{"id":"7a90284f-e05b-477f-b020-38430b0d64c0","arxiv_id":"2505.17643","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A contrastive pretraining step that aligns static EHR tables with discharge notes improves downstream TabNet predictions on MIMIC-IV readmission and critical outcome tasks.","lead":"This paper trains a contrastive learning model that links patient data tables with their discharge summaries, then uses only the table-based model for prediction. It reports better readmission and critical-outcome predictions than standard baselines on MIMIC-IV, but the design may let outcome information leak through the text.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Discharge summaries likely leak outcome mentions into EHR embeddings via the CLIP alignment; the critical outcome gain may be a leakage artifact rather than semantic transfer. A text-probe experiment would settle it.","rationale":"The reader identified label leakage through discharge summaries as the primary weakest assumption. I agree that this is the single most load-bearing concern: it directly threatens the critical outcome result (Table 2), which is the largest claimed improvement, and it undermines the mechanistic interpretation that the gain comes from general semantic transfer. The paper's own description of preprocessing makes the concern concrete: while structured features are scrubbed of target-related columns, the text is not scrubbed of outcome mentions. The contrastive objective guarantees that any predictive signal in the text embeddings is injected into the EHR encoder. The proposed text-probe test is feasible and decisive: if text embeddings alone can predict the labels, the leakage pathway is verified; if not, the concern is laid to rest. I also considered the absence of an EHR-only contrastive control as a possible concern, but that addresses the mechanism rather than the validity of the empirical result; leakage is more damaging. The reader's conditional verdict appropriately flags this risk, and our proposed check would supply the missing evidence, so the verdict should remain unchanged pending the test.","tokens_in":7997,"tokens_out":6880,"duration_ms":73954,"concrete_test":"A single decisive check: train a linear probe (e.g., logistic regression) on the mean-pooled [CLS] text embeddings from the pretrained Longformer to predict the critical outcome and readmission labels on the pretraining split. Report AUROC for both. If the text probe achieves AUROC substantially above chance (e.g., >0.8 for critical outcome), then the text carries direct outcome signal, confirming the leakage pathway. If the probe is near chance, the concern is defused. Optionally, to confirm causality, repeat CL pretraining with outcome-related sentences masked (e.g., remove sentences containing 'expired', 'died', 'death', 'DNR', 'comfort care', 'intubated', 'ventilated', 'code blue') and re-run the downstream fine-tuning; if the gain over the masked-pretraining TabNet drops below 5% AUROC, the published gain is attributable to leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that contrastive pretraining (Section 3.1) transfers clinically meaningful semantic information from discharge summaries into the EHR encoder. For this to hold, the text modality must not directly encode downstream labels. In MIMIC-IV, discharge summaries frequently contain explicit outcome information such as 'the patient expired', 'comfort care', 'DNR', or descriptions of critical events. The preprocessing in Section 4.1 removes numbers, dates, punctuation, and some administrative sections, but it does not scrub outcome-bearing sentences. Because the CLIP loss (Equation LCLIP) aligns EHR and text embeddings, any outcome signal present in the text embeddings is pulled into the EHR embeddings. At downstream fine-tuning, the EHR encoder then has a shortcut to the critical outcome label, so the 13.25% AUROC gain over XGBoost in Table 2 could reflect label leakage rather than general semantic transfer. The readmission gain (4.1%) is also vulnerable if summaries mention unplanned returns or discharge disposition. The paper correctly notes that text is not used at inference, but this does not address leakage during pretraining. This unverified premise—that the notes carry no direct label information—is the most load-bearing assumption for the empirical claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a multimodal contrastive learning framework that aligns static EHR features with discharge summary text via a CLIP-style loss, then fine-tunes the EHR encoder on two downstream tasks: 30-day hospital readmission and critical outcome prediction. The authors report that the CL-pretrained TabNet encoder achieves mean test AUCs of 0.809 and 0.821, outperforming XGBoost and a masked-pretraining-only TabNet baseline, and that the gains persist when fine-tuning on 50% of the training data. The central claim is that contrastive pretraining transfers clinically meaningful semantic information from clinical notes into the EHR encoder, enabling text-free inference. The paper includes code and pretrained weights in the supplementary material and reports statistical significance across five random seeds.","tokens_in":8218,"tokens_out":3037,"duration_ms":27119,"significance":"If the central claim holds, the framework would be a practical way to inject clinical knowledge from notes into tabular-EHR pipelines without requiring text at inference, which is attractive for deployment settings where text is unavailable or expensive. The paper's strengths include a clear and reproducible architecture, paired-data construction from MIMIC-IV, multiple-seed evaluation with significance tests, and public release of code and weights. However, the empirical claims rest on a load-bearing assumption that the discharge summaries carry no direct information about the downstream labels. Because MIMIC-IV discharge summaries frequently contain explicit outcome statements, such as 'the patient expired' or 'comfort care', the reported gains, especially the 13.25% AUROC improvement for critical outcome prediction, may partly or wholly reflect label leakage through the contrastive objective rather than general semantic transfer. The second load-bearing assumption, that the with-notes pretraining population is representative of the without-notes fine-tuning population, is acknowledged as a limitation but not quantified or corrected.","major_comments":[{"comment":"The text preprocessing in §4.1 removes dates, numbers, punctuation, and some administrative sections, but it does not remove outcome-bearing sentences from discharge summaries. MIMIC-IV discharge summaries often contain explicit statements about death, code status, and critical events (e.g., 'the patient expired', 'DNR', 'comfort care'). Because the CLIP loss in §3.1 aligns EHR embeddings with text embeddings, any such outcome signal in the text is pulled into the EHR encoder. At fine-tuning time, the EHR encoder can then exploit this shortcut, making the critical-outcome gain in Table 2 (13.25% over XGBoost) a potential measure of label leakage rather than semantic transfer. The paper's statement that text is not used at inference does not address leakage during pretraining. I recommend a direct probe experiment: train a linear classifier on the frozen EHR embeddings (or frozen text embeddings) to predict each downstream label and compare against chance; additionally, retrain the CL framework after removing outcome-related sections or sentences (e.g., 'Discharge Disposition', 'Hospital Course' sentences mentioning death/code status) and report whether the downstream gains persist. This experiment is essential because the entire empirical claim rests on the text modality not encoding the labels.","section":null},{"comment":"The contrastive pretraining uses 376,021 admissions with discharge summaries, while downstream fine-tuning and testing use a held-out portion of admissions without discharge summaries. The Discussion acknowledges that patients without notes may differ systematically in clinically relevant ways, but the paper does not quantify this selection bias or test its impact. This matters because the claim is that the CL-pretrained encoder 'transfers' knowledge to the note-free population; if the two populations differ in outcome prevalence or comorbidity, part of the observed AUC gain could be due to distribution shift rather than semantic transfer. I suggest reporting patient characteristics for both populations (age, sex, comorbidity counts, outcome rates), and either matching the fine-tuning population to the pretraining population on observable covariates or reporting results on both a with-notes held-out set and the without-notes set. This would clarify whether the fine-tuning improvements generalize beyond the specific population used for evaluation.","section":null},{"comment":"The paper reports p-values from two-tailed t-tests across five seeds, which is appropriate, but the test is underpowered with only five seeds and the standard deviations overlap across baselines in several cases. For example, in Table 1 the 100%-data XGBoost AUC is 0.777 ± 0.012 and the CL-TabNet AUC is 0.809 ± 0.016; the t-test is significant only because the paired seeds are correlated. I recommend reporting the full per-seed results in the main text (or at least the paired differences) and also computing a non-parametric test (e.g., Wilcoxon signed-rank) to confirm the significance claim. This is a minor statistical robustness request, but it directly concerns the confidence one can place in the headline improvements.","section":null},{"comment":"The paper evaluates on a single dataset (MIMIC-IV) and two binary classification tasks. The authors acknowledge this in the Discussion and list external validation as future work, which is good, but the manuscript's title and framing ('Enhanced Clinical Tasks') suggest broader applicability. For a clinical prediction claim, at least one external or temporal validation would substantially strengthen the paper. I would not make this a blocker if the leakage and population-matching issues are resolved, but it is a limitation that should be stated more prominently in the abstract or conclusion, not only in the future-work paragraph.","section":null}],"minor_comments":[{"comment":"There is a typo in the abstract: 'e.g,' should be 'e.g.,'. Also, in §3.1 the phrase 'As shown by Chen et al. (2020)' should be 'as shown by Chen et al. (2020)' (lowercase 'as') for sentence flow.","section":null},{"comment":"The reference for NODE is listed as 'Popov et al. (1909)' but the correct year is 2019; the arXiv number confirms this. Please correct the citation year.","section":null},{"comment":"The phrase 'and and broadens the applicability' contains a duplicated 'and'. Also, 'Tabtransformer' in §2.1 should be 'TabTransformer' with a capital T for the architecture name as used in the cited paper.","section":null},{"comment":"The evaluation setting section does not state the class imbalance or the prevalence of the positive class for either downstream task. Reporting prevalence is standard for clinical prediction papers and helps readers interpret AUC values. Please add this information.","section":null},{"comment":"The figures are informative but the text labels are small and some are cut off in the PDF rendering. Please ensure the figures are legible and that all text in the figures is readable at print size.","section":null}],"recommendation":"major_revision","confidential_remarks":"The leakage concern is the single most important issue and is directly tied to the central claim. The authors should be encouraged to run the proposed text-probe and outcome-scrubbing experiments; if those experiments show that the gains disappear, the paper would need substantial reframing. The selection-bias issue is secondary but also needs quantitative treatment. The paper is otherwise well-structured and the code/weights release is a plus. I would not reject outright because the architecture and evaluation protocol are reasonable and the issues are empirically testable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the report. I read the paper carefully and mostly agree with your conditional verdict, with one adjustment: I think the leakage risk is even more central than your report suggests, and I'd want a text-probe experiment before believing the critical outcome numbers.\n\nWhat's actually new: the paper is the first I've seen that applies a standard CLIP-style contrastive loss to static (non-temporal) EHR feature vectors and discharge summaries, then fine-tunes only the EHR encoder for downstream tasks. That's a useful extension, especially for settings where longitudinal sequences aren't available. The work is cleanly presented, the baselines (XGBoost and masked-pretrained TabNet) are appropriate, and the low-resource result—CL-pretrained TabNet at 50% data beating XGBoost at 100%—is genuinely interesting. The authors also honestly acknowledge selection bias and the two-task scope in their limitations.\n\nThe soft spots are real. The critical outcome result (AUC 0.821 vs 0.725) is large enough to be suspicious. Discharge summaries in MIMIC-IV routinely state outcomes: 'the patient expired,' 'comfort care,' 'DNR,' 'discharged to hospice.' The preprocessing strips numbers, dates and punctuation but not outcome-bearing sentences. The CLIP loss will happily align that signal with the EHR embedding, so the downstream EHR encoder has a shortcut to the label. The paper excludes target-related structured columns but never addresses outcome text. That's a load-bearing gap. A simple experiment—train a linear probe on the text embeddings to predict the outcome—would settle whether the info is in the notes. Similarly, the readmission gain may be inflated if summaries mention planned returns or discharge disposition, though it's smaller and more plausible as genuine transfer.\n\nSecond issue: the fine-tuning set is deliberately patients without discharge summaries, while pretraining used patients with them. The authors call this a leakage safeguard, but it's actually a distribution shift. Note-availability may correlate with severity and outcomes, and the pretrained encoder is being tested on a population it never saw. This could over- or under-estimate the transfer; it's not a clean evaluation.\n\nThird, no code, data, or weights are actually available despite the claim in Section 4.2. The supplementary material is not accessible in the arXiv submission. For a reproducibility-relevant claim, that's a problem.\n\nOverall: this is a competent, honest paper that makes a narrow but legitimate contribution. The readmission result is credible. The critical outcome result is not, without a leakage check. It deserves a serious referee, but I'd recommend rejection unless the authors add a text-probe analysis, address the population shift, and release artifacts.","headline":"A clean CLIP-style extension for static EHR + discharge summaries, but the critical-outcome gain likely leaks through outcome-bearing text; needs a text-probe before it's believable.","tokens_in":8751,"tokens_out":2908,"would_cite":false,"duration_ms":25884,"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":"By aligning static EHR features with discharge summaries through a CLIP-style contrastive loss, the paper shows that the pretrained EHR encoder alone beats XGBoost on 30-day readmission (0.809 vs 0.777 AUROC) and critical outcome…","keywords":["contrastive learning","electronic health records","discharge summaries","multimodal representation learning","CLIP loss","TabNet","readmission prediction","clinical decision support"],"falsifier":"Train the same contrastive framework on discharge summaries from which all sentences mentioning death, critical events, or readmission are removed; if the downstream AUROC gains over masked-pretrained TabNet shrink or vanish, the effect is label leakage, not semantic transfer.","tokens_in":7780,"feed_emoji":"🏥","tokens_out":7478,"duration_ms":65453,"temperature":0.7,"pith_summary":"The paper tries to show that pairing plain structured electronic health records (EHR) with discharge summaries during a contrastive pretraining step makes the EHR representation good enough to predict outcomes that tabular-only models handle poorly. The pretraining pulls the embedding of a patient's EHR close to the embedding of that same patient's discharge summary, and pushes different patients' pairs apart, using a CLIP-style loss. After fine-tuning only the EHR encoder on labeled data, the paper reports AUROC of 0.809 for 30-day readmission and 0.821 for critical outcomes, outperforming XGBoost (0.777 and 0.725) and a masked-pretrained TabNet baseline. The practical point is that clinical notes never have to be available at prediction time: they contribute their knowledge during pretraining, then the EHR-only model carries it forward.","feed_headline":"Notes-aligned EHR model lifts readmission AUC to 0.809","feed_subtitle":"Critical-outcome AUC reaches 0.821, and no clinical text is needed at inference time.","key_machinery":"The load-bearing object is the contrastive alignment itself: a CLIP-style loss over paired EHR and text embeddings, computed as the sum of EHR-to-text and text-to-EHR InfoNCE terms with a temperature parameter. The EHR side uses a TabNet encoder, first pretrained with masked feature reconstruction and then with its low-level embedding and feature-splitting layers frozen; the text side uses a Longformer encoder initialized with clinical-domain weights, processing discharge notes in 256-token chunks and mean-pooling the [CLS] embeddings. Two small MLP projection heads map both modalities into the shared 128-dimensional space where cosine similarity is compared. What this mechanism does is let the structured representation inherit information from the unstructured note during pretraining, while remaining purely tabular at inference.","core_discovery":"The central discovery is that cross-modal alignment of static EHR features with free-text discharge summaries transfers clinically meaningful signal into the EHR encoder. Concretely, a TabNet encoder pretrained with a contrastive objective—matching the embedding of each admission's structured record to its discharge summary while repelling unpaired records—then fine-tuned on two binary tasks reaches mean test AUROC 0.809 for 30-day readmission and 0.821 for critical outcome prediction. These numbers beat the same TabNet architecture initialized only with masked pretraining (0.759 and 0.745) and beat XGBoost (0.777 and 0.725). The gains persist when training data is cut in half, and the text modality is used neither during fine-tuning nor at inference; the paper attributes the improvement to semantic knowledge absorbed from discharge summaries.","pith_inferences":["A decisive check the paper leaves implicit is whether shuffling or paraphrasing the notes while keeping outcome-related content would preserve the gain; if it would, the mechanism is coarse label leakage rather than fine-grained semantic alignment.","The selection-bias caveat the paper acknowledges implies that the pretraining population (patients with discharge summaries) and the fine-tuning population (patients without notes) may differ, so applying the framework where note availability correlates with severity could overstate or understate the transfer benefit.","The same architecture could be extended to predict multiple outcomes at once, or to use the aligned text encoder for report generation from EHR, which the paper names as future work but leaves untested.","One could test whether the alignment transfers across hospitals and domains; the paper uses a single dataset, so cross-institution generalizability remains open."],"forward_implications":["An EHR-only model can absorb knowledge from clinical notes during pretraining and then make predictions without needing notes at inference, making deployment cheaper in settings where notes are unavailable downstream.","The 50%-data experiments imply the pretrained initialization preserves most of its advantage under data scarcity: with 3,000 training points it still beats baselines trained on the full set.","Static-form EHR is sufficient for cross-modal contrastive pretraining, so hospitals without longitudinal or time-series records can still use the approach.","Because the framework is modality-agnostic, the same alignment recipe can in principle be applied to other paired structured and unstructured clinical data, such as imaging metadata and radiology reports.","The method does not require task-specific pretraining labels; the alignment is self-supervised using naturally paired admission records and discharge summaries."],"supporting_citations":[{"why":"Supplies the TabNet encoder, its masked self-supervised pretraining, and the fine-tuning architecture that the whole method builds on.","marker":"Arik and Pfister (2021)"},{"why":"Defines the CLIP loss (dual temperature-scaled InfoNCE) used to align the EHR and text embeddings.","marker":"Radford et al. (2021)"},{"why":"Supplies Longformer, the long-document transformer that encodes the discharge summaries in 256-token chunks.","marker":"Beltagy et al. (2020)"},{"why":"Provides the clinical-domain Longformer weights that initialize the text encoder, giving it clinical language knowledge from the start.","marker":"Li et al. (2022)"},{"why":"Provides the paired EHR and discharge-summary dataset used for pretraining and evaluation.","marker":"Johnson et al. (2023)"},{"why":"Motivates the contrastive objective and projection-head design that the multimodal framework follows.","marker":"Chen et al. (2020)"},{"why":"Earlier EHR-text contrastive work that treats EHR as temporal sequences; the paper positions its static-EHR contribution against this line.","marker":"Ma et al. (2024b)"},{"why":"Another time-series EHR-text contrastive framework used as a contrast point for the static-EHR approach.","marker":"King et al. (2023)"}],"fun_headline_variants":["Contrastive EHR-text alignment lifts readmission AUC to 0.809","Aligned clinical notes boost EHR readmission prediction to 0.809 AUC","Multimodal contrastive learning improves EHR readmission prediction","Note-aligned TabNet beats XGBoost on readmission AUC","EHR encoder aligned with clinical texts lifts readmission AUC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The discharge summaries used during contrastive pretraining must not themselves reveal whether the patient was readmitted, died, or had a critical event; otherwise the reported improvement would measure information leakage rather than understanding transferred from text.","fun_headline_variants_meta":{"raw":{"variants":["Contrastive EHR-text alignment lifts readmission AUC to 0.809","Aligned clinical notes boost EHR readmission prediction to 0.809 AUC","Multimodal contrastive learning improves EHR readmission prediction","Note-aligned TabNet beats XGBoost on readmission AUC","EHR encoder aligned with clinical texts lifts readmission AUC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000354,"raw_usage":{"total_tokens":1893,"prompt_tokens":879,"completion_tokens":1014,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":922}},"tokens_in":495,"tokens_out":1014,"duration_ms":5827,"temperature":1.0,"reasoning_tokens":922,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:42:54.345894+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same contrastive framework on discharge summaries from which all sentences mentioning death, critical events, or readmission are removed; if the downstream AUROC gains over masked-pretrained TabNet shrink or vanish, the effect is label leakage, not semantic transfer.","supporting_citations":[{"cited_title":"Tabnet: Attentive interpretable tabular learning","cited_arxiv_id":null,"evidence_quote":"Supplies the TabNet encoder, its masked self-supervised pretraining, and the fine-tuning architecture that the whole method builds on."},{"cited_title":"Multimodal pretraining of medical time series and notes","cited_arxiv_id":null,"evidence_quote":"Another time-series EHR-text contrastive framework used as a contrast point for the static-EHR approach."}],"review_version":1}