{"id":"27c2020e-9364-4fed-ae8e-1929829ff789","arxiv_id":"2508.08273","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"Keyword distillation of clinical notes improved BERT predictions and explanation quality in small tests, but effects are modest and some important details are unverified.","lead":"TT-XAI tests whether turning long hospital discharge notes into short keyword lists helps a medical AI predict long hospital stays and explain itself. Across small tests the keyword approach scored better on accuracy and expert-rated clarity, but the study is small and leaves key data-processing details unverified.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Raw-text baseline is truncated (≤1024 tokens) while keyword inputs are extracted from full notes, so the classification gain may reflect information access rather than domain-aware distillation.","rationale":"The Reader's verdict of CONDITIONAL is appropriate, but the most load-bearing weakness is not the date-leak assumption; it is the unequal information access in the classification comparison. The paper's own description (Section 3.1) says discharge notes often exceed 6,000 tokens, yet Table 1 evaluates raw inputs only at up to 1,024 tokens, while the keyword inputs are extracted from the complete notes before truncation. This guarantees that the keyword model has effectively seen the whole document, whereas the raw model sees only a prefix. The reported improvement could therefore come from document-level access rather than from the domain-awareness of Rakun/Med7. This is a verifiable methodological flaw, not a speculative one. The explanation-fidelity comparison has a similar issue if the two LIME variants are run on different underlying models, which the paper does not clarify. The date-leak concern raised by the Reader is still worth checking, but the low raw-text accuracy (0.665 F1 at 512 tokens) suggests those direct LOS indicators were likely removed by preprocessing; otherwise a model could trivially compute the label. Thus I keep the same CONDITIONAL verdict, but the condition should require a full-length or matched-access raw-text baseline, plus a same-model explanation comparison, before the central distillation claim can be accepted.","tokens_in":10853,"tokens_out":5982,"duration_ms":69405,"concrete_test":"Retrain ModernBERT on complete raw discharge notes (e.g., 8192-token context or chunked with aggregation) and compare macro-F1 against the 512-token keyword model; also run a control using a random 512-token segment sampled from each full note and an extractive sentence-selection baseline. If keyword F1 does not exceed the random-segment and full-text baselines by a meaningful margin (e.g., >5 points), the distillation benefit is an artifact of truncation. Additionally, apply classical LIME and focused LIME to the same trained model to isolate the effect of the perturbation set from the effect of the underlying classifier.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 4.1 the 'raw full-text' baseline is not full text: Table 1 truncates raw discharge notes to 50–1024 tokens, while the keyword representation is built by Rakun from the entire note (which the authors state can exceed 6,000 tokens, Section 3.1) and then truncated. The keyword model therefore has document-level access through the extraction step; the raw model only sees a prefix. Clinically relevant sections such as the hospital course and discharge instructions often appear late in MIMIC-IV notes, so the reported F1 gain (0.665→0.767 at 512 tokens) may reflect information access, not the value of domain-aware keyword distillation. A fair control would feed the raw model a random 512-token segment from the full note or use a long-context model on the complete note. The explanation-fidelity comparison in Section 4.2 is additionally confounded if classical LIME is applied to the raw-text model and focused LIME to the keyword model, since deletion curves then compare different classifiers. The Reader's date-leak concern is plausible, but the low raw-text accuracy in Table 1 suggests those fields were probably removed; the truncation confound is guaranteed by the described pipeline and is the more load-bearing threat.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TT-XAI, a framework that distills long MIMIC-IV discharge notes for kidney-stone patients into keyword representations using the RaKUn algorithm and Med7 named-entity recognition, then uses these distilled inputs for two purposes: (1) fine-tuning a ModernBERT classifier to predict prolonged length of stay, and (2) guiding focused LIME explanations and LLM-generated chain-of-thought reasoning. The authors report that keyword inputs outperform raw truncated notes in macro-F1 across context lengths (e.g., 0.767 vs. 0.665 at 512 tokens), that focused LIME achieves lower deletion-curve AUC (0.668 vs. 0.742), and that both LLM-based scoring and a blinded three-specialist study favor keyword-augmented explanations (average 3.15 vs. 2.58). The paper concludes that lightweight keyword distillation simultaneously improves predictive performance, token-level explanation fidelity, and human-perceived explanation quality.","tokens_in":11057,"tokens_out":3605,"duration_ms":42904,"significance":"If the central claim holds, the paper offers a lightweight, domain-aware preprocessing step that could improve both accuracy and interpretability for clinical text models, which would be practically valuable given the length and noise of EHR notes. The manuscript has real strengths: the code is promised to be publicly available, the evaluation includes a blinded human specialist study in addition to automatic metrics, and the pipeline is simple enough to be reproducible. However, the significance is substantially tempered by evaluation-design issues. The classification comparison does not control for information access, because the raw-text baseline is truncated from the beginning of the note while the keyword representation is extracted from the full note; the explanation-fidelity comparison compares different classifiers under different perturbation spaces; and the LLM-based evaluation uses a judge from the same model family as the generator. These issues are load-bearing for the paper's headline claims and require additional experiments or explicit controls before the conclusions can be accepted.","major_comments":[{"comment":"The central classification claim is confounded by unequal information access. The 'Original' baseline is truncated to the first 50-1024 tokens of the raw discharge note, while the 'Keywords' representation is built by RaKUn from the entire note, which the authors state can exceed 6,000 tokens (Section 3.1), and is then truncated. Consequently, the keyword model can draw on document-level evidence, whereas the raw model only sees a prefix. Clinically relevant content such as the hospital course or discharge instructions may appear late in MIMIC-IV notes, so the reported F1 gains may reflect access to more of the document rather than the value of domain-aware distillation per se. A fair control should feed the raw model either a random 512-token segment sampled from the full note or the complete note through a long-context model; without such a control, the claim that 'distilled keyword inputs consistently outperform raw text' (Section 4.1) is not established.","section":"Section 4.1, Table 1"},{"comment":"The paper does not explicitly state that admission and discharge dates were removed from the discharge summaries during preprocessing. MIMIC-IV discharge notes typically contain 'Admission Date' and 'Discharge Date' fields, from which the binary LOS_long label can be computed directly. Section 3.1 mentions only 'basic filtering to remove administrative boilerplate and empty sections' and never rules out this direct label leak. The low raw-text accuracy in Table 1 (e.g., 0.638 at 512 tokens) suggests that dates were probably not present, but this must be stated explicitly and verified in the preprocessing description; otherwise the classification improvements in Section 4.1 and the downstream explanation findings inherit the uncertainty.","section":"Section 3.1"},{"comment":"The explanation-fidelity comparison is not apples-to-apples. Classical LIME is applied to the raw-text model, while focused LIME is applied to the keyword model, so the deletion curves in Figure 3 compare different classifiers operating on different input spaces. A lower AUC for focused LIME could reflect the fact that the keyword model's decision is more concentrated on a small salient set, rather than that focused LIME produces more faithful explanations. In addition, the focus set F is built from the same RaKun keywords whose benefit is being demonstrated, making the deletion test partially self-referential; and the evaluation is restricted to 20 preselected correctly classified long-stay examples, which limits generality. A cleaner test would compare classical and focused LIME on the same model (e.g., both on the keyword model, or both on a model that takes the full note with the same perturbation budget), and would report how the 20 examples were selected.","section":"Section 4.2, Section 3.4"},{"comment":"The LLM-based evaluation is self-referential and the reporting is inconsistent. Section 3.5 states that 'LLM LLaMA 3 70B then scored each explanation,' yet Section 4.3 and Table 2 describe these as 'expert scores.' More importantly, the scoring model belongs to the same model family (LLaMA-3) as the generator (DeepSeek-distilled LLaMA-3 14B), so apparent preferences for keyword-augmented prompts may reflect a within-family bias rather than a general quality difference. The human specialist study (Section 4.4, Table 3) is an independent check, but it uses only 10 explanation pairs and three raters, and one of the three specialists actually scored Method B lower (3.33 vs. 3.45), so the 'consensus' phrasing in Section 4.4 overstates the evidence. A larger, more diverse rater pool and a non-LLM or cross-family judge would be needed to support the claim that keyword augmentation consistently improves explanation clarity and clinical usefulness.","section":"Section 4.3, Section 3.5"}],"minor_comments":[{"comment":"The text says 'Experts rated each' but the experimental protocol in Section 3.5 specifies LLaMA-3 70B scoring; please replace 'expert' with 'LLM judge' or clarify that these are LLM-provided scores.","section":"Section 4.3, Table 2"},{"comment":"The deletion-test description says 'top-k influential tokens' but never specifies the value of k or the step size used to construct the deletion curve; please provide these details for reproducibility.","section":"Section 3.4"},{"comment":"It is unclear whether the 20 correctly classified long-stay examples used for deletion curves are the same 20 true-positive cases used for LLM reasoning; please state the overlap explicitly.","section":"Section 4.2 vs. Section 3.5"},{"comment":"No statistical significance tests are reported for the accuracy/F1 differences across folds; given the small cohort (467 admissions) and the overlapping standard deviations at several context lengths, a paired test or confidence interval would strengthen the 'consistently outperform' claim.","section":"Table 1"},{"comment":"The caption's phrase 'explanations aligned with tokens in the explanations' is circular and should be replaced with a precise description of what the deletion curve measures.","section":"Figure 3 caption"},{"comment":"Specialist 3 gave Method A a higher score than Method B (3.45 vs. 3.33); the text acknowledges this but should soften the phrase 'consensus clearly indicates' in Section 4.4 accordingly.","section":"Section 4.4, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a clinical NLP journal and the proposed pipeline is simple and interesting, but the current evaluation does not yet support the strongest claims. The truncation confound is the most serious issue because it is guaranteed by the described pipeline: the raw baseline cannot see beyond the first N tokens while the keyword extractor sees the whole note. The date-leak concern is secondary and likely false given the low raw-text accuracies, but the authors should state it explicitly. I would encourage the editor to require the additional baselines and the clarification of the LIME comparison before reconsidering the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a usable paper, not a field-changer. The new thing is the packaging: Rakun keyword extraction plus Med7 entities as both a preprocessing step for ModernBERT and a focus set for LIME, plus keyword-augmented chain-of-thought prompts for LLM explanations. Individually those parts are known; the combination applied to MIMIC-IV kidney-stone LOS prediction is new, and the blinded human evaluation with three specialists is a nice touch. The writing is clear and the code is promised.\n\nThe main result — keyword inputs beat raw text by 10 F1 points at 512 tokens — is probably real as stated, but the comparison is not apples-to-apples. Raw baselines are truncated to 50–1024 tokens, while Rakun extracts keywords from the full note (which can exceed 6,000 tokens) and then truncates the keyword string. So the keyword model effectively sees the whole document through the extractor; the raw model sees only the first N tokens. That is a guaranteed confound, not a hypothetical. A random-segment baseline or a long-context model on the full note would settle it. The date-leak worry in the reader report is plausible but less load-bearing: the low raw-text accuracy in Table 1 suggests the dates were probably stripped, but the paper never says so, which is sloppy.\n\nThe explanation-fidelity comparison has a similar issue: focused LIME perturbs only keywords while classical LIME perturbs all tokens, and the deletion test runs on different classifier inputs. Lower AUC may just reflect the perturbation set, not more faithful explanations. The LLM-as-judge is same-family (LLaMA-3 70B scoring LLaMA-3 14B), and the human study is 10 cases with no significance testing. Those are real but fixable gaps, not fatal ones. The paper honestly lists some limitations in Section 5, which earns some trust.\n\nIf I were a referee I would ask for the truncation control, a statement on dates, and significance testing on the human scores. The core idea — cheap keyword distillation to improve both prediction and interpretability on long clinical notes — is worth taking seriously. Small deployment settings would benefit if it holds. It deserves peer review with major revisions; I would not desk-reject it.","headline":"Keyword distillation is a plausible cheap win for clinical NLP, but the headline F1 gain is confounded by raw text being truncated while keywords are extracted from full notes.","tokens_in":11650,"tokens_out":2066,"would_cite":false,"duration_ms":23877,"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":"Keyword distillation of clinical notes simultaneously improves BERT classification and the faithfulness of LIME and LLM explanations.","keywords":["trustworthy AI","clinical NLP","electronic health records","keyword extraction","explainability","LIME","chain-of-thought reasoning","length of stay prediction"],"falsifier":"Inspect the distilled inputs for a random sample of the 467 notes and check whether date-like tokens or explicit stay-length phrases survive; if stripping those tokens makes the keyword advantage over raw text disappear, the classification gain is an artifact of label leakage rather than of distillation.","tokens_in":10588,"feed_emoji":"🩺","tokens_out":4787,"duration_ms":47473,"temperature":0.7,"pith_summary":"The paper claims that converting long, noisy discharge summaries into a compact keyword representation, built by the RaKUn graph-based extractor and Med7 clinical entities, improves both prediction and interpretability at once. On a 467-note MIMIC-IV kidney-stone cohort, distilled inputs raise ModernBERT's macro-F1 from 0.665 to 0.767 at a 512-token budget, with gains at every tested context length. The same focus set sharpens LIME into a 'focused LIME' whose deletion-curve AUC drops from 0.742 to 0.668, and steers LLaMA-3 chain-of-thought reasoning that three blinded specialists rate higher (3.15 vs 2.58 on a 1–5 scale). If correct, the result points to a lightweight preprocessing recipe that makes clinical classifiers both more accurate and more auditable.","feed_headline":"Distilling clinical notes to keywords lifts F1 by 10 points","feed_subtitle":"The same keyword focus set sharpens explanations and wins higher specialist ratings.","key_machinery":"The central mechanism is keyword distillation: the RaKUn algorithm builds a word co-occurrence graph per note, ranks terms by load centrality, and forms bigram and trigram keyphrases, which are then unioned with Med7 clinical named entities to form a focused 'focus set' of up to 512 phrases. This focus set plays a triple role: it is the input to the fine-tuned ModernBERT classifier, the restricted perturbation space for the focused LIME variant, and the guiding prefix in the keyword-augmented chain-of-thought prompts.","core_discovery":"The central claim is that raw discharge notes are information-dilute for both prediction and explanation: most of the signal for a prolonged-stay label is concentrated in a small set of clinically salient terms. By distilling each note into its top keyword phrases (RaKUn) plus medical entities (Med7), the paper reports that ModernBERT's macro-F1 rises from 0.665 to 0.767 at 512 tokens, that a LIME variant perturbing only this focused set yields a lower deletion-curve AUC (0.668 vs 0.742), and that LLM explanations generated from keyword-augmented prompts are rated higher by three blinded specialists (3.15 vs 2.58). The finding is that distillation, a simple unsupervised preprocessing step, improves machine prediction and human-perceived explainability simultaneously, rather than trading one off against the other.","pith_inferences":["The same distillation-and-focus-set recipe should transfer to other MIMIC-IV prediction tasks (readmission, mortality, sepsis) and to any long-document classifier where the label depends on a few explicit facts.","A direct ablation that explicitly strips date tokens and other stay-length phrasing from both raw and distilled inputs would isolate the distillation effect from any label leakage and should be run before clinical deployment.","The deletion-curve fidelity measure used for LIME could be adapted to grade the keyword-augmented LLM explanations themselves, unifying the two evaluation branches into a single metric."],"forward_implications":["At a 512-token budget, keyword-distilled inputs raise macro-F1 to 0.767 compared with 0.665 for raw text, implying that the same model can be run on far fewer tokens with better accuracy.","Focused LIME's lower deletion AUC (0.668 vs 0.742) means explanations rank decision-driving tokens more accurately, which should make post-hoc auditing of clinical classifiers more reliable.","Keyword-guided chain-of-thought prompts produce explanations that blinded specialists rate as clearer and more clinically useful, supporting their use in interactive clinical decision support.","Because the distillation is an unsupervised preprocessing step, it can be bolted onto existing clinical classifiers and explanation pipelines without retraining the underlying language model."],"supporting_citations":[{"why":"Supplies the RaKUn graph-based keyword extraction algorithm that produces the distilled focus set.","marker":"Škrlj et al., 2019"},{"why":"Supplies the Med7 clinical named-entity model that identifies medical entities for the focus set.","marker":"Kormilitzin et al., 2021"},{"why":"Provides the MIMIC-IV discharge summaries and admission metadata used to build the kidney-stone cohort and length-of-stay labels.","marker":"Johnson et al., 2021"},{"why":"Defines ModernBERT, the pretrained encoder that is fine-tuned on raw and distilled text.","marker":"Warner et al., 2024"},{"why":"Defines the LIME method that the paper adapts into its focused perturbation variant.","marker":"Ribeiro et al., 2016"},{"why":"Establishes chain-of-thought prompting, the reasoning strategy that the keyword-augmented prompts are built on.","marker":"Wei et al., 2022"}],"fun_headline_variants":["Keyword distillation sharpens clinical predictions and explanations","Distilling notes to keywords lifts F1 and human-rated trust","TT-XAI: Keyword focus beats full notes for accuracy and clarity","Short keywords, not long text, boost clinical AI explainability","Keywords from notes: AI gets sharper and more trustworthy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The preprocessing genuinely strips direct length-of-stay statements (admission and discharge dates and any explicit stay-length phrasing) from the notes, so the classifier must learn clinical indicators rather than read the answer off the page.","fun_headline_variants_meta":{"raw":{"variants":["Keyword distillation sharpens clinical predictions and explanations","Distilling notes to keywords lifts F1 and human-rated trust","TT-XAI: Keyword focus beats full notes for accuracy and clarity","Short keywords, not long text, boost clinical AI explainability","Keywords from notes: AI gets sharper and more trustworthy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1478,"prompt_tokens":876,"completion_tokens":602,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":521}},"tokens_in":492,"tokens_out":602,"duration_ms":6938,"temperature":1.0,"reasoning_tokens":521,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:17:15.193186+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the distilled inputs for a random sample of the 467 notes and check whether date-like tokens or explicit stay-length phrases survive; if stripping those tokens makes the keyword advantage over raw text disappear, the classification gain is an artifact of label leakage rather than of distillation.","supporting_citations":[],"review_version":1}