{"id":"de54b780-c4c9-4557-9394-9e274fe7d22c","arxiv_id":"2506.06244","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The EEG response to self-referential sentences separates healthy from depressed participants with an AUC of 0.707 and depressed from suicidal participants with an AUC of 0.624.","lead":"People with depression show different brain-wave patterns while reading short self-referential sentences, and a deep learning model tells them apart from healthy people with a score of 0.71. The work suggests EEG could support depression screening and even help distinguish depressed people with and without suicidal thoughts.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline AUCs are valid only if the five-fold CV splits by participant; §4.6 does not state this, and Algorithm 1's bootstrap makes the distinction consequential.","rationale":"The paper has real strengths: the temporal decoding uses explicit LOSOCV, the all-sentence and random-contrast baselines provide context, and the behavioral-response control (chance-level EEG prediction of the subject's own response) argues against a simple response-pattern confound. None of these controls, however, resolves the one ambiguity that matters most for the deep-learning AUCs: whether the five-fold cross-validation in Section 4.6 is participant-disjoint. Algorithm 1's language suggests it is, but the split construction is described in a single phrase and the code is not public. If trial-level leakage occurred, the 0.707 and 0.624 AUCs could reflect within-subject EEG correlations rather than group-level depression signatures. That is exactly the reader's weakest assumption, and I agree with it. A secondary internal inconsistency worth noting but not the primary stress point: Section 4.3 says EEG data from 6 participants were excluded during preprocessing, which would leave 140 rather than the 146 listed in Table 2; this should be clarified but does not by itself overturn the central result. Given that the leakage concern is addressable by inspecting or rerunning the CV, the appropriate verdict remains CONDITIONAL.","tokens_in":17305,"tokens_out":9355,"duration_ms":99195,"concrete_test":"Request the released code (or ask the authors for the exact CV splitter used for Table 1) and verify that no trial from a participant in a test fold appears in any training bootstrap. Then recompute the Sentence Sentiment Contrasting CvsDS AUC with strict Leave-One-Subject-Out CV and compare to 0.707; if the LOSO AUC drops below about 0.6 or its CI includes 0.5, the reported value reflects leakage. Also repeat the same check for the DvsS Response Time Contrasting result (0.624).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central classification claim (abstract, Table 1) is that a deep learning model distinguishes C vs D,S with AUC 0.707 and D vs S with AUC 0.624. For these numbers to estimate generalization to new people, the five-fold cross-validation must split participants, not trials. Section 4.6 says only 'a five-fold cross validation stratified for gender were used,' and Algorithm 1 describes per-subject bootstrapping (N=200 averaged trials per subject, B=20 trials per average) without ever stating that all trials of a test participant are excluded from training. If folds are trial-level, or if the same participant contributes trials to both train and test, then each bootstrapped train and test input shares single-trial EEG from the same person. Because EEG contains strong subject-specific noise, the classifier can exploit within-subject correlations rather than depression-related signal, inflating the reported AUC. The small sample (146 participants, roughly 117 training subjects per fold) makes this especially consequential. Algorithm 1's wording ('training subjects' vs 'test subjects') suggests subject-level split was intended, but the actual fold construction is not described, and code/data are only 'available upon request.' This is the load-bearing uncertainty: if the split is subject-disjoint, the result stands as reported; if not, the AUCs are uninterpretable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates whether EEG responses to self-referential affective sentences contain information about depression status and suicidal ideation. It reports group-level temporal decoding results using MVPA, deep learning classification results (AUC 0.707 for healthy vs. depressed, AUC 0.624 for depressed non-suicidal vs. suicidal under specific trial groupings), spatial/temporal ablations, subgroup generalization analyses, and correlations with questionnaire scores. The central claims are that stimulus-driven neural signatures of depression are present in the EEG and that a deep learning model can exploit them for classification. The methods involve PHQ-9/SIS-based grouping, a sentence-response task with 160 sentences, EEG preprocessing, bootstrapped trial generation, and five-fold cross-validated transformer-based classifiers.","tokens_in":17587,"tokens_out":4697,"duration_ms":51351,"significance":"If the classification results are valid, this is a worthwhile contribution to EEG-based depression screening and to understanding the temporal dynamics of self-referential affective processing. The study has several strengths: subject-level bootstrapped classification, permutation-based significance testing for the decoding analyses, confidence intervals for AUCs, a behavioral-confound control (EEG does not predict agree/disagree), and ablation experiments for trials, bootstrap parameters, and training-set composition. The main caveat is that the validity of the headline AUCs depends critically on the cross-validation splitting participants rather than trials; the manuscript strongly implies this but does not state it explicitly, and the data/code are not publicly available. There is also a direct inconsistency in the reported number of participants that must be resolved.","major_comments":[{"comment":"The central generalization claim depends on the five-fold cross-validation partitioning participants, but §4.6 only states 'a five-fold cross validation stratified for gender' and does not explicitly say that all trials of a test participant are excluded from training. Algorithm 1's 'training subjects'/'test subjects' wording strongly suggests a subject-level split, and the bootstrap procedure operates within subject; nevertheless, because single-trial EEG is strongly subject-correlated, a trial-level or within-subject split would inflate the reported AUCs in Table 1. Please state the fold-construction rule explicitly and, ideally, provide code or a subject-disjoint fold index in the supplement so that the reader can verify that no trial from a test subject appears in training.","section":"§4.6, Algorithm 1"},{"comment":"There is a direct numerical inconsistency in participant counts. Section 4.2 excludes 13+1 participants from 160, leaving 146, and Table 2 lists 49+47+50=146. Section 4.3 then states that 6 participants with extremely noisy EEG were excluded from the analysis, which would imply a final sample of 140. Please clarify whether the 6 noisy participants are excluded before Table 2, whether Table 2 reports the pre-EEG-quality sample, and whether all reported analyses use 146 or 140 participants; if 140, update group sizes, demographics, and all related text and figures.","section":"§4.2 vs §4.3, Table 2"},{"comment":"The statistical significance reported in Table 1 for the deep-learning AUCs is attributed to permutation tests, but the Methods section describes only bootstrap confidence intervals from resampling predictions. The permutation test procedure for the deep-learning AUCs should be specified (e.g., how the null distribution is generated, whether participants are permuted, and how multiple comparisons across conditions are handled). Without this, the p-value annotations are not verifiable.","section":"§4.7, Table 1"}],"minor_comments":[{"comment":"The deep-learning architecture is described as '2 convolution layers and 2 vanilla transformer layers,' but the text does not give the number of channels, kernel sizes, embedding dimensions, number of heads, or training epochs; these should be reported or the code should be made public.","section":"§4.6"},{"comment":"The statement that data and code are 'available upon request' is not sufficient for a machine-learning classification paper whose central result depends on exact CV and bootstrap details; please provide an open repository or a detailed supplementary protocol.","section":"Data/Code Availability"},{"comment":"The behavioral classifier (AUC 0.94 for CvsD,S and 0.88 for DvsS) is described only in the Discussion, with no method description or table. Since it is used to argue that EEG captures information beyond behavior, please describe its input features, training, and evaluation procedure.","section":"§3, behavioral baseline"},{"comment":"The sentence in 'EEG data corresponding to 6 participants was found to be extremely noisy and was excluded from the analysis' should specify whether these 6 are counted in the Table 2 demographics or removed before all analyses.","section":"§4.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is potentially publishable if the authors can confirm that the cross-validation is subject-disjoint and resolve the participant-count inconsistency. The CV issue is the main risk: Algorithm 1 implies subject-level splitting, but the main text should state it explicitly, and reproducibility would be much improved by releasing code or fold indices. I do not see a basis for rejection, but the current text does not allow a reader to fully verify the central claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honestly, the alert about the cross-validation split is a false alarm in the main. Algorithm 1 shows the bootstrapped trials are generated per training subject and per test subject, and the AUC is computed over test participants, so the folds are clearly subject-level even if Section 4.6 doesn't spell it out. That said, the authors should say the words \"participant-level folds\" in the text; right now you have to reverse-engineer it from the pseudocode.\n\nWhat's genuinely new: using a self-referential word-by-word sentence task as the stimulus for EEG-based classification of depression, and splitting the depressed group by suicidal ideation. The MVPA temporal decoding results (late cluster 544–900 ms for all sentences, earlier frontal effects for contrasts) are a solid characterization of when and where the group differences live. I also credit the paper for running the right controls: a behavioral-confound check (EEG does not predict agree/disagree), all-sentence and random-contrast baselines, permutation tests, confidence intervals, and ablation studies for trial count and bootstrap parameters. That is more careful than most papers at this level.\n\nSoft spots, in proportion. The 0.707 headline AUC is selected from many sentence-grouping conditions; no correction for multiple comparisons across conditions. The 95% CIs are per condition, not across all comparisons, so the true confidence is a bit wider. Labels come from questionnaire cutoffs (PHQ-9 ≥10, SIS>16), not clinical interviews; that's a reasonable screening definition, but it limits claims about MDD diagnosis. Data and code are only on request, so external verification will be slow. The D vs S separation at 0.624 is modest and should be treated as exploratory, as the authors mostly do.\n\nOne thing I disagree with the reader's report on: the CV split is adequately specified by Algorithm 1; the text could be clearer, but the intent is there. The stress-test note that the AUCs might be trial-level is not supported by the actual method.\n\nWho's this for? Researchers working on EEG biomarkers or affective neuroscience who want an example of a task-based paradigm with a thorough evaluation pipeline. It's not a ready clinical tool, and the authors don't oversell it. I'd send it to referees. A good referee will push for clearer fold disclosure, a multiple-comparison correction, and ideally open data, but none of that is a deal-breaker for peer review.","headline":"A credible task-based EEG depression classification study with strong controls; the headline CV-split worry is mostly answered by Algorithm 1, but the paper still needs clearer folding disclosure and is best read as a careful proof-of-concept, not a clinical tool.","tokens_in":18147,"tokens_out":2345,"would_cite":true,"duration_ms":24041,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that EEG responses to self-referential affective sentences carry reproducible signatures of depression, with a deep-learning classifier reaching AUC 0.707 for healthy versus depressed participants and 0.624 for depressed…","keywords":["Major depressive disorder","suicidal ideation","electroencephalography (EEG)","self-referential processing","affective sentence processing","multivariate pattern analysis","deep learning classification","event-related potentials"],"falsifier":"Run the classifier with an explicit participant-disjoint five-fold cross-validation, keeping all trials from each participant in exactly one fold, and compare the resulting AUC with the reported values; if the healthy-versus-depressed AUC falls to near 0.5, the central claim fails because the original numbers were inflated by within-subject leakage. A complementary check is a label-permutation null distribution over subjects to confirm that AUC 0.707 sits outside the permutation range.","tokens_in":17146,"feed_emoji":"🧠","tokens_out":8101,"duration_ms":77113,"temperature":0.7,"pith_summary":"This paper asks whether the brain's response to self-referential affective sentences can serve as an objective neural marker of depression, and reports group-level evidence that it can. Using 64-channel surface EEG recorded while 146 young adults read sentences one word at a time and then pressed agree or disagree, the authors find that healthy and depressed participants separate in multivariate decoding starting around 544 ms after the final word, and earlier when positive and negative sentence responses are contrasted. A deep-learning classifier fed bootstrapped trial averages reaches an area under the ROC curve (AUC) of 0.707 for healthy versus depressed participants and 0.624 for depressed participants with versus without suicidal ideation. The strongest separability comes from subtracting negative from positive sentiment trials, and spatial ablations point to anterior electrodes. If the participant-level split assumption is clean, the paradigm offers a scalable, low-cost complement to self-report screening.","feed_headline":"Brain waves when reading a sentence flag depression at AUC 0.707","feed_subtitle":"A model reading 1.1 seconds of brain signals around a sentence's final word separates healthy from depressed participants.","key_machinery":"The load-bearing machinery is the conjunction of a one-word-at-a-time self-referential sentence paradigm with subject-level bootstrapped EEG epochs and a contrast-based classifier. Each trial is epoched from 200 ms before to 900 ms after the final critical word, baseline corrected by z-scoring on the pre-stimulus window, and then B=20 single trials are sampled with replacement and averaged N=200 times per subject to make stable pseudo-trials; these feed a two-convolution-layer, two-transformer-layer network with sinusoidal positional encoding. For the time-resolved decoding, per-subject ERPs for each sentence grouping are passed at each 10 ms time step to an L1-regularized logistic regression in a leave-one-subject-out scheme, and cluster-based permutation testing identifies windows where AUC exceeds chance. The crucial design choice is contrast subtraction: positive minus negative sentence sentiment, agree minus disagree, and slow minus fast response time produce the inputs with the highest separability, and reveal much of the signal's spatial focus in anterior electrodes.","core_discovery":"The central claim is that the temporal dynamics of EEG responses to affective self-referential sentences contain reproducible signatures of depression, not just task behavior. In a between-subject multivariate pattern analysis, healthy controls and depressed participants (with or without suicidal ideation) are decodable from the averaged response to all sentences in a late window from roughly 544 ms to 900 ms after the onset of the final critical word, which the authors associate with semantic integration and decision making; contrasting positive minus negative sentence ERPs shifts decodability earlier, to 256-657 ms, with importance concentrated in anterior midline electrodes tied to emotional processing. The deep-learning results extend this from time-resolved decoding to full spatiotemporal classification: a transformer-based model trained on bootstrapped subject-level trials achieves AUC 0.707 (95% CI 0.66-0.75) for controls versus the combined depressed groups under the sentence-sentiment contrast, and AUC 0.624 for depressed non-suicidal versus depressed suicidal under the response-time contrast. The paper further claims that a model trained only on controls versus non-suicidal depressed assigns a mean depressed probability of 0.61 to the held-out suicidal group, suggesting that the neural signature generalizes across depression subgroups.","pith_inferences":["A likely next step left implicit by the paper is to combine the EEG contrast signal with the behavioral agree or disagree profile, which the paper reports reaches near-ceiling AUC; the neural model adds value precisely where self-report is unreliable.","The paper's opposite sentiment effects for healthy-versus-depressed versus suicidal-versus-non-suicidal classification suggest a testable hypothesis: sentence sentiment may separate symptom dimensions rather than diagnostic labels.","If the participant-level split assumption is confirmed, the natural replication is a shortened anterior-channel version of the task on an independent cohort, using only the early contrast window from 256 ms to 657 ms."],"forward_implications":["A shortened screening protocol remains viable: the response-type model keeps performance with up to 75% of a new participant's trials removed at test time.","Training data can be small: sentence-sentiment models keep stable AUC when up to 75% of depressed participants are removed from training, easing collection in clinical settings.","EEG predictions track symptom severity: model depression probabilities correlate more strongly with same-day PHQ-9 scores than with screening-time scores.","Depression-subtype labels may not be required: models trained on one depressed subgroup transfer to the other, with held-out suicidal participants still receiving elevated depressed probabilities.","Anterior electrodes contain much of the discriminative information, suggesting future studies could use a reduced electrode montage."],"supporting_citations":[{"why":"Defines the PHQ-9 threshold that separates controls from depressed participants.","marker":"[3]"},{"why":"Supplies prior evidence of reduced neural response to positive emotional material in depression, which the sentence-sentiment contrast extends.","marker":"[31]"},{"why":"Provides prior ERP evidence that depressed participants process self-referential negative adjectives differently, motivating the final-word manipulation.","marker":"[34]"},{"why":"Shows that self-referential sentence-response tasks produce ERP differences that this paper converts into a classification task.","marker":"[35]"},{"why":"Supplies the multivariate pattern analysis framework used for time-resolved EEG decoding.","marker":"[39]"},{"why":"Provides the SIS threshold used to divide depressed participants into non-suicidal and suicidal groups.","marker":"[46]"},{"why":"Describes the artifact subspace reconstruction method used to clean the EEG before classification.","marker":"[54]"},{"why":"Supplies the transformer architecture used in the deep-learning classifier.","marker":"[59]"}],"fun_headline_variants":["EEG sentence reading flags depression at AUC 0.707","Depression decoded from EEG to affective sentences (AUC 0.707)","Affective sentence EEG yields depression signature (AUC 0.707)","Neural responses to affective sentences decode depression (AUC 0.707)"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the five-fold cross-validation keeps all trials from each participant in a single fold; the paper does not state this explicitly, and if trials from the same person appear in both training and test sets, within-subject correlation could inflate the reported AUC.","fun_headline_variants_meta":{"raw":{"variants":["EEG sentence reading flags depression at AUC 0.707","Depression decoded from EEG to affective sentences (AUC 0.707)","Affective sentence EEG yields depression signature (AUC 0.707)","Neural responses to affective sentences decode depression (AUC 0.707)"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001343,"raw_usage":{"total_tokens":5463,"prompt_tokens":954,"completion_tokens":4509,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":4430}},"tokens_in":570,"tokens_out":4509,"duration_ms":30061,"temperature":1.0,"reasoning_tokens":4430,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:57:57.709823+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the classifier with an explicit participant-disjoint five-fold cross-validation, keeping all trials from each participant in exactly one fold, and compare the resulting AUC with the reported values; if the healthy-versus-depressed AUC falls to near 0.5, the central claim fails because the original numbers were inflated by within-subject leakage. A complementary check is a label-permutation null distribution over subjects to confirm that AUC 0.707 sits outside the permutation range.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the PHQ-9 threshold that separates controls from depressed participants."},{"cited_title":"Y., Deldin, P","cited_arxiv_id":null,"evidence_quote":"Supplies prior evidence of reduced neural response to positive emotional material in depression, which the sentence-sentiment contrast extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides prior ERP evidence that depressed participants process self-referential negative adjectives differently, motivating the final-word manipulation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that self-referential sentence-response tasks produce ERP differences that this paper converts into a classification task."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the multivariate pattern analysis framework used for time-resolved EEG decoding."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the SIS threshold used to divide depressed participants into non-suicidal and suicidal groups."},{"cited_title":"R.et al.Real-time neuroimaging and cognitive monitoring using wearable dry eeg.IEEE transactions on biomedical engineering62, 2553–2567 (2015)","cited_arxiv_id":null,"evidence_quote":"Describes the artifact subspace reconstruction method used to clean the EEG before classification."}],"review_version":1}