Pith. sign in

REVIEW 3 major objections 5 minor 42 references

Latent compression is the best way to fuse ECG, X-ray, and notes into an autoregressive EHR model, yet the added modalities do not reliably beat the EHR-only baseline on zero-shot ICU mortality.

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 ↗

Latent compression improves multimodal fusion, but ICU mortality predictions from EHR plus ECG, CXR, and notes are generally no better than EHR-only.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Useful, honest ablation of compression and encoder choice for multimodal EHR generation, but single-run metrics and test-selected hyperparameters make the rankings provisional. the 3 major comments →

arxiv 2607.22264 v1 pith:VT5UIUP4 submitted 2026-07-24 cs.LG

Autoregressive EHR Foundation Models with Multimodal Inputs

Yuxuan Liu , Joshua Placidi , Jinpei Han , Alfred John Balston , Marek Rei , A. Aldo Faisal This is my paper
classification cs.LG
keywords EHR foundation modelsmultimodal fusionlatent compressiongated cross-attentionzero-shot clinical predictionICU mortalitytemporal alignmentencoder choice
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.

The reading

The paper builds a framework for conditioning an autoregressive EHR foundation model on three auxiliary clinical modalities — ECG waveforms, chest X-ray images, and clinical notes — using frozen pretrained encoders, a Perceiver-style latent compression module, and gated cross-attention with past-only temporal masking. Through controlled ablations, it shows that compressing each modality into a small set of learned latent tokens (k=8, l=4) outperforms both uncompressed cross-attention and mean pooling on zero-shot ICU mortality prediction, while cutting attention cost from O(NT) to O(Nk+kT). It also shows that the choice of pretrained encoder matters within each modality: domain-adapted encoders consistently beat general-purpose ones. Most surprisingly, the multimodal additions do not reliably improve over an EHR-only baseline: only notes encoded with BioMedBERT exceed it on both AUROC and AUPRC, and a unified three-modality model underperforms it. The paper concludes that careful fusion design and clinically contextual evaluation matter more than the sheer number of modalities.

Core claim

The central claim is that latent compression plus gated cross-attention is the right way to inject auxiliary modalities into an autoregressive EHR transformer, but that doing so does not guarantee gains. In the compression ablation, the best latent configuration (k=8, l=4) reaches AUROC 0.8751 and AUPRC 0.4599, beating no-compression and mean pooling. In the encoder ablation, only the EHR+notes model with BioMedBERT exceeds the EHR-only baseline on both AUROC and AUPRC (0.8751/0.4599 vs 0.8651/0.4561), while the unified model with all three modalities falls to 0.8460 AUROC. The paper interprets the failure as an optimization and signal-balance problem, not as evidence that the modalities lac

What carries the argument

The key machinery is a latent compression module inspired by Perceiver: each modality's encoder output (e.g., a long ECG waveform or document) is distilled into k learned latent vectors through l layers of cross-attention, reducing fusion cost to O(Nk + Tk). These latents feed gated cross-attention blocks (a Flamingo-style per-layer learnable scalar gate) interleaved with the decoder's self-attention, and a past-only masking rule prevents the model from attending to modality events acquired after the prediction point.

Load-bearing premise

The paper's rankings treat single-run, rollout-based AUROC/AUPRC differences, often smaller than 0.01, as meaningful signal, even though the authors state the evaluation is stochastic and noisy and was not repeated across seeds.

What would settle it

Re-run the two key comparisons — latent (k=8,l=4) versus mean pooling, and EHR+notes (BioMedBERT) versus EHR-only — across many random seeds or with paired bootstrapping. If the 95% confidence intervals on AUROC/AUPRC overlap for those pairs, the paper's central rankings and its 'no guaranteed improvement' conclusion would not survive.

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

If this is right

  • Latent compression is essential for effective and efficient fusion of long modality sequences; it improves performance while bounding compute.
  • Encoder choice has a clear within-modality effect, so effort spent selecting or adapting the frozen modality encoder may yield larger returns than additional fusion-side complexity.
  • Adding modalities does not guarantee improvement over a strong EHR-only baseline; future work should target more flexible fusion and modality-aware training.
  • The negative result is plausibly an optimization and signal-balance problem (modality competition), not evidence that the modalities lack mortality-relevant information.
  • Temporally sparse modalities like CXR tend to hurt after past-only masking, while notes with a domain-adapted encoder help, suggesting availability and informativeness jointly determine utility.

Where Pith is reading between the lines

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

  • If the observed single-run differences (often <0.01 AUROC/AUPRC) are within rollout noise, the ranking of compression methods and the 'no guaranteed improvement' conclusion could both be artifacts; multi-seed or bootstrap replication would settle this.
  • The latent bottleneck's benefit may generalize beyond ECG/CXR/notes to other long-sequence modalities (continuous vitals, genomics) in EHR models, acting as a regularizer as well as a compute saver.
  • A missingness-aware training objective — rewarding the model for exploiting available modalities or conditioning on their absence — might convert the negative result into gains by countering modality competition.
  • Cohort-stratified evaluation restricted to time points where a modality is actually present (e.g., patients with a recent CXR) could reveal signal that the aggregate negative result obscures.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 5 minor

Summary. The paper proposes a multimodal conditioning framework for autoregressive EHR foundation models. Frozen pretrained encoders (ECG, CXR, clinical notes) produce token sequences that are compressed into k learned latent vectors through l cross-attention layers and then fused into a GPT-2-style EHR decoder via gated cross-attention with temporal alignment. On MIMIC-IV, the authors evaluate zero-shot ICU mortality using rollout sampling. They report three main findings: learned latent compression outperforms uncompressed cross-attention and mean pooling; encoder choice has a clear within-modality effect; and adding auxiliary modalities does not reliably improve over an EHR-only baseline. The paper is candid that the rollout evaluation is stochastic and that experiments were not repeated across seeds, and it frames the results as broad trends.

Significance. If the reported trends hold, this is a useful controlled ablation for a practically important problem: it shows that fusion architecture and encoder selection matter and that simply adding modalities can hurt. The strengths of the paper are the explicit temporal-alignment masking, missing-modality masks, controlled comparisons over multiple pretrained encoders, and an unusually transparent limitations section. The main negative result — that multimodal conditioning does not beat a strong EHR-only baseline — is a valuable and falsifiable clinical finding. However, the current empirical support is thin: single-run rollout metrics, no confidence intervals, no repeated seeds, and test-set-based selection of the winning configuration. The paper is therefore best read as a proof-of-concept with an important negative result that needs stronger statistical grounding before the comparative claims are established.

major comments (3)
  1. [Section 3, Tables 1-2] The paper concedes in Section 3 that the rollout-based evaluation is stochastic and noisy and that experiments were not repeated across multiple random seeds. Combined with K=20 rollouts per prompt (Section 2.4), each per-prompt score is a multiple of 0.05, so the reported AUROC/AUPRC values are discrete and subject to Monte Carlo error. Several headline comparisons are within or close to this plausible noise range: EHR-only vs. EHR+notes (BioMedBERT) is 0.8651 vs. 0.8751 AUROC and 0.4561 vs. 0.4599 AUPRC; the unified model is 0.8460 vs. 0.8651 AUROC. Without confidence intervals, bootstraps, or seed-level replication, the ranking of compression methods and the negative result about added modalities are not strongly supported. The authors' own caveat ('broad trends rather than definitive rankings') should be reflected in the abstract and conclusion, which currently assert that latent com
  2. [Section 3, Table 1] The choice k=8, l=4 is selected as the setting with the best AUPRC on the very table used to demonstrate its superiority, and Section 2.1 does not state a validation split for this sweep. Since Table 1 appears to report test-table performance, this is a selection-on-the-test-set problem. Moreover, the 'best per modality' encoders used in the unified row of Table 2 are also identified from test AUROC/AUPRC in the same table. This biases every subsequent comparison in favor of the selected configuration and against the alternatives. Please either introduce a validation split for all hyperparameter and encoder selections and report only final test performance, or explicitly re-frame Tables 1 and 2 as exploratory post-selection results.
  3. [Section 3, Table 1, Abstract, Conclusion] The paper notes that no single latent configuration dominates all metrics, yet the abstract and conclusion state that latent compression is essential and outperforms pooling. In Table 1, k=16, l=2 gives AUROC 0.8771, higher than the selected k=8, l=4 AUROC 0.8751, while lower on AUPRC. The claimed superiority is therefore contingent on the a priori choice to prioritize AUPRC and on the seed-free comparison. Please state this contingency explicitly and temper the 'essential' language, or provide evidence that the qualitative outcome is invariant across reasonable metric choices and seeds.
minor comments (5)
  1. [Section 2.1-2.2] The symbol K_m is introduced as the maximum number of modality events per modality but its value is never specified. Please define it in the data-preprocessing description.
  2. [Abstract] Grammar: 'the best latent-compression configurations outperforms both...' should be 'configuration outperforms' or 'configurations outperform'.
  3. [Section 3, Table 1] The sentence 'best latent compression configuration improves all three' should be made precise: relative to no compression and mean pooling, and noting that the AUROC maximizer is actually k=16,l=2, not k=8,l=4.
  4. [Section 2.2/Table 2] The CXR encoder row is labeled 'BioMedCLIP ViT'; BioMedCLIP is a contrastive vision-language model, not a plain ViT. Clarify the exact architecture and projection used.
  5. [Section 2.1] MIMIC-IV v3.1, MIMIC-IV-ECG, MIMIC-CXR-JPG, and MIMIC-IV-Note are cited, but no dataset versions or access details are provided for the auxiliary datasets. Adding these details would improve reproducibility.

Circularity Check

0 steps flagged

No circularity: all central claims are empirical benchmark results, not derivations from their inputs.

full rationale

The paper's central claims are empirical comparisons on MIMIC-IV: latent compression versus pooling/no compression, encoder choice, and the effect of adding modalities on zero-shot ICU mortality. The model is trained with next-token cross-entropy on EHR tokens only; auxiliary modalities are conditioning inputs from frozen pretrained encoders. The zero-shot ICU mortality evaluation uses rollout sampling and compares against an EHR-only baseline, so the reported AUROC/AUPRC values are measured outcomes rather than quantities constructed from the fitted parameters. The selection of k=8, l=4 as the best latent-compression configuration is a model-selection choice made from Table 1, not a fitted parameter renamed as a prediction; it may risk selection-induced optimism, but it does not make the subsequent comparison circular. The authors' own caveat in Section 3 — 'the rollout-based evaluation is stochastic and can be noisy, and due to computational constraints we were unable to repeat all experiments across multiple random seeds' — is a statistical robustness limitation, not evidence that any claim reduces to its inputs. The self-citations (e.g., CSFM, Gu et al. 2026) are to externally pretrained encoder artifacts whose performance is measured here, not to an invoked uniqueness theorem or ansatz that is doing the derivational work. No equation in the paper equates an output to its input by construction; no fitted value is relabeled as a prediction. The negative result about modality non-additivity is an empirical finding consistent with prior external work, not a renaming of it. Overall circularity score: 0.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical or theoretical entities. Its load-bearing free choices are the compression hyperparameters k and l and the per-modality encoder selection, all chosen by performance on the test tables. The domain assumptions about timestamp reliability, transferability of frozen features, and rollout noise are the main unverified premises.

free parameters (3)
  • Number of latent tokens k = 8
    Swept over {8, 16} and selected because it gives the best AUPRC in Table 1. This is a design choice fitted to the evaluation metric, not derived.
  • Number of latent cross-attention layers l = 4
    Swept over {1, 2, 4} and selected jointly with k on test AUPRC. The authors acknowledge no single setting dominates all metrics.
  • Per-modality encoder selection = BioMedCLIP (CXR), CSFM (ECG), BioMedBERT (notes)
    The 'best per modality' encoders are chosen based on their within-modality performance in Table 2, then used in the unified three-modality model. This selection is itself part of the empirical claim.
axioms (4)
  • domain assumption MIMIC-IV tokenized via MEDS following ETHOS yields a faithful chronological EHR trajectory with reliable timestamps.
    Used to define the causal temporal masking in Section 2.2. If timestamps are imprecise or rounded, past-only masking may either leak future information or incorrectly discard usable modality events.
  • domain assumption Frozen pretrained encoders produce embeddings that transfer to this EHR task after learned latent compression.
    All encoders are frozen (Section 2.3); no fine-tuning or feature alignment is performed. The encoder-choice conclusions assume that the compressed features are meaningfully comparable across encoders.
  • domain assumption Rollout sampling with K=20 stochastic futures accurately estimates mortality probability well enough to rank configurations.
    Appendix A defines the evaluation; Section 3 explicitly notes the rollout evaluation is stochastic and noisy with no repeated seeds. Every comparison in Tables 1 and 2 relies on this stability assumption.
  • standard math Standard transformer and attention machinery behaves as described in the cited literature.
    The architecture combines GPT-2, Perceiver-style iterative attention, and Flamingo gating without formal verification; the paper assumes these published mechanisms compose correctly.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Autoregressive EHR Foundation Models with Multimodal Inputs." pith.science (2026). https://pith.science/paper/VT5UIUP4

@misc{pith2026260722264,
  author       = {Pith},
  title        = {Pith review of: Autoregressive EHR Foundation Models with Multimodal Inputs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VT5UIUP4}},
  note         = {Machine review of arXiv:2607.22264}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Autoregressive foundation models trained on tokenized electronic health records (EHRs) can support zero-shot clinical prediction, yet most operate on structured event codes alone, and do not incorporate multiple modalities in a principled way. We present a framework for conditioning such models on auxiliary clinical modalities, including ECG waveforms, chest X-ray images, and clinical notes, using modality-specific latent compression and gated cross-attention with temporal alignment. We investigate two key design choices: (1) how to compress long per-modality sequences (e.g., ECG time series) before they enter the multi-modal cross-attention. This feature may be essential to reduce compute overheads and may be beneficial for generalization; (2) how the choice of pretrained encoder for each modality impacts downstream performance. Through controlled ablations on MIMIC-IV, we show that the best latent-compression configurations outperforms both uncompressed cross-attention and mean pooling. Encoder choice has a clear within-modality effect, with stronger pretrained encoders consistently outperforming weaker alternatives. We further show that merely adding auxiliary modalities does not guarantee improvement on ICU mortality prediction over an EHR-only baseline. This implies that careful design of the fusion architecture and an appropriate evaluation in the clinical context are required.

Figures

Figures reproduced from arXiv: 2607.22264 by A. Aldo Faisal, Alfred John Balston, Jinpei Han, Joshua Placidi, Marek Rei, Yuxuan Liu.

Figure 1
Figure 1. Figure 1: Electronic health records are represented as sequences of clinical events spanning multiple modalities. 1. Introduction Recent work has shown that autoregressive transformers trained on tokenized electronic health records (EHRs) can learn reusable patient representations that support zero-shot clinical prediction (Renc et al., 2024; Waxler et al., 2025). These models linearize longitudinal EHR events (i.e.… view at source ↗
Figure 2
Figure 2. Figure 2: Multimodal EHR trajectory model. Auxiliary modalities (ECG, CXR, notes) are encoded, compressed into latent tokens, and fused into a GPT-2-style decoder via gated cross-attention. Temporal masking ensures each EHR token attends only to past modality events. pleural effusions. Clinical notes capture information that is not well represented in coded tables, including presenting symptoms, differential diagnos… 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

42 extracted references · 8 linked inside Pith

  1. [1]

    arXiv preprint arXiv:1607.06450 , year =

    Layer Normalization , author =. arXiv preprint arXiv:1607.06450 , year =

    Pith/arXiv arXiv

  2. [2]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    What Makes Training Multi-Modal Classification Networks Hard? , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  3. [3]

    Proceedings of the 39th International Conference on Machine Learning (ICML) , volume=

    Characterizing and Overcoming the Greedy Nature of Learning in Multi-modal Deep Neural Networks , author=. Proceedings of the 39th International Conference on Machine Learning (ICML) , volume=

  4. [4]

    Proceedings of the 39th International Conference on Machine Learning (ICML) , pages=

    Modality Competition: What Makes Joint Training of Multi-modal Network Fail in Deep Learning? (Provably) , author=. Proceedings of the 39th International Conference on Machine Learning (ICML) , pages=

  5. [5]

    Proceedings of the 38th International Conference on Machine Learning , series =

    Perceiver: General Perception with Iterative Attention , author =. Proceedings of the 38th International Conference on Machine Learning , series =. 2021 , editor =

    2021

  6. [6]

    Advances in Neural Information Processing Systems , editor=

    Flamingo: a Visual Language Model for Few-Shot Learning , author=. Advances in Neural Information Processing Systems , editor=. 2022 , url=

    2022

  7. [7]

    2020 , eprint =

    Yu Gu and Robert Tinn and Hao Cheng and Michael Lucas and Naoto Usuyama and Xiaodong Liu and Tristan Naumann and Jianfeng Gao and Hoifung Poon , title =. 2020 , eprint =

    2020

  8. [8]

    Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT) , year =

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding , author =. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT) , year =

    2019

  9. [9]

    OpenAI Technical Report , year =

    Language Models are Unsupervised Multitask Learners , author =. OpenAI Technical Report , year =

  10. [10]

    Workshop on Time Series Learning for Health (TS4H) at ICLR , year =

    Medical Event Data Standard (MEDS): Facilitating Machine Learning for Health , author =. Workshop on Time Series Learning for Health (TS4H) at ICLR , year =

  11. [11]

    npj Digital Medicine , year =

    Zero-shot health trajectory prediction using transformers , author =. npj Digital Medicine , year =. doi:10.1038/s41746-024-01235-0 , url =

    doi:10.1038/s41746-024-01235-0

  12. [12]

    2024 , publisher =

    MIMIC-IV (version 3.1) , author =. 2024 , publisher =. doi:10.13026/kpb9-mt58 , url =

    doi:10.13026/kpb9-mt58 2024

  13. [13]

    2023 , publisher =

    MIMIC-IV-ED (version 2.2) , author =. 2023 , publisher =. doi:10.13026/5ntk-km72 , url =

    doi:10.13026/5ntk-km72 2023

  14. [14]

    Circulation , year =

    PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals , author =. Circulation , year =

  15. [15]

    Advances in neural information processing systems , volume=

    wav2vec 2.0: A framework for self-supervised learning of speech representations , author=. Advances in neural information processing systems , volume=

  16. [16]

    JAMIA open , volume=

    Ecg-fm: An open electrocardiogram foundation model , author=. JAMIA open , volume=. 2025 , publisher=

    2025

  17. [17]

    Nature Machine Intelligence , volume=

    Cardiac health assessment across scenarios and devices using a multimodal foundation model pretrained on data from 1.7 million individuals , author=. Nature Machine Intelligence , volume=. 2026 , publisher=

    2026

  18. [18]

    Nature medicine , volume=

    The artificial intelligence clinician learns optimal treatment strategies for sepsis in intensive care , author=. Nature medicine , volume=. 2018 , publisher=

    2018

  19. [19]

    Journal of biomedical informatics , volume=

    Deep learning for temporal data representation in electronic health records: A systematic review of challenges and methodologies , author=. Journal of biomedical informatics , volume=. 2022 , publisher=

    2022

  20. [20]

    Heliyon , volume=

    Multimodal risk prediction with physiological signals, medical images and clinical notes , author=. Heliyon , volume=. 2024 , publisher=

    2024

  21. [21]

    Journal of the American Medical Informatics Association , volume=

    The first step is the hardest: Pitfalls of representing and tokenizing temporal data for large language models , author=. Journal of the American Medical Informatics Association , volume=. 2024 , publisher=

    2024

  22. [22]

    arXiv preprint arXiv:2508.09182 , year=

    MedPatch: Confidence-Guided Multi-Stage Fusion for Multimodal Clinical Data , author=. arXiv preprint arXiv:2508.09182 , year=

    Pith/arXiv arXiv

  23. [23]

    Authorea Preprints , year=

    Foundation models for biosignals: A survey , author=. Authorea Preprints , year=

  24. [24]

    Machine Learning for Health , pages=

    CEHR-BERT: Incorporating temporal information from structured EHR data to improve prediction tasks , author=. Machine Learning for Health , pages=. 2021 , organization=

    2021

  25. [25]

    Artificial Intelligence in Medicine , volume=

    Leveraging VQ-VAE tokenization for autoregressive modeling of medical time series , author=. Artificial Intelligence in Medicine , volume=. 2024 , publisher=

    2024

  26. [26]

    Briefings in bioinformatics , volume=

    Multimodal deep learning for biomedical data fusion: a review , author=. Briefings in bioinformatics , volume=. 2022 , publisher=

    2022

  27. [27]

    Scientific Reports , volume=

    Automated PD-L1 status prediction in lung cancer with multi-modal PET/CT fusion , author=. Scientific Reports , volume=. 2024 , publisher=

    2024

  28. [28]

    arXiv preprint arXiv:2509.09290 , year=

    Modality-Agnostic Input Channels Enable Segmentation of Brain lesions in Multimodal MRI with Sequences Unavailable During Training , author=. arXiv preprint arXiv:2509.09290 , year=

    Pith/arXiv arXiv

  29. [29]

    Machine Learning for Health (ML4H) , pages=

    Med-flamingo: a multimodal medical few-shot learner , author=. Machine Learning for Health (ML4H) , pages=. 2023 , organization=

    2023

  30. [30]

    Nejm Ai , volume=

    Towards generalist biomedical AI , author=. Nejm Ai , volume=. 2024 , publisher=

    2024

  31. [31]

    arXiv preprint arXiv:2510.23639 , year=

    Integrating Genomics into Multimodal EHR Foundation Models , author=. arXiv preprint arXiv:2510.23639 , year=

    arXiv

  32. [32]

    arXiv preprint arXiv:2506.07044 , year=

    Lingshu: A Generalist Foundation Model for Unified Multimodal Medical Understanding and Reasoning , author=. arXiv preprint arXiv:2506.07044 , year=

    Pith/arXiv arXiv

  33. [33]

    arXiv preprint arXiv:2306.05109 , year=

    Yet another icu benchmark: A flexible multi-center framework for clinical ml , author=. arXiv preprint arXiv:2306.05109 , year=

    Pith/arXiv arXiv

  34. [34]

    Scientific reports , volume=

    Deep patient: an unsupervised representation to predict the future of patients from the electronic health records , author=. Scientific reports , volume=. 2016 , publisher=

    2016

  35. [35]

    arXiv preprint arXiv:2508.12104 , year=

    Generative medical event models improve with scale , author=. arXiv preprint arXiv:2508.12104 , year=

    arXiv

  36. [36]

    arXiv preprint arXiv:2505.22964 , year=

    Exploring Scaling Laws for EHR Foundation Models , author=. arXiv preprint arXiv:2505.22964 , year=

    Pith/arXiv arXiv

  37. [37]

    Advances in Neural Information Processing Systems , volume=

    Ehrshot: An ehr benchmark for few-shot evaluation of foundation models , author=. Advances in Neural Information Processing Systems , volume=

  38. [38]

    Type: dataset , volume=

    Mimic-iv-ecg: Diagnostic electrocardiogram matched subset , author=. Type: dataset , volume=

  39. [39]

    arXiv preprint arXiv:1901.07042 , year=

    MIMIC-CXR-JPG, a large publicly available database of labeled chest radiographs , author=. arXiv preprint arXiv:1901.07042 , year=

    Pith/arXiv arXiv 1901

  40. [40]

    PhysioNet , author=

    MIMIC-IV-Note: Deidentified free-text clinical notes. PhysioNet , author=

  41. [41]

    arXiv preprint arXiv:2303.00915 , year=

    BiomedCLIP: A Multimodal Biomedical Foundation Model Pretrained from Fifteen Million Scientific Image-Text Pairs , author=. arXiv preprint arXiv:2303.00915 , year=

    Pith/arXiv arXiv

  42. [42]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Masked Autoencoders Are Scalable Vision Learners , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.