Pith. sign in

REVIEW 2 major objections 6 minor 40 references

The paper claims that current LLMs, when asked to diagnose from emergency-department records available at admission, mostly retrieve diagnoses that appear verbatim in the record; diagnoses that must be inferred are recovered only 3–31% of t

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 00:22 UTC pith:QQEXUV7Z

load-bearing objection A well-built admission-anchored benchmark with unusually thorough robustness checks; the main caveat is that the headline extraction-vs-inference gap is scored by a judge from the same model family as the post-trained system, so the exact margins need a sensitivity check before I'd trust them. the 2 major comments →

arxiv 2607.28788 v1 pith:QQEXUV7Z submitted 2026-07-30 cs.AI

EarlyDx: An Admission-Anchored Benchmark for Open-Ended Generation of Evidence-Supported ED-Encounter Diagnoses

Jiahui Li , Ruili Fang , Zishuai Liu , Yutong Guo , Nan Yang , Wenzhan Song , Jin Lu , Fei Dou This is my paper
classification cs.AI
keywords early diagnosisemergency departmentLLM evaluationevidence groundingextraction vs inferenceopen-ended generationadmission-time recordsclinical benchmarks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

EarlyDx is a benchmark for open-ended diagnosis generation at hospital admission, built from 154,834 emergency-department encounters and restricted to records available at admission time. Its key move is to not trust recorded diagnoses as ground truth: an LLM auditor tags each free-text diagnosis as supported, partially supported, or unsupported by the admission-time evidence, and the primary score counts only fully supported labels. On this benchmark, zero-shot LLMs—general and medical alike—score mostly by extraction: they recover 37–65% of diagnoses that appear verbatim in the record but only 3–31% of diagnoses that must be inferred. Fine-tuning a small model on evidence-grounded chain-of-thought labels raises inference-dependent recall to 56%, but no system reaches a clinician's operating point on time-critical conditions. The paper's point is that admission-time diagnostic competence, measured this way, is largely a lexical-retrieval skill in current LLMs.

Core claim

The central claim is that no evaluated system—frontier general-purpose, medical-specialized, or in-domain post-trained—reliably synthesizes admission-time evidence into open-ended diagnoses. Zero-shot systems are competitive on explicit labels (diagnoses named in the input) but collapse on implicit labels, which form 57% of supported diagnoses and must be inferred: general models recover 9–31% and medical-specialized models 3–6%. Post-training a 4-billion-parameter model on gold-conditioned rationales recovers 56% of implicit labels while keeping explicit recall high, showing the gap is partly teachable but far from closed. On six time-critical conditions the clinician's reference point is 7

What carries the argument

The load-bearing piece is the evidence verifier: an LLM auditor that classifies each ED-encounter diagnosis label as supported (a direct admission-time finding substantiates it), partially supported (only indirect cues, e.g., prior history or home medications), or unsupported (depends on later or outside information). The primary evaluation track scores only fully supported labels. The second instrument is the explicit/implicit partition of supported labels: a diagnosis is explicit if its full title or every content word appears verbatim in the input, implicit otherwise. Comparing recall on the two subsets is what separates extraction from inference. Supporting machinery is the admission-anc

Load-bearing premise

The central extraction-versus-inference conclusion depends on the LLM auditor's 'supported' label being a valid proxy for what a clinician could infer from the admission-time record; because ED diagnoses carry no per-diagnosis timestamps, some 'supported' labels may depend on information that only became available after admission, which would make the implicit-recall gap and the post-training benefit partly artifacts of the label boundary.

What would settle it

Take a sample of encounters and have emergency clinicians, reading only the admission-time record, independently mark which recorded ED diagnoses they could have inferred at that moment; if the auditor's 'supported' labels agree poorly with clinician-inferable labels, or if the explicit/implicit recall gap shrinks when using clinician-marked labels, the paper's central dissociation is an artifact of the LLM auditor's boundary. A second check: obtain per-diagnosis timestamps from a hospital that records them and verify that 'supported' labels were actually documented before or at admission; if

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Evaluating early-diagnosis models against discharge-coded labels overstates their ability; admission-anchored, evidence-verified labels measure a different and harder capability.
  • Small task-aligned models can substantially outperform much larger zero-shot models on this task (0.51 vs 0.40 supported F1), so parameter count is not the bottleneck; post-training on evidence-grounded rationales is.
  • A system's score on explicit labels is not predictive of its score on implicit labels; any diagnostic benchmark should report them separately to avoid crediting extraction.
  • Risk-weighted evaluation changes conclusions: systems ranked first by aggregate F1 can still be worst on the sensitivity-precision trade-off for time-critical conditions.
  • Because the human clinician also over-lists relative to the coded reference (2.58 vs 1.4 diagnoses per encounter), part of the precision gap is a property of the billing-code reference, not of the models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Testable extension: if the explicit/implicit dissociation replicates on other clinical corpora, it offers a cheap audit—re-score any claimed diagnostic-reasoning result on labels not stated in the input; large recall drops would indicate retrieval rather than reasoning.
  • The paper's proposal to let systems emit 'suspected, pending confirmation' diagnoses under an asymmetric loss could invert the risk-weighted ranking: zero-shot models with broad differentials would be credited rather than penalized, and the 'no system matches the clinician' conclusion would need re-examination.
  • The monotone performance decline with patient age across all systems, including models never trained on EarlyDx, suggests a label-density confound (older patients have 1.72 vs 1.42 diagnoses per encounter) rather than model bias; stratifying by number of reference labels is a natural control.
  • The moderate inter-rater agreement on the supported/partial boundary (κ=0.51–0.64) means the exact boundary is noisy; the paper's own stricter human-confirmed gold standard widened the post-training advantage, so the main conclusions appear robust to boundary choice—a claim worth testing on a larger human-audited sample.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. The paper introduces EarlyDx, a benchmark for open-ended, evidence-grounded diagnosis generation from 154,834 MIMIC-IV emergency department encounters. Inputs are clipped to records available at hospital admission time t0, supervision is taken from ED-encounter diagnoses (not discharge codes), and each free-text diagnosis label is annotated by an LLM auditor as supported, partially supported, or unsupported by the admission-time evidence. Evaluation uses a fixed LLM-as-judge (qwen3.5-27B) with cached outputs. The central empirical claim is that zero-shot LLMs—frontier general and medical-specialized alike—perform largely by extracting diagnoses that appear in the input, recovering only 3–31% of implicit (non-verbatim) supported labels, whereas a fine-tuned Qwen3.5-4B model attains 56% implicit recall and a supported-track F1 of 0.51 versus 0.40 for the best zero-shot model. The paper also reports that no system reaches a clinician's operating point on six time-critical conditions. Extensive robustness checks are included: dual-judge agreement, manual gold-matching review, human re-audit of evidence classes, timestamp sensitivity, cardinality controls, contamination probes, and subgroup analyses.

Significance. If the empirical claims hold, EarlyDx is a valuable contribution. It addresses genuine deficiencies in existing medical-LLM benchmarks: closed code sets, exclusion of free-text evidence, and supervision by discharge diagnoses that leak post-admission information. The methodology for evidence-grounded label verification is carefully developed, and the paper is admirably thorough—it ships a reproducible pipeline with cached judge/verifier outputs, version-pinned models, human audits of both evidence classes and judge decisions, and multiple sensitivity analyses. The extraction-versus-inference dissociation is a clinically meaningful finding that would be of broad interest. However, the central quantitative claims rest on a single LLM judge whose model family overlaps with the post-trained system, and the existing controls do not fully rule out judge-family bias. The significance is therefore conditional: if an independent-family judge reproduces the rankings, this is a strong paper; if not, the headline numbers are artifacts. The label-boundary construct is another risk, though it is substantially mitigated by the human re-audit and ranking-invariance analysis in Appendix B.

major comments (2)
  1. [§3.5, Table 2, Fig. 9] The primary judge is qwen3.5-27B and the only post-trained model is Qwen3.5-4B, from the same family. All headline numbers—supported-track F1 (0.51 vs 0.40), implicit-label recall (56% vs 3–31%), and the claimed post-training benefit—are computed through this judge. The two controls do not fully exclude family bias: the 'second, independent judge' is unnamed, so its family is unverified; and the manual gold review covers only 100 pairs, far too few to detect a bias concentrated in the implicit subset (57% of supported labels) where the main gap resides. The Appendix B re-audit concerns the evidence-class boundary, not the judge's semantic matching, and also uses the same judge for scoring. Concretely, re-score all 6,975 test predictions with a judge from a different family (e.g., GPT-5.5, Claude, or a non-Qwen open-weight model) and report rank invariance, especially on implicit labels;
  2. [§3.3, §5.2, Fig. 9] The explicit/implicit partition is purely lexical: a supported label is 'implicit' if its full title or all content words do not appear verbatim. This conflates two distinct things—'not stated in words' and 'requires diagnostic inference beyond reading off a specific finding.' A diagnosis such as hyperkalemia may be lexically implicit yet directly evidenced by a potassium value, so a model that extracts it from a lab result is scored as 'inference.' This construct-validity concern is load-bearing for the central extraction-vs-inference dissociation. The human clinician's null gap (45% vs 48%) suggests the partition does carry something real, and Appendix B's ranking invariance helps, but the analysis does not directly test whether the implicit labels are truly inferential rather than merely non-verbal. Please re-analyze the partition controlling for whether the supporting evidence is a s
minor comments (6)
  1. [Abstract] The URL 'here' is a placeholder; a permanent link should be provided.
  2. [§3.5] The second judge is described only as 'independent'; specify the exact model, version, and confirm it is not from the Qwen family.
  3. [Appendix D] The micro-F1 values in Table 6 (e.g., 0.546) differ from those in Table 2 for the same model (0.51). Clarify whether Table 6 is computed on a subset or with a different configuration.
  4. [Appendix J] The section headings 'B.1 Evidence Verifier', 'B.2 ...' are misnumbered; these should be J.1, J.2, etc.
  5. [Fig. 11] The number of true positives among the 46 critical alerts raised by Claude in 200 encounters should be stated explicitly.
  6. [Appendix G] The text uses 'MiMo-V2.5' while reference [27] is titled 'MiMo-V2-Flash'; align the version name.

Circularity Check

0 steps flagged

No circularity found; the benchmark's labels, evidence partition, and evaluation are not defined in terms of the fitted outputs, and the judge-family overlap is an evaluation-bias concern rather than a by-construction circular step.

full rationale

Walking the claimed derivation chain: EarlyDx constructs admission-time inputs from MIMIC-IV, takes ED-encounter ICD titles as open-vocabulary reference labels, and uses an LLM auditor (MiniMax M3) to mark each label supported/partial/unsupported, with a four-clinician re-audit in Appendix B (76% agreement, kappa 0.51-0.64) and a stricter gold-standard analysis showing invariant rankings. The extraction-vs-inference partition is defined lexically (explicit = full title or every content word appears verbatim; implicit otherwise), not by the judge, so the headline implicit-recall gap is not an input to the construction. The post-trained Qwen3.5-4B is fine-tuned on the training split with gold-conditioned rationales and evaluated on a patient-disjoint test split; this is standard supervised evaluation, not a fitted parameter being renamed as a prediction. No equation in the paper defines a reported quantity in terms of the fitted quantity, no load-bearing uniqueness claim is imported from the authors' own prior work, and no ansatz is smuggled in via citation. The strongest concern visible in the text is that the evaluation judge is qwen3.5-27B while the post-trained model is Qwen3.5-4B, so the judge and the headline system share a model family; if the judge is systematically more lenient toward Qwen-style output, the post-training advantage and the 56% implicit recall could be inflated. The paper's controls (an unnamed second judge with 89.7% exact-count agreement, and a 100-pair manual gold review with 94% agreement) mitigate but do not fully close this, and the paper itself releases cached judge outputs making the question testable. That is a genuine evaluation-bias risk and a limitation of the paper, but it is not a circular reduction: the benchmark's labels and evidence classes do not reduce to the judge's matching decisions, and the ranking-invariance analyses provide independent content. Accordingly, no circular step is identified; the appropriate finding is 'no significant circularity', with the judge-family concern registered as a correctness risk rather than a circularity score driver.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The benchmark's central claims rest on five stated assumptions about the fidelity of MIMIC-IV-ED billing labels, timestamp filtering, LLM-based evidence classification, LLM-based semantic matching, and lexical explicitness. The authors stress-test most of them (storetime sensitivity, human re-audit, second judge, manual matching), but none is a derived result; each is a domain assumption that a critic could probe further. The only hand-chosen parameter is the evidence window W=0, with sensitivity checks at 6/24 h.

free parameters (1)
  • Evidence cutoff window W = 0 h (main benchmark); 6/24 h sensitivity
    All headline numbers use W=0, meaning evidence is clipped to records charted at or before admission time. This hand-chosen cutoff determines input completeness; Appendix L shows F1 rises with W but the central comparison survives.
axioms (5)
  • domain assumption MIMIC-IV-ED diagnosis table entries are accurate records of ED-encounter diagnoses.
    Section 3.3 uses ICD-10 titles from the ED diagnosis table as reference labels after filtering symptom/ill-defined codes. These are administrative billing codes; if they misrepresent what a clinician actually concluded at admission, the gold set is mis-specified. The authors do not adjudicate remaining codes.
  • domain assumption Admission-time clipping at t0 with charttime filtering preserves the information actually available at admission.
    Section 3.2 filters on charttime, which precedes storetime by a median of 1.2–1.8 h; Appendix D shows stricter storetime filtering lowers F1 by 0.034, so results are an upper bound but rankings survive.
  • domain assumption The LLM auditor's supported/partial/unsupported classification validly operationalizes 'inferable from admission-time evidence'.
    Section 3.3 and Appendix B; human re-audit shows 76% agreement at ≥3/4 experts and κ=0.51–0.64, so the primary track's boundary is moderately reliable but far from perfect.
  • domain assumption The LLM-as-judge semantic match is a valid measure of diagnosis equivalence.
    Section 3.5; exact agreement with a second judge 89.7% and with manual review 94% on 100 pairs. Still an LLM proxy for clinical equivalence, and the judge's family overlaps with the post-trained model.
  • domain assumption Lexical explicitness separates extraction from inference.
    Section 5.2 defines explicit as full title or every content word occurring verbatim; paraphrased mentions count as implicit, so the extraction/inference dissociation may be overstated. The authors note 43% is a lower bound on the extractable fraction.

pith-pipeline@v1.3.0-alltime-deepseek · 23503 in / 17128 out tokens · 175957 ms · 2026-08-03T00:22:49.419626+00:00 · methodology

0 comments
read the original abstract

Clinical diagnosis at hospital admission must be made rapidly from limited, incomplete evidence. Existing diagnosis-prediction benchmarks are poorly suited to this setting: they restrict prediction to closed code sets, exclude free-text notes, and supervise with discharge diagnoses that incorporate the full inpatient course. We introduce EarlyDx, a large-scale benchmark for open-ended early diagnosis, built from 154,834 emergency department encounters in MIMIC-IV. Each encounter is restricted to records available at admission time $t_0$ and supervised by the diagnoses recorded during the ED encounter rather than at discharge. An LLM auditor further verifies every free-text label as supported, partially supported, or unsupported by that evidence; the primary evaluation scores only fully supported labels. Under a semantic LLM-as-judge protocol, no evaluated system --- frontier general, medical-specialized, or in-domain post-trained --- synthesizes admission-time evidence reliably. Zero-shot models score largely by extraction, recovering only 3-31% of diagnoses that must be inferred rather than read from the record; post-training raises inference-dependent recall to 56%, but a sizeable margin remains, and on time-critical conditions no system attains a clinician's balance of sensitivity and precision. We release the full construction and evaluation pipeline at here.

Figures

Figures reproduced from arXiv: 2607.28788 by Fei Dou, Jiahui Li, Jin Lu, Nan Yang, Ruili Fang, Wenzhan Song, Yutong Guo, Zishuai Liu.

Figure 1
Figure 1. Figure 1: Information lost when free-text diagnoses are con [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the EarlyDx framework. From MIMIC-IV emergency-department encounters, we build an admission [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Modality coverage in EarlyDx at 𝑊 =0: fraction of the 154,834 encounters for which each input field is present. Presentation and triage vitals are universal; ECG and radi￾ology cover 74% and 65%; in-window laboratory results only 16%, since admitted-patient results are timestamped at or after 𝑡0. Colors group modalities by type. readable in the record (storetime) by a median of 1.2 h for labora￾tory result… view at source ↗
Figure 4
Figure 4. Figure 4: Agreement between two independent LLM judges [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Agreement between the LLM judge and manual [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The EarlyDx fine-tuning framework. For each en [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Complete-match rate—the fraction of encounters [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Cardinality-controlled zero-shot baselines (all [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: reports recall on each subset.1 Every system performs worse on implicit labels, confirming the partition isolates genuinely harder instances. The informative result is a dissociation between families: zero-shot models are competitive on explicit labels (37– 65%) yet collapse on implicit ones, recovering 9–31% (general) and 3– 6% (medical), so their apparent competence is largely attributable to extraction.… view at source ↗
Figure 10
Figure 10. Figure 10: Average number of predicted diagnoses per en [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Supported-label recall at a matched prediction [PITH_FULL_IMAGE:figures/full_fig_p010_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Cohort composition and evidence availability by [PITH_FULL_IMAGE:figures/full_fig_p012_13.png] view at source ↗
Figure 15
Figure 15. Figure 15: Exploratory primary supporting-evidence modal [PITH_FULL_IMAGE:figures/full_fig_p013_15.png] view at source ↗
Figure 14
Figure 14. Figure 14: Exploratory primary supporting-evidence modal [PITH_FULL_IMAGE:figures/full_fig_p013_14.png] view at source ↗
Figure 16
Figure 16. Figure 16: Training loss for the 2B and 4B rational-SFT mod [PITH_FULL_IMAGE:figures/full_fig_p014_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Chain-of-thought traces for a supported (left) and a partially supported (right) diagnosis. Supported diagnoses [PITH_FULL_IMAGE:figures/full_fig_p016_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Human vs. our fine-tuned model’s chain-of-thought for the same admission (asthma/COPD exacerbation). Both reach [PITH_FULL_IMAGE:figures/full_fig_p016_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Clinical review of gold-conditioned rationales. [PITH_FULL_IMAGE:figures/full_fig_p017_19.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 12 linked inside Pith

  1. [1]

    Juan Miguel Lopez Alcaraz, Hjalmar Bouma, and Nils Strodthoff. 2024. MDS- ED: Multimodal Decision Support in the Emergency Department–a Benchmark Dataset for Diagnoses and Deterioration Prediction in Emergency Medicine. ArXiv Preprint(2024)

    2024

  2. [2]

    Juan Miguel Lopez Alcaraz, Hjalmar Bouma, and Nils Strodthoff. 2025. En- hancing clinical decision support with physiological waveforms—a multimodal benchmark in emergency care.Computers in biology and medicine192 (2025), 110196

    2025

  3. [3]

    Anthropic. 2025. Claude Opus 4.8. https://www.anthropic.com/claude. Large language model. Accessed via API, model IDclaude-opus-4-8

    2025

  4. [4]

    Aaron Blakeman, Aaron Grattafiori, Aarti Basant, Abhibha Gupta, Abhinav Khat- tar, Adi Renduchintala, Aditya Vavre, Akanksha Shukla, Akhiad Bercovich, Alek- sander Ficek, et al. 2025. NVIDIA Nemotron 3: Efficient and Open Intelligence. arXiv preprint arXiv:2512.20856(2025)

    Pith/arXiv arXiv 2025

  5. [5]

    Emma Chen, Aman Kansal, Julie Chen, Boyang Tom Jin, Julia Reisler, David E Kim, and Pranav Rajpurkar. 2023. Multimodal clinical benchmark for emergency care (mc-bec): A comprehensive benchmark for evaluating foundation models in emergency medicine.Advances in Neural Information Processing Systems36 (2023), 45794–45811

    2023

  6. [6]

    Haodong Chen, Guido Zuccon, and Teerapong Leelanupab. 2025. Beyond GeneGPT: A Multi-Agent Architecture with Open-Source LLMs for Enhanced Genomic Question Answering. InProceedings of the 2025 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region. 143–152

    2025

  7. [7]

    Junying Chen, Zhenyang Cai, Ke Ji, Xidong Wang, Wanlong Liu, Rongsheng Wang, Jianye Hou, and Benyou Wang. 2024. Huatuogpt-o1, towards medical complex reasoning with llms.arXiv preprint arXiv:2412.18925(2024)

    Pith/arXiv arXiv 2024

  8. [8]

    Arsene Fansi Tchango, Rishab Goel, Zhi Wen, Julien Martel, and Joumana Ghosn

  9. [9]

    Dyke Ferber, Lars Hilgers, Christiane Höper, Benedict Kinny-Köster, Jan-Niklas Eckardt, Katharina Egger-Heidrich, Marius Bill, Martin MK Schneider, Jan Clus- mann, Lejla Kadric, et al. 2026. Towards autonomous medical artificial intelligence agents.Nature(2026), 1–10

    2026

  10. [10]

    JA Hirsch, G Nicola, G McGinty, RW Liu, RM Barr, MD Chittle, and L Manchikanti

  11. [11]

    Namgyu Ho, Laura Schmid, and Se-Young Yun. 2023. Large language models are reasoning teachers. InProceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers). 14852–14882

    2023

  12. [12]

    Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. 2023. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. InFindings of the Association for Computational Linguistics: ACL 2023. 8003–8017

    2023

  13. [13]

    Kexin Huang, Jaan Altosaar, and Rajesh Ranganath. 2019. Clinicalbert: Modeling clinical notes and predicting hospital readmission.arXiv preprint arXiv:1904.05342 (2019)

    Pith/arXiv arXiv 2019

  14. [14]

    Qiao Jin, Won Kim, Qingyu Chen, Donald C Comeau, Lana Yeganova, W John Wilbur, and Zhiyong Lu. 2023. Medcpt: Contrastive pre-trained transformers with large-scale pubmed search logs for zero-shot biomedical information retrieval. Bioinformatics39, 11 (2023), btad651

    2023

  15. [15]

    Alistair EW Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, Ayad Shammout, Steven Horng, Tom J Pollard, Sicheng Hao, Benjamin Moody, Brian Gow, et al

  16. [16]

    Suhyeon Lee, Won Jun Kim, Jinho Chang, and Jong Chul Ye. 2024. LLM-CXR: Instruction-finetuned LLM for CXR image understanding and generation. In International Conference on Learning Representations, Vol. 2024. 29745–29765

    2024

  17. [17]

    Lucie Charlotte Magister, Jonathan Mallinson, Jakub Adamek, Eric Malmi, and Aliaksei Severyn. 2023. Teaching small language models to reason. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). 1773–1781

    2023

  18. [18]

    MiniMax, :, Aili Chen, Aonian Li, Bangwei Gong, Binyang Jiang, Bo Fei, Bo Yang, Boji Shan, Changqing Yu, Chao Wang, Cheng Zhu, Chengjun Xiao, Chengyu Du, Chi Zhang, Chu Qiao, Chunhao Zhang, Chunhui Du, Congchao Guo, Da Chen, Deming Ding, Dianjun Sun, Dong Li, Enwei Jiao, Haigang Zhou, Haimo Zhang, Han Ding, Haohai Sun, Haoyu Feng, Huaiguang Cai, Haichao Z...

    Pith/arXiv arXiv 2025

  19. [19]

    Aishik Nagar, Arun-Kumar Kaliya-Perumal, Yu-Hsuan Han, Andrew Sheng-Han Huang, Kristen Kee, Yushi Cao, Yiming Chen, and Hongchao Jiang. 2026. CLR- voyance: Reinforcing Open-Ended Reasoning for Inpatient Clinical Decision Support with Outcome-Aware Rubrics.arXiv preprint arXiv:2605.09584(2026)

    Pith/arXiv arXiv 2026

  20. [20]

    Kimberly J O’malley, Karon F Cook, Matt D Price, Kimberly Raiford Wildes, John F Hurdle, and Carol M Ashton. 2005. Measuring diagnoses: ICD code accuracy. Health services research40, 5p2 (2005), 1620–1639

    2005

  21. [21]

    Samuel Schmidgall, Rojin Ziaei, Carl Harris, Eduardo Reis, Jeffrey Jopling, and Michael Moor. 2024. Agentclinic: a multimodal agent benchmark to evaluate ai in simulated clinical environments.arXiv preprint arXiv:2405.07960(2024)

    Pith/arXiv arXiv 2024

  22. [22]

    Andrew Sellergren, Sahar Kazemzadeh, Tiam Jaroensri, Atilla Kiraly, Madeleine Traverse, Timo Kohlberger, Shawn Xu, Fayaz Jamil, Cían Hughes, Charles Lau, et al. 2025. Medgemma technical report.arXiv preprint arXiv:2507.05201(2025)

    Pith/arXiv arXiv 2025

  23. [23]

    Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al

  24. [24]

    Karan Singhal, Tao Tu, Juraj Gottweis, Rory Sayres, Ellery Wulczyn, Mohamed Amin, Le Hou, Kevin Clark, Stephen R Pfohl, Heather Cole-Lewis, et al . 2025. Toward expert-level medical question answering with large language models. Nature medicine31, 3 (2025), 943–950

    2025

  25. [25]

    Alex Stinard. 2026. ClinicalBench: Stress-Testing Assertion-Aware Retrieval for Cross-Admission Clinical QA on MIMIC-IV.arXiv preprint arXiv:2605.11143 (2026)

    Pith/arXiv arXiv 2026

  26. [26]

    Sameer Sundrani, Julie Chen, Boyang Tom Jin, Zahra Shakeri Hossein Abad, Pranav Rajpurkar, and David Kim. 2023. Predicting patient decompensation from continuous physiologic monitoring in the emergency department.NPJ digital medicine6, 1 (2023), 60

    2023

  27. [27]

    Core Team, Bangjun Xiao, Bingquan Xia, Bo Yang, Bofei Gao, Bowen Shen, Chen Zhang, Chenhong He, Chiheng Lou, Fuli Luo, Gang Wang, Gang Xie, Hailin Zhang, Hanglong Lv, Hanyu Li, Heyu Chen, Hongshen Xu, Houbin Zhang, Huaqiu Liu, Jiangshan Duo, Jianyu Wei, Jiebao Xiao, Jinhao Dong, Jun Shi, Junhao Hu, Kainan Bao, Kang Zhou, Lei Li, Liang Zhao, Linghao Zhang,...

    Pith/arXiv arXiv 2026

  28. [28]

    Qwen Team. 2026. Qwen3. 5-omni technical report.arXiv preprint arXiv:2604.15804(2026)

    Pith/arXiv arXiv 2026

  29. [29]

    Tao Tu, Shekoofeh Azizi, Danny Driess, Mike Schaekermann, Mohamed Amin, Pi-Chuan Chang, Andrew Carroll, Charles Lau, Ryutaro Tanno, Ira Ktena, et al

  30. [30]

    Dave Van Veen, Cara Van Uden, Maayane Attias, Anuj Pareek, Christian Blueth- gen, Malgorzata Polacin, Wah Chiu, Jean-Benoit Delbrouck, Juan Zambrano Chaves, Curtis Langlotz, et al . 2023. RadAdapt: Radiology report summariza- tion via lightweight domain adaptation of large language models. InThe 22nd Workshop on Biomedical Natural Language Processing and ...

    2023

  31. [31]

    Shirui Wang, Zhihui Tang, Huaxia Yang, Qiuhong Gong, Tiantian Gu, Hongyang Ma, Yongxin Wang, Wubin Sun, Zeliang Lian, Kehang Mao, et al. 2025. A novel evaluation benchmark for medical LLMs illuminating safety and effectiveness in clinical domains.npj Digital Medicine(2025)

    2025

  32. [32]

    Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. 2022. Star: Boot- strapping reasoning with reasoning.Advances in Neural Information Processing Systems35 (2022), 15476–15488

    2022

  33. [33]

    Aohan Zeng, Xin Lv, Zhenyu Hou, Zhengxiao Du, Qinkai Zheng, Bin Chen, Da Yin, Chendi Ge, Chenghua Huang, Chengxing Xie, et al. 2026. Glm-5: from vibe KDD ’26, August 2026, Location Li et al. coding to agentic engineering.arXiv preprint arXiv:2602.15763(2026)

    Pith/arXiv arXiv 2026

  34. [34]

    Yuting Zhang, Kaishen Yuan, Hao Lu, Yutao Yue, Jintai Chen, and Kaishun Wu. 2026. Medtvt-r1: A multimodal llm empowering medical reasoning and diagnosis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 35248–35259

    2026

  35. [35]

    dx ": " SUBARACHNOID HEMORRHAGE

    Dimitrios Zikos, Aashara Shrestha, and Leonidas Fegaras. 2019. Estimation of the mismatch between admission and discharge diagnosis for respiratory patients, and implications on the length of stay and hospital charges.AMIA Summits on Translational Science Proceedings2019 (2019), 192. A A cardinality-tolerant view: recall at a matched budget Micro-F1 penal...

    2019

  36. [2016]

    ICD-10: history and context.American Journal of Neuroradiology37, 4 (2016), 596–599

    2016

  37. [2022]

    Ddxplus: A new dataset for automatic medical diagnosis.Advances in neural information processing systems35 (2022), 31306–31318

    2022

  38. [2023]

    MIMIC-IV, a freely accessible electronic health record dataset.Scientific data10, 1 (2023), 1

    2023

  39. [2024]

    Towards generalist biomedical AI.Nejm Ai1, 3 (2024), AIoa2300138

    2024

  40. [2025]

    Openai gpt-5 system card.arXiv preprint arXiv:2601.03267(2025)

    Pith/arXiv arXiv 2025