the participants were derived from validated survey tools that address known associations with depression/suicidality, but could limit the overall heterogeneity of sentence meanings which may cause exposure effects. While interpretability in the sensor space is provided through spatial ablations, future work could perform source space reconstruction that would improve explainability and allow for anatomical interpretations. Moreover, refining these methods for use in time-constrained and resource-limited clinical settings would enhance translational relevance. Supplementary information Supplementary Material Supplementary Material