Pith. sign in

REVIEW 3 major objections 3 minor 1 cited by

This paper argues that ICU risk prediction from vital signs alone can be improved by training with ICD-ontology-aware contrastive learning and distilling clinical-note knowledge into a vitals-only student.

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

2026-08-02 15:56 UTC pith:6G62VCYO

load-bearing objection A clean, well-specified two-stage framework whose central claim about label efficiency and SOTA is not yet firmly supported: the improvements over baselines sit inside overlapping confidence intervals, and the stage-1 pretraining has a supervision confound the paper does not fully control. the 3 major comments →

arxiv 2604.16878 v3 pith:6G62VCYO submitted 2026-04-18 cs.LG

OC-Distill: Ontology-aware Contrastive Learning with Cross-Modal Distillation for ICU Risk Prediction

Zhongyuan Liang , Junhyung Jo , Hyang-Jung Lee , Sang Kyu Kim , Irene Y. Chen This is my paper
classification cs.LG
keywords ICU risk predictioncontrastive learningknowledge distillationICD ontologyclinical time serieslabel efficiencymultimodal learningMIMIC
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 paper argues that ICU risk prediction from vital-sign streams is held back by two design choices: contrastive pretraining that treats every patient as an equally strong negative, and training that ignores clinical notes. It proposes OC-Distill, a two-stage framework whose first stage uses the ICD diagnosis hierarchy to compute patient similarity and down-weights clinically similar negative pairs during contrastive pretraining, and whose second stage distills knowledge from a teacher trained on both clinical notes and vital signs into a student that sees only vital signs. On MIMIC benchmarks for in-hospital mortality and length-of-stay prediction, the paper reports improved label efficiency and state-of-the-art performance among methods using only vital signs at inference. A sympathetic reader would take the central claim to be that clinical domain structure can fruitfully shape contrastive objectives, and that complementary modalities can be used at training time without burdening deployment.

Core claim

OC-Distill's central claim is that replacing the uniform negative weighting of standard contrastive learning with weights derived from the ICD code hierarchy produces clinically organized, label-efficient representations, and that distilling soft targets from a notes-plus-vitals teacher into a vitals-only student adds further downstream gains. Patient similarity is computed as a symmetric best-match Jaccard overlap of root-excluded ICD code paths; a monotone decreasing transformation Phi maps this similarity to contrastive weights w_ij = Phi(Sim(A_i, A_j)), which multiply the negative-pair terms in the NT-Xent denominator. Stage two trains a teacher on fused note and vital representations, t

What carries the argument

The carrying object is the ontology-aware patient similarity score and its insertion into a weighted contrastive loss. The ICD hierarchy is treated as a rooted tree with 14,870 nodes; two ICD codes a and b get similarity s(a,b) as the Jaccard overlap of their root-excluded ancestor paths, and two patients get Sim(A,B) as the symmetric average of best matches across their diagnosis sets. The default transformation is the power function Phi(s)=(1-s)^gamma with gamma=5, so highly similar diagnosis pairs are strongly down-weighted in the contrastive denominator while dissimilar pairs are repelled almost as in SimCLR. This reweighting is what turns an augmentation-based unsupervised objective int

Load-bearing premise

The load-bearing premise is that ICD path-based similarity is a valid and unbiased measure of clinical relatedness, so down-weighting diagnosis-similar negatives helps the model learn physiology rather than merely copying code structure that correlates with the target labels.

What would settle it

Retrain Stage 1 with the same contrastive objective but replace the ICD hierarchy with a random depth-preserving permutation of the diagnosis tree, so that similarity scores are computed on shuffled labels. If the linear-probe AUROC gains over SimCLR remain as large as with the true ICD tree, the claimed mechanism is unsupported; if the gains vanish, the hierarchy itself is doing the work.

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

Share X LinkedIn Reddit HN

If this is right

  • If the pretraining gains are real, ICU risk models can reach strong performance with substantially less labeled data, since linear-probe gains are largest at 1-10% labeled training data.
  • Because the student deploys on vital signs alone, the approach avoids the practical failure of multimodal fusion at inference while still capturing note-derived signal in the trained representation.
  • The reweighting scheme is loss-agnostic, so the same ICD-derived similarity weighting can be inserted into other contrastive objectives, including supervised contrastive learning.
  • The teacher-student distillation is modality-agnostic, meaning other complementary modalities beyond clinical notes could serve as the teacher.
  • The authors' MIMIC-IV validation of the vitals-only student suggests the student transfers directly to a dataset that lacks longitudinal nursing notes.

Where Pith is reading between the lines

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

  • The label-efficiency result suggests the ICD ontology acts as a form of weak supervision; a natural testable extension is applying the same hierarchical weighting to other diagnosis taxonomies on non-ICU clinical time series.
  • The paper leaves open which clinical concepts are actually transferred during distillation; an interpretability study of the student's gradient or attention patterns could identify whether the gains come from concrete concepts in the notes or from a generic smoothing of the teacher's predictions.
  • A clean leakage test would be to compute patient similarity weights from ICD codes that cannot contain the outcome label, for example, codes from a prior admission or codes stripped of target-correlated categories; if gains persist under that restriction, the mechanism is less suspicious.
  • The paper's own Section 6 limitation statement concedes that ICD coding can be noisy, missing, or upcoded, so the most decisive follow-up should evaluate whether gains persist when the similarity function is replaced by a diagnosis-agnostic hierarchy built from physiology variables.

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 / 3 minor

Summary. OC-Distill proposes a two-stage framework for ICU risk prediction from vital signs only at deployment. Stage 1 pretrains a vitals encoder with a contrastive objective in which negative pair weights are derived from ICD ontology-based patient similarity (Jaccard overlap of root-excluded ICD paths). Stage 2 fine-tunes the encoder via knowledge distillation from a teacher trained on clinical notes plus vitals. The paper reports improved label efficiency and state-of-the-art performance among vitals-only-inference methods on MIMIC-III (in-hospital mortality and length-of-stay), with additional MIMIC-IV validation. The empirical claims rest on Tables 2–5 and Figure 6, and the paper includes a code repository and bootstrap confidence intervals.

Significance. If the ontology-aware pretraining mechanism is what drives the gains, the idea is a useful and generalizable contribution: it combines a clinically meaningful hierarchy with a deployment-friendly distillation stage. The paper is clearly written, provides code, and reports consistency across label fractions and horizons. However, the central claims are not cleanly established. The comparison with SimCLR conflates the ontology structure with additional, target-correlated ICD supervision; the pretraining split and ICD-code timing are not specified; and the headline SOTA differences in Table 4 have overlapping bootstrap CIs with no significance testing. These issues are fixable and do not undermine the plausibility of the framework, but they need to be addressed before the claims can be accepted.

major comments (3)
  1. [§4.2 and §3.1 (Eq. 3, w_ij = Φ(Sim(A_i,A_j)))] The pretraining protocol is under-specified and the main comparison is confounded. It is never stated whether Stage-1 pretraining was restricted to the 16,861 training stays or whether ICD codes were taken only from the first T hours. If test-set patients' ICD codes enter the pretraining batch, the linear-evaluation and fine-tuning results in Tables 2–4 are leakage-contaminated. Even if pretraining is split-clean, the ICD-derived weights provide diagnosis-based supervision that SimCLR does not receive; since ICD codes are strongly predictive of mortality and LOS, the comparison does not isolate the effect of the ontology hierarchy. The Discussion acknowledges ICD noise/upcoding but does not address this confound. Please add a control that breaks the link between the weight values and the ICD codes while preserving the weight distribution (e.g., permuted or randomized weights), and clarif
  2. [§5.2.1, Table 4] The claim that OC-Distill achieves state-of-the-art performance among vitals-only-inference methods is not statistically supported. The reported AUROC/AUPRC differences over the Transformer baseline have overlapping bootstrap CIs (e.g., mortality 100% AUROC 0.793 [0.769–0.816] vs 0.788 [0.762–0.811]; LOS 100% AUROC 0.682 [0.670–0.695] vs 0.676 [0.664–0.688]), and no p-values or difference-in-CI tests are provided. Because the bootstrap CIs are computed from test predictions rather than over training runs, they also do not reflect seed variability; this matters especially at 50% labels. Please provide significance tests (e.g., DeLong or bootstrap tests for AUROC differences) and, if feasible, multiple seeds.
  3. [§5.1.2, Tables 2–3] The label-efficiency claim is based primarily on linear evaluation at 1%, 5%, and 10% labeled data. While the trends are directionally consistent, the reported CIs overlap for most comparisons (e.g., 1% mortality AUROC: Ontology-Aware 0.673 [0.643–0.702] vs SimCLR 0.635 [0.603–0.666]; 5% AUPRC: 0.319 [0.271–0.372] vs 0.279 [0.235–0.328]). No multiple-comparison correction is applied across tasks, metrics, and label fractions. The claim of consistent gains would be strengthened by reporting bootstrap difference distributions or paired tests, and by clarifying whether the bootstrap CIs are over patients or training runs.
minor comments (3)
  1. [Table 7 (Appendix B)] In the 10% row, the In-Hospital Mortality AUPRC is identical (0.313 with CI 0.267–0.369) for SimCLR, Flat Diagnosis CL, and Ontology-Aware CL. This is likely a copy-paste error and should be corrected, since it affects the appendix's cross-horizon robustness story.
  2. [§5.1.3] The text says 'sharpened settings (e.g., γ=5, τ=0.3)' but the threshold transform is denoted δ, not τ, in §3.1.2. Please fix the notation to avoid confusion with the temperature τ used elsewhere.
  3. [Figure 5] The x-axis labels 'Power =1' through 'Power =5' omit the γ symbol, and the right-pair axes would benefit from clearer labels distinguishing the transformation family from its hyperparameter.

Circularity Check

0 steps flagged

No significant circularity: central label-efficiency/SOTA claims are externally benchmarked; ICD-derived similarity is an external input, not a restatement of the downstream labels.

full rationale

The claimed derivation chain is not circular. Stage-1 pretraining defines patient similarity Sim(A,B) externally from the ICD hierarchy (§3.1.1) and inserts weights w_ij = Φ(Sim(A_i,A_j)) into an NT-Xent denominator (§3.1.2). The downstream targets (in-hospital mortality, LOS) never appear in the pretraining objective, so the prediction is not equivalent by construction to the similarity input. The baselines include a flat-diagnosis reweighting under the same loss, which isolates the ontology contribution, and the final student is evaluated on held-out MIMIC test splits and on MIMIC-IV without retraining (Appendix D), providing external benchmark support. The self-citations (Liang et al. 2025a,b,c; Ginsberg et al. 2022; Ma et al. 2026) are related-work or interpretability references and are not load-bearing. The only circularity-adjacent element is Section 5.1.1, where the observation that embedding neighbors show higher diagnosis similarity is in part a restatement of the weighted objective that explicitly down-weights high-Sim negatives; this is a self-consistency check, not a central prediction, and does not by itself support the headline label-efficiency claim. A residual correctness risk, not circularity, is that the label-efficiency comparison gives the proposed method access to ICD-code supervision while denying it to SimCLR, so part of the gain may reflect target-correlated auxiliary diagnosis information; the paper acknowledges ICD noise/upcoding but does not control for this confound.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No new physical or ontological entities are postulated. The framework depends on two domain assumptions: ICD codes are a trustworthy similarity signal, and clinical notes can be distilled into a vitals-only encoder. Hyperparameters γ, λ_distill, T, and p are tuned on validation and are genuine free parameters in the reported performance.

free parameters (5)
  • γ (power-transform exponent) = 5
    Selected via validation ablation (Figure 5); higher γ gives sharper down-weighting and better linear-probe AUROC.
  • λ_distill distillation weight = grid {1,5,10}, best varies by task
    Tuned on validation AUROC; Figure 6 shows performance is sensitive to this weight.
  • Distillation temperature T = grid {1,2,5}
    Controls softness of teacher targets; chosen by validation.
  • LLM summary probability p = grid {0,0.5,1.0}; 0.5 best in Appendix A
    Probability of using GPT-4o summary versus raw notes in teacher training.
  • Similarity transform family Φ = Power transform (vs exponential, threshold)
    Design choice among parameterized transforms; ablation in Figure 5.
axioms (5)
  • domain assumption The ICD ontology from BioPortal is accurate and complete, and Jaccard overlap over root-excluded paths is a valid clinical similarity measure.
    This similarity weights every contrastive negative in Section 3.1.1. If coding is noisy, missing, or biased by upcoding, the representation signal is corrupted (acknowledged in Discussion).
  • domain assumption Clinical notes carry information about mortality and length of stay that is not already available in vital signs and that can be transferred by distillation.
    Stage 2 gains in Section 5.2 depend on the teacher providing soft targets that add signal beyond hard labels and vitals.
  • domain assumption The MIMIC-III train/test split is patient-disjoint and no test-set information, including ICD codes, is used during contrastive pretraining.
    Contrastive batches require ICD codes to compute similarity weights, but the paper never states whether test patients are excluded during pretraining.
  • domain assumption GPT-4o zero-shot summaries are faithful to the original notes and do not introduce hallucinated or temporally leaking content.
    Teacher training randomly includes generated summaries (Section 3.2.1); unfaithful summaries could bias soft targets.
  • domain assumption A BERT-style vitals encoder pretrained on 48-hour windows can be applied to 72- and 96-hour inputs without special positional-encoding handling.
    Cross-time results in Figure 4 use longer inputs than the pretraining horizon, but the paper does not describe how positional embeddings are adapted.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of OC-Distill: Ontology-aware Contrastive Learning with Cross-Modal Distillation for ICU Risk Prediction." pith.science (2026). https://pith.science/paper/6G62VCYO

@misc{pith2026260416878,
  author       = {Pith},
  title        = {Pith review of: OC-Distill: Ontology-aware Contrastive Learning with Cross-Modal Distillation for ICU Risk Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6G62VCYO}},
  note         = {Machine review of arXiv:2604.16878}
}
Share X LinkedIn Reddit HN
read the original abstract

Early prediction of severe clinical deterioration and remaining length of stay can enable timely intervention and better resource allocation in high-acuity settings such as the ICU. This has driven the development of machine learning models that leverage continuous streams of vital signs and other physiological signals for real-time risk prediction. Despite their promise, existing methods have important limitations. Contrastive pretraining treats all patients as equally strong negatives, failing to capture clinically meaningful similarity between patients with related diagnoses. Meanwhile, downstream fine-tuning typically ignores complementary modalities such as clinical notes, which provide rich contextual information unavailable in physiological signals alone. To address these challenges, we propose OC-Distill, a two-stage framework that leverages multimodal supervision during training while requiring only vital signs at inference. In the first stage, we introduce an ontology-aware contrastive objective that exploits the ICD hierarchy to quantify patient similarity and learn clinically grounded representations. In the second stage, we fine-tune the pretrained encoder via cross-modal knowledge distillation, transferring complementary information from clinical notes into the model. Across multiple ICU prediction tasks on MIMIC, OC-Distill demonstrates improved label efficiency and achieves state-of-the-art performance among methods that use only vital signs at inference.

Figures

Figures reproduced from arXiv: 2604.16878 by Hyang-Jung Lee, Irene Y. Chen, Junhyung Jo, Sang Kyu Kim, Zhongyuan Liang.

Figure 1
Figure 1. Figure 1: Overview of our two-stage framework. (a) We compute patient similarity from the ICD [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of contrastive weights for patient pairs. Flat diagnosis matching yields zero [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of contrastive weights for patient pairs. Flat diagnosis matching yields zero similarity (weight = 1) for half of pairs. Ontology-aware similarity produces a smoother distribution with gradations, while a power transformation sharpens the weights. Percentages indicate the fraction with weight < 1. Embedding-space similarity analysis. We next evaluate whether the pretrained en￾coder produces em… view at source ↗
Figure 3
Figure 3. Figure 3: Diagnosis similarity distributions for em [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 3
Figure 3. Figure 3 [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cross-time generalization of linear-probe [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 4
Figure 4. Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Linear probing performance under different weight transformations, evaluated using 5% [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 5
Figure 5. Figure 5: Linear probing performance under different weight transformations, evaluated using 5% labeled training data. Left pair: effect of the power-transform exponent γ. Right pair: comparison across weight functions. Similar trends are observed across other labeled-data fractions. Full fine-tuning. We lastly evaluate pretraining performance under full fine-tuning. In this setting, we initialize the model from pre… view at source ↗
Figure 6
Figure 6. Figure 6: Effect of λdistill on model performance, evaluated using 50% labeled training data. Across tasks, distillation consistently improves over the no-distillation baseline. Similar trends are observed across other labeled-data fractions. vised labels alone. Performance is strongest at moderate-to-large λdistill, suggesting that distillation is most effective when soft targets contribute meaningfully to training… view at source ↗
Figure 6
Figure 6. Figure 6: Effect of λdistill on model performance, evaluated using 50% labeled training data. Across tasks, distillation consistently improves over the no-distillation baseline. Similar trends are observed across other labeled-data fractions. 50% Labels 100% Labels Task Initialization AUROC AUPRC AUROC AUPRC In-Hospital Mortality Scratch 0.766 0.334 0.791 0.379 Pretrained 0.776 0.361 0.793 0.385 Length of Stay Scrat… view at source ↗

Review history (3 revisions) →

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. MMOOC: A Comprehensive Benchmark for Out-of-Context Evaluation in Multimodal Large Language Models

    cs.CV 2026-07 conditional novelty 6.0

    MMOOC, a 41K-pair benchmark, shows current multimodal LLMs struggle to both refuse truly out-of-context questions and correctly answer questions that remain answerable despite misleading or shifted context.

Reference graph

Works this paper leans on

20 extracted references · 13 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Publicly available clinical bert embeddings.arXiv preprint arXiv:1904.03323,

    Emily Alsentzer, John R Murphy, Willie Boag, Wei-Hung Weng, Di Jin, Tristan Naumann, and Matthew McDermott. Publicly available clinical bert embeddings.arXiv preprint arXiv:1904.03323,

    Pith/arXiv arXiv 1904

  2. [5]

    As shown in Table 11, the conclusion is stable across all choices ofK: nearest-neighbor pairs consistently exhibit higher diagnosis similarity than randomly sampled pairs, and all differences remain statistically significant. KKNN Mean Random Mean Effect SizerMann–Whitneyp 1 0.251 0.200 0.214<0.0001 3 0.248 0.200 0.205<0.0001 5 0.246 0.199 0.198<0.0001 Ta...

    2023

  3. [6]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531,

    Pith/arXiv arXiv

  4. [10]

    Reco-kd: Region-and context-aware knowledge distillation for efficient 3d medical image segmentation.arXiv preprint arXiv:2601.08301,

    Qizhen Lan, Yu-Chun Hsu, Nida Saddaf Khan, and Xiaoqian Jiang. Reco-kd: Region-and context-aware knowledge distillation for efficient 3d medical image segmentation.arXiv preprint arXiv:2601.08301,

    arXiv

  5. [13]

    A multimodal transformer: Fusing clinical notes with structured ehr data for interpretable in-hospital mortality prediction

    Weimin Lyu, Xinyu Dong, Rachel Wong, Songzhu Zheng, Kayley Abell-Hart, Fusheng Wang, and Chao Chen. A multimodal transformer: Fusing clinical notes with structured ehr data for interpretable in-hospital mortality prediction. InAMIA Annual Symposium Proceedings, volume 2022, page 719,

    2022

  6. [14]

    Do sparse autoencoders identify reasoning features in language models?arXiv preprint arXiv:2601.05679,

    George Ma, Zhongyuan Liang, Irene Y Chen, and Somayeh Sojoudi. Do sparse autoencoders identify reasoning features in language models?arXiv preprint arXiv:2601.05679,

    Pith/arXiv arXiv

  7. [16]

    Integrating physiological time series and clinical notes with deep learning for improved icu mortality prediction.arXiv preprint arXiv:2003.11059,

    19 Ontology-aware CL with Cross-Modal Distillation for ICU Risk Prediction Satya Narayan Shukla and Benjamin M Marlin. Integrating physiological time series and clinical notes with deep learning for improved icu mortality prediction.arXiv preprint arXiv:2003.11059,

    Pith/arXiv arXiv 2003

  8. [19]

    3d cgan based cross-modality mr image synthesis for brain tumor segmentation

    Biting Yu, Luping Zhou, Lei Wang, Jurgen Fripp, and Pierrick Bourgeat. 3d cgan based cross-modality mr image synthesis for brain tumor segmentation. In2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018), pages 626–630. IEEE,

    2018

  9. [1994]

    A learning based hypothesis test for harmful covariate shift.arXiv preprint arXiv:2212.02742,

    Tom Ginsberg, Zhongyuan Liang, and Rahul G Krishnan. A learning based hypothesis test for harmful covariate shift.arXiv preprint arXiv:2212.02742,

    Pith/arXiv arXiv

  10. [2009]

    Multi-view contrastive learning for robust domain adaptation in medical time series analysis.arXiv preprint arXiv:2506.22393,

    YongKyung Oh and Alex Bui. Multi-view contrastive learning for robust domain adaptation in medical time series analysis.arXiv preprint arXiv:2506.22393,

    arXiv

  11. [2015]

    Equitable electronic health record prediction with fame: Fairness-aware multimodal embedding.arXiv preprint arXiv:2506.13104,

    Nikkie Hooman, Zhongjie Wu, Eric C Larson, and Mehak Gupta. Equitable electronic health record prediction with fame: Fairness-aware multimodal embedding.arXiv preprint arXiv:2506.13104,

    arXiv

  12. [2018]

    Clinical intervention prediction and understanding using deep net- works.arXiv preprint arXiv:1705.08498,

    Harini Suresh, Nathan Hunt, Alistair Johnson, Leo Anthony Celi, Peter Szolovits, and Marzyeh Ghassemi. Clinical intervention prediction and understanding using deep net- works.arXiv preprint arXiv:1705.08498,

    Pith/arXiv arXiv

  13. [2019]

    Enhancing in-hospital mor- tality prediction using multi-representational learning with llm-generated expert sum- maries.arXiv preprint arXiv:2411.16818,

    Harshavardhan Battula, Jiacheng Liu, and Jaideep Srivastava. Enhancing in-hospital mor- tality prediction using multi-representational learning with llm-generated expert sum- maries.arXiv preprint arXiv:2411.16818,

    Pith/arXiv arXiv

  14. [2020]

    Integrating physiological time series and clinical notes with transformer for early prediction of sepsis.arXiv preprint arXiv:2203.14469,

    Yuqing Wang, Yun Zhao, Rachael Callcut, and Linda Petzold. Integrating physiological time series and clinical notes with transformer for early prediction of sepsis.arXiv preprint arXiv:2203.14469,

    Pith/arXiv arXiv

  15. [2021]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186,

    2019

  16. [2022]

    Local mdi+: Local feature importances for tree-based models.arXiv preprint arXiv:2506.08928, 2025a

    18 Ontology-aware CL with Cross-Modal Distillation for ICU Risk Prediction Zhongyuan Liang, Zachary T Rewolinski, Abhineet Agarwal, Tiffany M Tang, and Bin Yu. Local mdi+: Local feature importances for tree-based models.arXiv preprint arXiv:2506.08928, 2025a. Zhongyuan Liang, Arvind Suresh, and Irene Y. Chen. Treatment non-adherence bias in clinical machi...

    Pith/arXiv arXiv

  17. [2023]

    Medpatch: Confidence-guided multi-stage fusion for multimodal clinical data.arXiv preprint arXiv:2508.09182,

    17 Ontology-aware CL with Cross-Modal Distillation for ICU Risk Prediction Baraa Al Jorf and Farah Shamout. Medpatch: Confidence-guided multi-stage fusion for multimodal clinical data.arXiv preprint arXiv:2508.09182,

    Pith/arXiv arXiv

  18. [2024]

    Maven-t: Multi-agent environment-aware enhanced neural trajectory predictor with reinforcement learning.arXiv preprint arXiv:2604.10169,

    Wenchang Duan. Maven-t: Multi-agent environment-aware enhanced neural trajectory predictor with reinforcement learning.arXiv preprint arXiv:2604.10169,

    Pith/arXiv arXiv

  19. [2025]

    Bridging electronic health records and clinical texts: Contrastive learning for enhanced clinical tasks.arXiv preprint arXiv:2505.17643,

    Sara Ketabi and Dhanesh Ramachandram. Bridging electronic health records and clinical texts: Contrastive learning for enhanced clinical tasks.arXiv preprint arXiv:2505.17643,

    Pith/arXiv arXiv

  20. [2026]

    Retrieval-augmented dynamic prompt tuning for incomplete multimodal learning

    Jian Lang, Zhangtao Cheng, Ting Zhong, and Fan Zhou. Retrieval-augmented dynamic prompt tuning for incomplete multimodal learning. InProceedings of the AAAI Confer- ence on Artificial Intelligence, volume 39, pages 18035–18043, 2025a. Jian Lang, Rongpei Hong, Zhangtao Cheng, Ting Zhong, Yong Wang, and Fan Zhou. Re- deeming modality information loss: Retri...

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.