{"id":"8b955510-71e0-4132-94dd-bb9f8a336c87","arxiv_id":"2411.12707","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Barttender encodes tabular EHR features as grayscale bars appended to medical images, letting deep learning models compare imaging versus non-imaging data with comparable accuracy and added explainability.","lead":"Researchers turned electronic health record numbers into gray bars attached to chest X-rays, then trained neural networks on images with and without the X-ray to compare which data type predicts disease better. The method matches standard tabular models in accuracy and adds a way to see which features the model relies on.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Preprocessing pipeline likely fits on test data, undermining the out-of-sample comparison.","rationale":"The reader identified the same load-bearing concern: the absence of a clear statement that quantile normalization and feature selection are fit on training data alone. This is indeed the most critical assumption for the paper's central claim. If test information leaks into preprocessing, the reported parity between Barttender and XGBoost/logistic regression is not a valid out-of-sample result, and the gIoU explanations are also suspect. The concern is fixable (the authors could rerun with train-only preprocessing), so a conditional verdict is appropriate. I concur with the reader's CONDITIONAL verdict, hence UNCHANGED. Other concerns (e.g., the loose use of 'comparable', the arbitrariness of gIoU parameters) are secondary because they do not directly invalidate the core empirical comparison to the same extent as a preprocessing leak. The proposed test reruns the pipeline with proper train/test separation and would settle whether the leakage actually alters the conclusions.","tokens_in":16467,"tokens_out":7431,"duration_ms":70098,"concrete_test":"Re-run the MIMIC experiments (Table 2) with a strict leakage-free protocol: for each of the 10 folds, fit quantile normalization on the training portion only, and re-select the 8 tabular features using logistic regression on the training portion only (or use features fixed by a nested cross-validation on the training folds). Then evaluate on the held-out test set and compare AUC, F1, and MCC to Table 2. If the Image Barttender no longer performs within 0.02-0.03 AUC of XGBoost/Logistic Regression with image biomarkers, or if the Blank Barttender's 0.10 AUC advantage over tabular XGBoost shrinks, the leakage is consequential and the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Barttender performs comparably to scalar-based baselines depends on an unbiased out-of-sample evaluation. The paper does not establish this. In Section 3.2, feature selection uses '≈22,000 samples' to select 8 tabular features via logistic regression Z-scores. The MIMIC evaluation sample (n = 2667) is almost certainly a subset of that larger pool, so test labels and features are used to select the 8 features. Similarly, Section 2.1 describes quantile normalization without stating that the quantiles are fit on training data only; the experimental setup (Appendix B.2) mentions holding out test samples for model training but says nothing about preprocessing. If normalization and feature selection are computed on the full dataset, including the test split, all reported metrics in Table 2 are in-sample estimates and the 'comparable performance' conclusion is not a valid generalization. This is the weakest load-bearing assumption because the entire empirical case for the method rests on this comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Barttender, a framework that converts tabular EHR features into grayscale bars appended to medical images (Image Barttender) or to blank images (Blank Barttender), trains a DenseNet-121 on both, and compares the resulting models with logistic regression and XGBoost baselines trained on scalar representations. It introduces gIoU, an attribution-aggregation metric intended as a global feature importance measure for image-based models, and reports proof-of-concept experiments on CheXpert demographics and MIMIC cardiomegaly prediction. The central claim is that Barttender performs comparably to traditional scalar-based methods while offering local and global explainability.","tokens_in":16652,"tokens_out":5928,"duration_ms":55861,"significance":"If the evaluation is unbiased, Barttender is a useful contribution: it offers an approachable, explainable alternative for comparing imaging versus tabular utility, and it avoids the need for domain-specific image biomarker extraction. The paper is honest about limitations, including the absence of causal claims for gIoU and the need for clinical validation. Reproducibility is aided by the use of public datasets and released code, and the use of two attribution methods strengthens the explainability analysis. The main risk is procedural: the manuscript does not establish that normalization and feature selection are fit on training data alone, and this bears directly on the validity of the headline performance comparisons.","major_comments":[{"comment":"The manuscript does not state whether quantile normalization (Section 2.1) and the logistic-regression feature selection are fit on training folds only. Since the MIMIC evaluation sample (n = 2667) is drawn from the same Duvieusart et al. cohort used to build the ≈22,000-sample feature-selection pool, the text as written does not rule out test information entering preprocessing. The statement in Appendix B.2 that 'test samples were held out across all folds' refers only to model training. If preprocessing was fit on the full cohort, the AUCs in Table 2 are in-sample estimates and the central 'comparable performance' claim is not supported. Please either state explicitly that all normalization and feature selection were nested inside the cross-validation loop, or re-run the pipeline accordingly and report the resulting metrics.","section":"Section 3.2, 'Feature Selection' and Appendix B.2"},{"comment":"The claim that Blank Barttender 'has predictive performance similar to that of a logistic regression' is not fully supported by Table 1: for Cardiomegaly, Blank Barttender AUC is 0.53 ± 0.01 versus 0.57 ± 0.00 for logistic regression, a gap larger than the reported confidence intervals, while F1 and MCC rankings also differ (F1: 0.48 vs 0.38; MCC: 0.10 vs 0.07). The text should either report a statistical comparison of the AUC difference and justify the equivalence threshold, or soften the claim to 'similar on some metrics and some labels.'","section":"Section 4.1 / Table 1"},{"comment":"The gIoU metric is computed from the model's own attribution maps, so it cannot independently validate that the model attends to the bars for the intended reason; a model that encodes a spurious shortcut in the bars would still produce high gIoU for those bars. The statement in Section 1 that 'the model learned robust features from the tabular data' therefore goes beyond what gIoU alone can show. Please add a sanity check, such as permuting or ablating individual bars and verifying that gIoU and model performance change in the expected direction, or compare gIoU rankings against a held-out feature-importance measure.","section":"Section 2.3 / Section 5.1"}],"minor_comments":[{"comment":"There is an unresolved reference 'Section ??' in the sentence 'As mentioned in Section ??'; this should be replaced with a specific pointer to the related work or to Duvieusart et al. (2022).","section":"Section 3.2"},{"comment":"There are several typos, including 'deep learing' in Section 2.1, 'prediciton' in Section 2.2, 'clincal' in Section 5, 'recaptitulated' in the introduction, and 'Urea Nitrogren' in Figure 1; these should be corrected.","section":"Section 2.1 and Figure 1"},{"comment":"The caption of Table 3 says 'computed using gIoU' but the table appears to report raw Integrated Gradients values; please clarify the relationship between the table entries and the gIoU formula.","section":"Table 3"},{"comment":"Please specify the optimizer, learning rate, early stopping criterion, and the exact horizontal/vertical placement and width of the bars (described only as 'about 20%' of the image); these details are needed for reproducibility.","section":"Appendix B.2"},{"comment":"Appendix B.1 contains a self-reference to 'section B.1' in the sentence 'preprocessed as directed by Duvieusart et al. (2022) (details can be found in the appendix, section B.1)'; this should point to the relevant subsection.","section":"Appendix B.1"},{"comment":"All logistic regression entries in Table 2 report ±0.00 for every metric; if these values are rounded to two decimals, please state the rounding rule or report more significant digits so that the confidence intervals are meaningful.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The leakage concern in major comment 1 is the main risk to the paper's central claim. If the authors can show that all preprocessing and feature selection were fit on training folds only, the empirical comparison likely survives, though the Table 1 parity claim should also be reconciled. The paper's fit with the ML4H venue is appropriate, and the work is a reasonable proof-of-concept contribution if these issues are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is simple and well-presented: encode tabular EHR features as grayscale bars appended to the X-ray, train the same CNN on that image and on a blank-image control, and compare utilities. The blank control is a genuinely nice experimental touch—it isolates the tabular signal from the imaging signal. And the paper does a good job of positioning itself against prior tabular-to-image encodings (Zhu, Damri, Medeiros Neto, Alkhodari), with code and public data available. The gIoU aggregation is a reasonable descriptive tool, and the authors are appropriately careful that it supports correlation claims, not causal ones. For a proof-of-concept, the results are plausible: image Barttenders beat blank ones, and blank performance roughly tracks logistic regression and XGBoost.\n\nThe soft spots are real but mostly fixable. The biggest one is the preprocessing leakage that the stress-test flagged. Section 3.2 says feature selection uses roughly 22,000 samples, and the evaluation sample is 2,667. The text does not state that the 2,667 test samples were excluded from that feature-selection pool, and the natural reading is that they were included. Section 2.1 describes quantile normalization without saying the quantiles are fit on training data only, and Appendix B.2 only mentions holding out test samples for model training. So the reported numbers in Table 2 are likely in-sample with respect to feature selection and normalization. This affects all models equally (the same 8 features go into Barttender, XGBoost, and logistic regression), so the comparative ranking might survive, but the absolute out-of-sample performance and the generalizability claim are undermined. That needs to be clarified or corrected.\n\nThe 'comparable' language is also a bit loose. On CheXpert, the Blank Barttender AUC for Cardiomegaly is 0.53 versus logistic regression's 0.57—not obviously 'comparable' on AUC, though the MCC ranking is similar. And the encoding parameters (bar area fraction, attention threshold, blur kernel) are arbitrary and not swept, so sensitivity is unknown. That is a minor concern, not a fatal one.\n\nOverall, this is a solid proof-of-concept for clinical ML researchers who want a quick, interpretable benchmark of imaging versus tabular utility. The contribution is modest but real, and the evaluation is honest in its limitations. I would send this to peer review with a request to fix the train/test separation in preprocessing and to be more precise about what 'comparable' means. If the leakage is corrected and the numbers hold, it would be a useful addition to the multimodal benchmarking toolbox.","headline":"Clean idea—tabular bars plus blank-image control—but the preprocessing may leak test set information, so the headline 'comparable performance' claim needs a fix before I'd trust it.","tokens_in":754,"tokens_out":802,"would_cite":false,"duration_ms":34471,"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":"Grayscale bars let deep learning compare X-rays with EHR data directly.","keywords":["medical image analysis","tabular data","deep learning","explainability","gIoU","CheXpert","MIMIC","multimodal machine learning"],"falsifier":"Re-run the MIMIC workflow with quantile normalization and the logistic-regression feature selection performed separately inside each training fold. If the Blank Barttender no longer matches or beats XGBoost on tabular data, or if the Image Barttender's advantage shrinks, the reported parity would be traceable to leakage rather than to the bar representation.","tokens_in":16245,"feed_emoji":"🩻","tokens_out":4849,"duration_ms":43312,"temperature":0.7,"pith_summary":"Barttender is a way to put tabular health-record data on the same footing as medical images inside a deep learning model. It turns each non-imaging variable, such as age, lab values, or vital signs, into a grayscale bar, appends those bars to a chest X-ray, and trains a convolutional network on the combined image. A companion \"Blank Barttender\" appends the same bars to a blank image, so the difference in performance between the two isolates what the image adds. The paper reports that this approach matches the standard alternative, training XGBoost or logistic regression on scalar image biomarkers, while adding local and global explanations through a new metric called gIoU. A sympathetic reader would take away that imaging and tabular data can be compared directly, without losing image information to scalar summaries.","feed_headline":"Grayscale bars let deep learning compare X-rays with EHR data","feed_subtitle":"A new gIoU metric shows whether images or tabular risk factors drive each prediction.","key_machinery":"The machinery is the grayscale-bar encoding of tabular data plus the gIoU aggregation rule. Each continuous variable is quantile-normalized to $[0,1]$ and drawn as a bar whose intensity runs from white to black; categorical variables are assigned spaced values, and missing values are drawn as a red bar so missingness itself is visible to the model. Bars are concatenated to a real X-ray (Image Barttender) or to a blank image (Blank Barttender), and the same DenseNet-121 is trained on both. The gIoU measure weights the mean attribution intensity in a region by the Intersection over Union ($\\text{IoU} = |A \\cap B|/|A \\cup B|$) between the model's high-attention mask and that region, then averages over the test set. This turns pixel-level saliency maps into a population-level coefficient-like importance score, and the same construction at the sample level yields local explanations.","core_discovery":"The central claim is that converting non-imaging tabular features into grayscale bars and appending them to medical images produces a deep learning model whose predictive performance is comparable to established scalar-based comparisons, while remaining interpretable. On the MIMIC cardiomegaly task, the Image Barttender reaches an AUC of $0.86 \\pm 0.02$ and an MCC of $0.51 \\pm 0.05$, close to XGBoost with image biomarkers and tabular data ($0.83 \\pm 0.02$ AUC) and logistic regression with the same inputs ($0.88 \\pm 0.00$ AUC). The Blank Barttender, which uses only the bars, outperforms XGBoost and logistic regression trained on the same tabular features on most metrics, which the authors suggest may reflect nonlinear interactions among bars that shallow models miss. On CheXpert, the Blank Barttender replicates the relative ranking of logistic regression across conditions, and the bars' ordering does not change performance, evidence that the model attends to the bars' content rather than their position. The paper also introduces gIoU, an IoU-scaled, test-set-averaged attribution intensity, as a global feature importance measure, and shows that even when the X-ray dominates globally, individual patients can still have high bar importances.","pith_inferences":["A natural extension the authors leave implicit: because quantile normalization preserves only rank order, the Blank Barttender's edge over XGBoost suggests the convolutional network may be exploiting nonlinear interactions among bar regions; one could test this by training on bars whose values are randomly shuffled across features.","gIoU is defined for any region of an attribution map, not only bars, so it could serve as a general-purpose global importance metric for other image tasks where saliency maps are already computed.","The framework could be used as a cheap screening step before building multimodal fusion models: if the Image Barttender shows no gain over the Blank Barttender, investing in image-based deep learning for that task may be unwarranted.","A clinically meaningful next study would present the Image Barttender and its sample-level attributions to clinicians and measure whether it changes their interpretation of a prediction; the authors name this as future work but do not test it."],"forward_implications":["Researchers can compare the utility of imaging versus tabular data without building domain-specific pipelines that convert images into scalar biomarkers, shortening the analysis cycle.","The gap between Image and Blank Barttender performance gives a direct, model-controlled estimate of the added predictive value of images over risk factors.","gIoU offers a population-level feature importance ranking for image-based models, while sample-level IoU flags individual patients for whom tabular features drive the prediction.","Because missing values are encoded as a distinct bar color, the approach can model missing data without imputation and can surface missingness as a predictive signal.","The same bar-encoding recipe can be applied to other disease targets, such as diabetes or cardiovascular risk, where tabular risk scores and imaging could be compared."],"supporting_citations":[{"why":"Supplies the CheXpert dataset and the DenseNet baseline context for the proof-of-concept experiments.","marker":"(Irvin et al., 2019)"},{"why":"Provides the MIMIC cardiomegaly sample, the image biomarkers CTR and CTAR, and the XGBoost setup that Barttender is compared against.","marker":"(Duvieusart et al., 2022)"},{"why":"Supplies the MIMIC-IV and MIMIC-CXR-JPG data from which the multi-modal cardiomegaly sample is built.","marker":"(Johnson et al., a,b)"},{"why":"Provides the patient demographics (age, race, sex) used as the CheXpert tabular features.","marker":"(Banerjee et al., 2022)"},{"why":"Provides the pre-split CheXpert sample used in the experiments.","marker":"(Glocker et al., 2023)"},{"why":"Introduces gradient-based saliency maps, one of the two attribution methods aggregated by gIoU.","marker":"(Simonyan, 2013)"},{"why":"Introduces Integrated Gradients, the second attribution method used to validate the gIoU findings.","marker":"(Sundararajan et al., 2017)"}],"fun_headline_variants":["Barttender: bar-coded EHR data for interpretable imaging AI","New gIoU metric reveals which medical data truly drives predictions","Barttender turns EHR numbers into bars for fair image-vs-tabular AI","Barttender's bars make image and EHR data directly comparable","Compare X-rays and lab values fairly with Barttender's bars"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison is unbiased only if the quantile normalization of continuous variables and the selection of the eight tabular features via logistic regression Z-scores were fit on training data alone, with the test set excluded; the paper does not state that this was done.","fun_headline_variants_meta":{"raw":{"variants":["Barttender: bar-coded EHR data for interpretable imaging AI","New gIoU metric reveals which medical data truly drives predictions","Barttender turns EHR numbers into bars for fair image-vs-tabular AI","Barttender's bars make image and EHR data directly comparable","Compare X-rays and lab values fairly with Barttender's bars"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001105,"raw_usage":{"total_tokens":4638,"prompt_tokens":1005,"completion_tokens":3633,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":3539}},"tokens_in":621,"tokens_out":3633,"duration_ms":22729,"temperature":1.0,"reasoning_tokens":3539,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:13:36.112812+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the MIMIC workflow with quantile normalization and the logistic-regression feature selection performed separately inside each training fold. If the Blank Barttender no longer matches or beats XGBoost on tabular data, or if the Image Barttender's advantage shrinks, the reported parity would be traceable to leakage rather than to the bar representation.","supporting_citations":[{"cited_title":"Papież, and Adam Mahdi","cited_arxiv_id":null,"evidence_quote":"Provides the MIMIC cardiomegaly sample, the image biomarkers CTR and CTAR, and the XGBoost setup that Barttender is compared against."},{"cited_title":"Reading Race: AI Recognises Patient's Racial Identity In Medical Images","cited_arxiv_id":"2107.10356","evidence_quote":"Provides the patient demographics (age, race, sex) used as the CheXpert tabular features."}],"review_version":1}