{"id":"d89a99a1-b0df-477a-9343-edb3978767e4","arxiv_id":"2505.00228","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"ReXGradient-160K provides 160,000 chest X-ray studies with free-text reports from 109,487 patients across 79 medical sites, with public and private test splits.","lead":"This paper introduces ReXGradient-160K, a publicly available chest X-ray dataset with 160,000 studies, 273,000 images, and 109,000 patients from three U.S. health systems. It is the largest public chest X-ray collection by patient count and includes paired radiology reports for training and benchmarking report-generation models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-institutional patient deduplication is unverified; the 'unique patient' count underlying the 'largest' claim may be inflated.","rationale":"The reader's weakest assumption (patient overlap between splits) is not actually load-bearing: Table 1's per-split patient counts sum exactly to the overall 109,487, which, assuming 'Patients' means unique patients per split, implies no patient appears in more than one split. The paper should state this explicitly, but it is not a substantive risk to the central claim. The more significant gap is that the central claim—largest patient count—depends on 109,487 'unique' patients, yet the paper describes no cross-institutional deduplication method. Multi-site data collection without a common patient identifier commonly double-counts patients who receive care at multiple sites. Since the cited alternatives have roughly 65,000 patients, even moderate duplication could falsify the headline. This is a concrete, checkable threat to the paper's primary assertion, so the manuscript should be revised to disclose the patient-matching methodology and the released data should be audited before the dataset is accepted as a benchmark resource.","tokens_in":3856,"tokens_out":13050,"duration_ms":131877,"concrete_test":"After the dataset is released, download the metadata file and compute the number of distinct PatientIDs overall and per InstitutionName. If the sum of per-institution distinct PatientIDs exceeds the overall total, cross-institutional duplicates exist and the 109,487 count is inflated. Additionally, request a data-provenance statement from the authors specifying whether any cross-institutional patient-matching algorithm was used; if none was used, recompute the 'largest' comparison against MIMIC-CXR and CheXpert Plus using the per-institution union with duplicates removed.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline claim that ReXGradient-160K is the largest publicly available chest X-ray dataset in terms of patients rests on the figure of 109,487 unique patients across three health systems. Section 2.1 (Data De-identification) describes pseudonymization of names, IDs, and dates, and states that studies for the same patient undergo the same date shift, but it does not explain how patient identity was matched across the three contributing health systems. If the same individual received imaging at more than one health system and was assigned different de-identified PatientIDs per system, that person would be counted multiple times. The Introduction compares against MIMIC-CXR (65,379 patients) and CheXpert Plus (64,725 patients), so even a modest duplication rate could push the true unique count below these values and falsify the 'largest' claim. The paper must disclose whether cross-institutional patient linkage (e.g., using a master patient index or matching on PHI before de-identification) was performed. Without such disclosure, the 109,487 figure is an unverified upper bound, not a demonstrated unique-patient count.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces ReXGradient-160K, a multi-institutional chest X-ray dataset described as containing 160,000 studies with paired radiology reports from 109,487 unique patients across 3 U.S. health systems (79 medical sites). The dataset is split into training (140,000 studies), validation (10,000 studies), and public test (10,000 studies) sets, with an additional private test set of 10,000 studies reserved for the ReXrank benchmark. Images are de-identified and converted from DICOM to PNG, and reports are processed with GPT-4o into four sections: Indication, Comparison, Findings, and Impression. The central claim is that this is the largest publicly available chest X-ray dataset in terms of the number of patients.","tokens_in":4069,"tokens_out":4254,"duration_ms":43273,"significance":"If the unique-patient count and split construction are verified, this dataset is a valuable community resource: it would exceed MIMIC-CXR (65,379 patients) and CheXpert Plus (64,725 patients) in patient count, add multi-site diversity across 79 sites, and provide paired free-text reports at scale with open release on Hugging Face. The paper's internal arithmetic is consistent (273,004 images; patient counts summing to 109,487), and the demographic table suggests broadly balanced splits. I found no circularity: the dataset's size and composition are not derived from the authors' own benchmarks, and the ReXrank citations are contextual rather than load-bearing. The main risks are transparency about patient identity across institutions and about split disjointness, both of which are fixable with additional disclosure and analysis.","major_comments":[{"comment":"The unique-patient count of 109,487 is the basis for the 'largest publicly available chest X-ray dataset' claim, but the paper does not explain how patient identity was matched across the three contributing health systems. The de-identification description states that names, IDs, and dates are pseudonymized, but it does not say whether the same individual receiving imaging at more than one system was assigned the same de-identified PatientID. If cross-system matching was not performed, a patient treated at multiple sites would be counted multiple times, and because the comparator datasets are MIMIC-CXR (65,379 patients) and CheXpert Plus (64,725 patients), even a modest duplication rate could invalidate the headline comparison. Please state explicitly whether cross-institutional patient linkage was performed before de-identification (e.g., via a master patient index), or, if not, report per-system unique-patient counts and revise the claim to describe the sum of site-specific patients rather than unique persons.","section":"Section 2.1 (Data De-identification)"},{"comment":"The paper never states whether the training, validation, and test splits are disjoint by patient. The row-wise patient counts (95,716 + 6,964 + 6,807 = 109,487) strongly suggest that no patient appears in more than one split, but this is not declared. If a single patient appears in multiple splits, any benchmark results obtained from these splits would be inflated and the reported split statistics would be misleading. The authors should add an explicit statement of patient-level disjointness and describe the procedure used to assign patients to splits, including how longitudinal studies for the same patient were handled.","section":"Section 2, Table 1"},{"comment":"The extraction of four report sections using GPT-4o is described, but no validation of this extraction is reported. The text mentions a 'robust post-processing validation step' but gives no human evaluation, sample size, agreement metric, or failure analysis. Since downstream users will rely on the structured reports for training and evaluating report-generation models, the manuscript should report extraction accuracy on a held-out sample, describe how missing sections were handled, and quantify the frequency of extraction errors. Without this, the reliability of the report content as a dataset feature is unverified.","section":"Section 2.3 (Report Structure)"},{"comment":"The paper omits the study-selection criteria used to obtain the 160,000 studies from the three health systems. It is not stated whether this represents a complete consecutive sample, a random sample, or a filtered subset of available studies. This matters for assessing sampling bias and for reproducing the dataset construction, and it is directly relevant to the claim of multi-site representativeness. Please specify the inclusion and exclusion criteria, the date range of acquisition, and the number of contributing sites per health system.","section":"Section 2 (Dataset Composition)"}],"minor_comments":[{"comment":"The bullet for 'Comparison' is identical in wording to the bullet for 'Indication' ('Provides relevant patient background and reason for examination'). This appears to be a copy-paste error and should be corrected to describe the comparison with prior studies.","section":"Section 2.3 (Report Structure)"},{"comment":"The paper states that images were downsampled to 25% of their original dimensions but does not report the resulting typical resolution in pixels. Reporting the final image dimensions, or at least the range of dimensions, would help users plan model training.","section":"Section 2.2 (Image Characteristics)"},{"comment":"The metadata list includes PatientWeight, StudyDate, and EthnicGroup, but the units for weight and the date format after the 365-day shift are not specified. These should be documented in the data record description.","section":"Section 3 (Data Records)"},{"comment":"The paper says access requires signing the license on Hugging Face but does not state the license type. Please name the license (e.g., CC BY-NC 4.0 or a custom research license).","section":"Section 3 (Data Records)"},{"comment":"The statement that dates are 'shifted to within 365 days of the actual date' is ambiguous; HIPAA-compliant date shifting typically uses a random shift of up to 364 days applied consistently per patient. Clarify the exact procedure and how leap years are handled.","section":"Section 2.1 (Data De-identification)"}],"recommendation":"major_revision","confidential_remarks":"The dataset appears to be a genuinely useful resource, and the authors have disclosed the commercial relationship with Gradient Health. The main risk to the paper's central claim is the unverified cross-institutional unique-patient count; this is fixable by adding a detailed deduplication methodology and by tempering the 'largest' claim if linkage was not performed. I recommend requesting these clarifications before publication rather than rejecting, since the underlying dataset release is valuable regardless of whether the exact uniqueness count is confirmed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: ReXGradient-160K is a genuinely large, multi-site chest X-ray dataset with paired reports, and the scale is new. But the unique-patient count — the basis for the 'largest' claim — is not yet demonstrated, because the paper never says how patients were matched across the three health systems. That's a documentation gap, not proof of an inflated number, but it needs to be closed before the headline is taken at face value.\n\nWhat's good: 160,000 studies, 273,004 images, 109,487 patients across 3 health systems and 79 sites; public train/val/test splits plus a private test set for ReXrank; demographics look balanced across splits; internal counts are consistent (95,716 + 6,964 + 6,807 = 109,487). The release plan on HuggingFace is clear. This is a real new artifact, not a re-coordinatization of existing data, and no circular reasoning is involved. The authors cite their own ReXrank and MAIDA work, but the dataset stands independent of those references.\n\nSoft spots, in order of weight. First, cross-institutional patient de-duplication: the de-identification section covers pseudonymization but says nothing about linking records for the same person across sites. If a patient appears in more than one health system, they'd be counted multiple times. Even a few percent duplication would erode the margin over MIMIC-CXR's 65,379 and CheXpert Plus's 64,725 patients. This must be disclosed. Second, the paper doesn't explicitly state that the train/val/test splits are patient-disjoint. The counts sum exactly, which suggests they are, but that should be one line. Third, GPT-4o section extraction is described with a post-processing step but no validation—no human annotation or spot-check—so we don't know how clean the report sections are. Fourth, study-selection criteria are absent; how were the 160,000 studies chosen? These are all addressable documentation issues, not evidence that the data itself is bad.\n\nWho should read it: researchers working on report generation, chest X-ray representation learning, and multi-site generalization. This dataset is likely to become a useful resource even if the paper needs revision. I'd cite it once the dedup question is answered. A serious editor should send it to peer review; the issues are verification and documentation, not conceptual soundness. The authors should be asked to add a clear 'patient identity handling across sites' paragraph, a statement about disjoint splits, and a small validation study of the section extractor.","headline":"A genuinely large multi-site chest X-ray dataset with paired reports, worth using, but the paper needs to document patient identity handling before the 'largest' claim can be checked.","tokens_in":4570,"tokens_out":2789,"would_cite":true,"duration_ms":27496,"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":"The paper presents ReXGradient-160K, a public collection of 160,000 chest X-ray studies paired with free-text radiology reports from 109,487 patients across 79 medical sites, and claims it is the largest public chest X-ray dataset by…","keywords":["chest X-ray","radiology report generation","public dataset","multi-institutional","de-identification","benchmark split","medical imaging AI","ReXGradient-160K"],"falsifier":"Download the released metadata and check every anonymized patient identifier across the training, validation, and test splits; if any patient ID occurs in more than one split, then the splits are not patient-disjoint and the benchmark claims would need to be recomputed under exclusive splits.","tokens_in":3690,"feed_emoji":"🩻","tokens_out":8246,"duration_ms":73342,"temperature":0.7,"pith_summary":"ReXGradient-160K is a new public dataset of 160,000 chest X-ray studies paired with free-text radiology reports, drawn from 109,487 patients at 79 medical sites in three U.S. health systems. The paper's central claim is that, counting unique patients, this makes it the largest publicly available chest X-ray dataset to date. The authors see it as a resource for training and evaluating AI systems that generate radiology reports, because its size, multi-site origin, and consistent four-section report format support both large-scale training and tests of generalization across institutions. The dataset is released with fixed training, validation, and public test splits, as well as a private test set used for an external benchmark.","feed_headline":"By patient count, new chest X-ray dataset tops public records","feed_subtitle":"It pairs 160,000 X-ray studies with reports from 79 sites for radiology AI training.","key_machinery":"The central object is the dataset corpus itself: 273,004 unique chest X-ray images grouped into 160,000 studies, with each study linked to a four-section radiology report. The load-bearing design choices are the de-identification pipeline (HIPAA-compliant text and pixel de-identification, pseudonymized names, IDs, and date shifting that preserves within-patient time ordering), the standardized image preprocessing (DICOM-to-PNG conversion and downsampling to 25% of original dimensions), and the GPT-4o-based extraction of reports into a uniform section structure. These choices make the data usable for large-scale training and for benchmark evaluation, which is the work the central claim needs them to do.","core_discovery":"On its own terms, the discovery is a resource: a corpus of 273,004 de-identified chest radiographs organized into 160,000 studies, each associated with a report whose sections—indication, comparison, findings, impression—have been standardized. The paper reports that these studies come from 109,487 unique patients across 79 medical sites in three U.S. health systems, and it argues that this patient count makes ReXGradient-160K the largest public chest X-ray collection with paired reports. The dataset is intended to provide a more diverse and larger training and evaluation base for report-generation models, and to support studying how well such models generalize across institutions.","pith_inferences":["The 'largest public chest X-ray dataset' claim depends on the counting metric: if measured by number of studies rather than patients, an earlier collection contains more studies but fewer patients, so future comparisons should state the metric explicitly.","The report-section extraction relies on an unmeasured GPT-4o step; if extraction quality is uneven, models trained on the standardized reports could inherit errors that the paper does not quantify.","A natural next step is to test whether the private benchmark set and public splits are truly site-diverse; the paper does not report per-site image counts, so site-level leakage or imbalance remains unverified."],"forward_implications":["If the dataset is as large and as patient-diverse as reported, training corpora for chest X-ray report generation grow to more than one hundred thousand patients, with the same report vocabulary applied across sites.","The public test split gives a fixed ground for comparing report-generation models, and the private test set provides a held-out evaluation that avoids tuning directly to public test answers.","The multi-site composition makes it possible to measure how model performance shifts across health systems, a limitation of most earlier single-institution collections.","The uniform report sections allow downstream systems to use indication, findings, and impression without extra parsing, which should lower the barrier to reproducible experiments."],"supporting_citations":[{"why":"Supplies the earlier large single-institution chest X-ray collection whose patient count ReXGradient-160K claims to exceed.","marker":"[Johnson et al., 2019]"},{"why":"Provides the CheXpert Plus collection, another large public dataset with reports that serves as a scale and design comparison.","marker":"[Chambon et al., 2024]"},{"why":"Used as the early smaller image-report dataset that motivates larger report-generation resources.","marker":"[Demner-Fushman et al., 2016]"},{"why":"Supplies the multi-institutional data-sharing rationale the paper invokes against single-institution limitations.","marker":"[Saenz et al., 2024]"},{"why":"Defines the leaderboard whose private test set ReXGradient-160K feeds, linking the dataset to evaluation.","marker":"[Zhang et al., 2024]"}],"fun_headline_variants":["Patient-count record for public chest X-ray data","160k studies, 109k patients, 79 sites: new X-ray set","Largest chest X-ray dataset by patient count released","New chest X-ray dataset: most patients among public sets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on the all-important assumption that no patient appears in more than one of the training, validation, and test splits, but the paper never explicitly states that the splits are patient-disjoint; if a patient spans splits, the evaluation numbers become optimistic.","fun_headline_variants_meta":{"raw":{"variants":["Patient-count record for public chest X-ray data","160k studies, 109k patients, 79 sites: new X-ray set","Largest chest X-ray dataset by patient count released","New chest X-ray dataset: most patients among public sets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000268,"raw_usage":{"total_tokens":1581,"prompt_tokens":873,"completion_tokens":708,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":638}},"tokens_in":489,"tokens_out":708,"duration_ms":7274,"temperature":1.0,"reasoning_tokens":638,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:46:48.453393+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Download the released metadata and check every anonymized patient identifier across the training, validation, and test splits; if any patient ID occurs in more than one split, then the splits are not patient-disjoint and the benchmark claims would need to be recomputed under exclusive splits.","supporting_citations":[{"cited_title":"Preparing a collection of radiology examinations for distribution and retrieval","cited_arxiv_id":null,"evidence_quote":"Used as the early smaller image-report dataset that motivates larger report-generation resources."},{"cited_title":"The maida initiative: establishing a framework for global medical-imaging data sharing","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-institutional data-sharing rationale the paper invokes against single-institution limitations."}],"review_version":1}