CardioCoT: Hierarchical Reasoning for Multimodal Survival Analysis

Haoyang Su; Jinyi Xiang; Lian-Ming Wu; Shaohao Rui; Xiaosong Wang

REVIEW 4 major objections 5 minor 1 cited by

CardioCoT: Hierarchical Reasoning for Multimodal Survival Analysis

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read CardioCoT combines cardiac MRI with LLM-generated hierarchical clinical reasoning to predict MACE recurrence risk, reporting a C-index of 0.8342—a 7.53% gain over state-of-the-art multimodal survival baselines.

desk verdict A useful interpretability framework whose headline c-index gain is probably inflated by outcome leakage in the MACE follow-up reasoning input. read the letter →

arxiv 2505.19195 v1 pith:P4V3TIX6 submitted 2025-05-25 cs.AI cs.CV

Shaohao Rui , Haoyang Su , Jinyi Xiang , Lian-Ming Wu , Xiaosong Wang This is my paper

classification cs.AIcs.CV

keywords survivalanalysisMACErecurrenceriskhierarchicalreasoningchain-of-thoughtmultimodalfusioncardiacMRIself-refinementinterpretability

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces CardioCoT, a two-stage survival analysis pipeline that predicts recurrence of major adverse cardiovascular events (MACE) after acute myocardial infarction. Its claim is that adding hierarchical, LLM-generated clinical reasoning—radiological diagnosis, cardiac complications, and MACE follow-up—to cardiac MRI features improves risk stratification and provides interpretable intermediate evidence. On an in-house dataset of 871 patients with 104 recurrence events, CardioCoT reports a C-index of 0.8342, a 7.53% gain over the strongest multimodal baseline. The authors argue this makes the model clinically usable because the reasoning steps can be inspected, not just the risk score.

What carries the argument

The load-bearing mechanism is the evidence-augmented self-refinement loop that produces hierarchical reasoning trajectories. A Thinker model produces initial reasoning and conclusion; an Oracle verifies consistency with evidence; mismatches trigger Review or Rethinking prompts, and after a maximum of two failed attempts an Inference Correction prompt reveals the correct diagnosis and forces a corrected chain. The resulting diagnosis, complication, and MACE follow-up conclusions are formatted as text, encoded with a long-context T5 text encoder, and fused with MRI image features by gated attention pooling; the fused representation is trained with the same survival loss used by the baseline.

What would settle it

Run a version of the CardioCoT second stage where the MACE follow-up reasoning text is removed or replaced with a length-matched placeholder at test time and re-measure the C-index; a large drop from 0.8342 toward the 0.7589 baseline would show that the gain is carried by the label-bearing follow-up text rather than by the reasoning itself.

Watch

Extended reading notes

Core claim

The central discovery is that intermediate reasoning trajectories can act as a bridge between unstructured clinical text and a survival model, outperforming end-to-end prompting or direct embedding. In the first stage, a Thinker model is prompted to produce diagnosis, complications, and MACE follow-up conclusions, with an Oracle checking each against evidence and triggering review, rethinking, or inference correction until a consistent chain is formed. These trajectories are used to fine-tune several open-source LLMs and VLMs. In the second stage, embeddings of the radiological findings and the three reasoning levels are fused with 2D DenseNet121 features from MRI via attention pooling and optimized with a survival loss. The paper reports that this beats state-of-the-art multimodal survival methods and that all three reasoning levels contribute in ablation.

Load-bearing premise

The reported gain rests on the assumption that the MACE follow-up text fed to the survival model is produced without access to the true recurrence label; otherwise the 0.8342 C-index is inflated by label leakage.

Editorial extensions

If this is right

When a 70B domain-tuned LLM supplies the reasoning text, the survival model reaches a C-index of 0.8342, while smaller fine-tuned models still beat existing multimodal baselines.
Ablation shows each hierarchical level—diagnosis, complications, and MACE follow-up—adds predictive value, with the full combination giving the best C-index and survival AUC.
Attention analysis indicates the reasoning modules carry substantial, sample-variable weight, meaning the textual reasoning encodes information not directly visible in the MRI.
The same two-stage recipe should transfer to other composite clinical endpoints where structured follow-up and imaging are both available.
Because the reasoning text is produced in the first stage and only encoded in the second, the survival model itself remains small and trainable on limited data.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper leaves open whether the MACE follow-up text is generated with the true recurrence label visible; if it is, the C-index gain is inflated by outcome leakage.
If the trajectory-generation stage is truly label-free, the framework implies a practical division of labor: a large model writes inspectable reasoning once, and a small second-stage model consumes it, avoiding end-to-end LLM survival modeling.
A testable extension is to redact or shuffle the follow-up segment of the reasoning at inference and measure the C-index drop; a large drop would confirm that segment carries the predictive signal.
The hierarchical reasoning chain also gives clinicians a concrete artifact to audit when a predicted risk disagrees with their judgment, which could support error analysis and calibration in practice.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

CardioCoT is a well-written two-stage system for MACE recurrence prediction: GPT-4o generates hierarchical reasoning over radiological diagnosis, complications, and MACE follow-up, with a self-refinement loop; then a survival model fuses MRI embeddings with T5 embeddings of that reasoning text. The clinical motivation is real, and the interpretability side is a genuine plus—the qualitative examples show how the reasoning would look to a clinician. The authors also did honest legwork: ablations for each reasoning module, attention-score analysis, and fine-tuning across several LLM/VLMs with LoRA. The code is public.

I checked the stress-test concern and it holds up. The MACE follow-up reasoning input, I_M, is generated from "the recurrence status of MACE over a specified time period"—which is, in survival analysis, the event label. The self-refinement protocol explicitly includes an Inference Correction branch that says "The correct answer is {...}" when the reasoning doesn't match the evidence. If that evidence includes the true recurrence status, then I_M encodes the outcome, and feeding it into the Stage 2 survival model is label leakage. The paper never states that the outcome is withheld when I_M is generated at test time, and the gain from adding the M module in Table 2 (+0.0127 over D+C) is exactly what leakage would look like. There is also an operational problem: at the time of prediction, future MACE status is unknown, so the pipeline as described could not run prospectively.

Even setting leakage aside, the evidence base is thin: one in-house dataset, 871 patients and 104 events, a single 6:2:2 split, no confidence intervals, no external validation, and no non-reasoning LLM-text baseline that controls for the extra text modality. The 7.53% improvement over PORPOISE/MMDB is therefore not defensible as stated.

I don't think this is a fundamentally bad idea. The two-stage design is sensible, and the hierarchical reasoning with self-refinement is a reasonable way to get interpretable clinical predictions. The fix is clear: define the prediction-time information set, generate I_M without the true outcome, use temporal or at least repeated splits, report CIs, and add a control LLM feature baseline. If they do that, the result may survive; the current version doesn't support the headline.

Who is this for? Researchers working on LLM-assisted clinical risk prediction, and anyone who wants a clear example of how easy label leakage is in generated reasoning features. I'd send it to a serious referee—the leakage question is exactly what peer review should adjudicate—but I would not cite the reported numbers.

Referee Report

4 major / 5 minor

Summary. The paper proposes CardioCoT, a two-stage framework for MACE recurrence risk prediction after acute myocardial infarction. In Stage 1, GPT-4o generates hierarchical reasoning trajectories over radiological diagnosis, complications, and MACE follow-up, using an evidence-augmented self-refinement loop; these trajectories are used to fine-tune open-source LLMs/VLMs. In Stage 2, the resulting textual reasoning features (IF, ID, IC, IM) are encoded with T5, fused with DenseNet121 MRI features via attention pooling, and used for survival prediction. On an in-house dataset of 871 patients with 104 MACE events, the method reports a C-index of 0.8342 versus 0.7589 for the MMDB baseline, a 7.53% improvement, along with Kaplan-Meier stratification and attention-based interpretability analyses.

Significance. If the reported gain were genuine, CardioCoT would be a useful contribution: it explicitly targets interpretability in survival analysis, combines multimodal imaging with language-model reasoning, and provides a modular ablation study. The authors also release a code repository, which aids reproducibility. However, the central claim is not defensible as stated because the MACE follow-up reasoning input appears to encode the outcome being predicted, and no uncertainty quantification or external validation supports the 7.53% improvement. The interpretability results are also weakened by the same leakage concern.

major comments (4)

[Section 2.1, Section 2.3, Eq. (1)] The paper states that "MACE follow-up tracks the recurrence status of MACE over a specified time period" and then uses the resulting reasoning trajectory IM as an input to the Stage 2 survival model. In a prospective survival analysis, the recurrence status over the follow-up period is exactly the event label the model is supposed to predict. The manuscript never states that the true recurrence label is withheld when IM is generated, so the reported C-index gain (Table 1, 0.8342 vs 0.7589) can be explained by label leakage rather than by genuine reasoning. This is a load-bearing flaw: as described, the method cannot be deployed for prospective risk prediction because future MACE status is unknown at prediction time.
[Eq. (1), Section 2.1] The self-refinement protocol in Eq. (1) includes an Inference Correction branch with the prompt "The correct answer is {···}. You need to correct the previous reasoning process and conclusion to ensure the final answer is right." If this branch is applied to the MACE follow-up stage, as the description of hierarchical reasoning integrating "relevant complications and MACE follow-up data one after another" suggests, then the correct recurrence label is explicitly provided during reasoning generation. The ablation gain from the MACE module (Table 2: D+C+M 0.8008 vs D+C 0.7881) is therefore plausibly a circular-feature artifact. The authors must either prove that the outcome is never used at test time or redesign the reasoning generation to hide the label.
[Section 3.1, Section 3.4, Table 1] The central 7.53% C-index improvement is reported without confidence intervals, without repeated cross-validation, and without external validation, on a dataset with only 104 events. Given the small event count, differences such as 0.8008 (CardioCoT-InternVL2.5) versus 0.8115 (CardioCoT-LLama3.1 70B) are not shown to be statistically significant, and the same holds for the comparison against the MMDB baseline. At minimum, the authors should report bootstrap confidence intervals and a significance test, and ideally validate on an independent cohort, before claiming state-of-the-art performance.
[Section 3.2, Table 1] The baselines PORPOISE and MMDB are multimodal survival models that do not use the reasoning trajectories. There is no fair non-reasoning LLM/VLM baseline where the same underlying model (e.g., Qwen2-VL or Llama3.1) is given the same inputs except without the hierarchical reasoning text, or where the reasoning text is replaced with a non-reasoning summary. Without such a control, the contribution of the reasoning mechanism itself, as opposed to the additional text modality or larger model capacity, is not identified.

minor comments (5)

[Throughout] There are typographical issues, including "T able 1" and "InterVL2.5" (presumably InternVL2.5), and inconsistent spacing in the references section.
[Section 3.3] The hyperparameter beta in the survival loss is set to 0.4, and the number of MRI slices is fixed at 10, but no sensitivity analysis is provided for either choice, so the robustness of the reported results to these settings is unclear.
[Section 2.3, Eq. (2)] The attention pooling formula has a minor notation issue: the denominator uses a summation index j while the numerator uses m, making the expression slightly ambiguous; clarifying the indices would improve readability.
[Section 3.1] The definition of the four survival intervals as "based on uncensored patients" is not fully specified; the authors should state how the interval boundaries are computed and whether censoring is handled consistently with the survival loss.
[Section 3.4, Fig. 4] The attention score analysis in Fig. 4 is qualitative and based on a few examples; reporting the distribution of attention weights over the whole test set would strengthen the interpretability claim.

Circularity Check

1 steps flagged · score 8.0 of 10

Central C-index gain appears to be an artifact of outcome leakage: MACE follow-up reasoning (I_M) is an input to the survival model but is generated from the recurrence status the model is supposed to predict.

self definitional [Section 2.1 (Hierarchical Evidence-augmented Reasoning Generation), Eq. (1); Section 2.2; Section 2.3 (Reasoning-Enhanced Survival Analysis)]
"Meanwhile, MACE follow-up tracks the recurrence status of MACE over a specified time period. ... Finally, we generate reasoning-enhanced endpoint-related evidence I_D, I_C, and I_M, which serve as part of multimodal inputs for final survival analysis. ... Inference Correction: The correct answer is {···}. You need to correct the previous reasoning process and conclusion to ensure the final answer is right."

The survival model's endpoint is MACE recurrence risk, and I_M is defined as reasoning over MACE follow-up, which 'tracks the recurrence status of MACE'—i.e., the event label itself. This label-derived reasoning is then explicitly included as a multimodal input to the same survival model that predicts recurrence. Eq. (1)'s Inference Correction branch even injects 'the correct answer' into the reasoning generator. Unless the paper states that recurrence status is withheld during test-time generation of I_M, the reported C-index gain (0.8342 vs 0.7589) is attributable to the model reading the outcome from its own input, making the feature construction circular by definition.

full rationale

The paper's central claimed improvement is the C-index gain from CardioCoT over PORPOISE and MMDB. The load-bearing step is the use of I_M, the 'MACE follow-up' reasoning trajectory, as an input to the final survival model. Section 2.1 defines MACE follow-up as tracking the recurrence status of MACE, and Section 2.2 says I_M is 'endpoint-related evidence' that serves as part of the multimodal inputs for final survival analysis. Since the endpoint is MACE recurrence, I_M is constructed from the very outcome the model predicts. The self-refinement loop in Eq. (1) further provides 'the correct answer' during reasoning correction, so the generation process is explicitly label-conditioned. The paper never states that the true recurrence status is hidden when I_M is generated at test time, and the described pipeline integrates follow-up data into the reasoning process. Thus, the reported improvement is plausibly a label-leakage artifact rather than an independently derived predictive signal. No other circularity was found: the use of GPT-4o for trajectory generation, SFT on open-source models, and the survival loss from PORPOISE [5] are external tools, not circular arguments. The MMDB baseline [21] involves a co-author, but the comparison itself is not used as evidence for the method's validity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method relies on several hand-chosen hyperparameters and unstated assumptions about outcome handling and data independence. The most serious is the unspecified treatment of MACE follow-up labels in the reasoning generation pipeline, which if violated would make the reported improvement an artifact.

free parameters (3)

beta (survival loss weighting) = 0.4
Section 3.3: 'beta is set to 0.4 for fair comparison'; controls trade-off in the survival loss and can affect C-index.
number of MRI slices sampled = 10
Section 3.3: each MRI scan is sampled at 10 evenly spaced slices; chosen by hand, not justified.
self-refinement max iterations N = 2
Section 2.1: iterative refinement continues at most N=2; hand-chosen and affects reasoning quality.

assumptions (4)

domain assumption The random 6:2:2 split produces independent training and test patients with no information leakage.
Section 3.1; no patient-level grouping or repeated splits are described, so this assumption is unverified.
domain assumption GPT-4o's self-refined reasoning provides a reliable and unbiased source of clinical evidence.
Section 2.1; the entire first stage depends on GPT-4o's outputs, which are not validated for downstream correctness beyond the Oracle's exact-match check.
domain assumption The MACE follow-up status is not part of the reasoning given to the survival model at inference; only the fine-tuned model's prediction is used.
Sections 2.1 and 2.2; this is never explicitly stated, and if false the target label leaks into the model inputs.
domain assumption The 871-patient in-house cohort is representative of the broader acute myocardial infarction population.
Section 3.1; no external validation or demographic breakdown is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CardioCoT: Hierarchical Reasoning for Multimodal Survival Analysis." pith.science (2026). https://pith.science/paper/P4V3TIX6

@misc{pith2026250519195,
  author       = {Pith},
  title        = {Pith review of: CardioCoT: Hierarchical Reasoning for Multimodal Survival Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P4V3TIX6}},
  note         = {Machine review of arXiv:2505.19195}
}

read the original abstract

Accurate prediction of major adverse cardiovascular events recurrence risk in acute myocardial infarction patients based on postoperative cardiac MRI and associated clinical notes is crucial for precision treatment and personalized intervention. Existing methods primarily focus on risk stratification capability while overlooking the need for intermediate robust reasoning and model interpretability in clinical practice. Moreover, end-to-end risk prediction using LLM/VLM faces significant challenges due to data limitations and modeling complexity. To bridge this gap, we propose CardioCoT, a novel two-stage hierarchical reasoning-enhanced survival analysis framework designed to enhance both model interpretability and predictive performance. In the first stage, we employ an evidence-augmented self-refinement mechanism to guide LLM/VLMs in generating robust hierarchical reasoning trajectories based on associated radiological findings. In the second stage, we integrate the reasoning trajectories with imaging data for risk model training and prediction. CardioCoT demonstrates superior performance in MACE recurrence risk prediction while providing interpretable reasoning processes, offering valuable insights for clinical decision-making.

Figures

Figures reproduced from arXiv: 2505.19195 by the authors.

**Figure 1.** Comparison of LLM/LLM utilization in survival analysis: 1) Promptbased: risk estimation via direct survival prompt; 2) Format-based: structured covariates extraction from unstructured data; 3) Embedding-based: semantic embedding extraction from pre-trained LLM/VLM; 4) Ours: hierarchical reasoning enabling evidential support and interpretable analysis. LLM/VLM to predict survival probabilities at specific time [10]… view at source ↗

**Figure 2.** The proposed CardioCoT consists of two stages: (1) GPT-4o generates hierarchical evidence-augmented reasoning with self-refinement, fine-tuned on advanced LLMs/VLMs; (2) MRI scans and reasoning outputs are encoded by vision and text encoders, fused via an attention module for final risk prediction. augmented self-refinement scheme to build a hierarchical set of reasoning trajectories for tuning our CardioCoT model … view at source ↗

**Figure 3.** The KM analysis curves results. is determined by the median predicted risk. A larger gap between the predicted survival probability curves for these groups indicates better risk stratification capability. Our method achieves a p-value of 7.37e-6, which significantly outperforms the baseline methods, demonstrating its strong effectiveness in risk stratification. Additionally, given the introduction of hierarchical i… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Attention quantification of different parts for survival analysis [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Qualitative result of a case for prediction complications with reasoning. experimental results highlight the critical role of each module and demonstrate that integrating all three reasoning components effectively improves overall performance. 4 Conclusion In this pape…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

A Composite Alignment-Aware Framework for Myocardial Lesion Segmentation in Multi-sequence CMR Images
eess.IV 2025-07 conditional novelty 6.0 of 10

CAA-Seg combines selective slice alignment with hierarchical feature fusion to improve infarction, edema, and myocardium segmentation on a 397-patient multi-sequence CMR dataset.

Reference graph

Works this paper leans on

27 extracted references · 10 canonical work pages · cited by 1 Pith paper

[1]

arXiv preprint arXiv:2303.08774 (2023)

Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

arXiv 2023
[2]

arxiv, 2022

Agrawal, M., Hegselmann, S., Lang, H., Kim, Y., Sontag, D.: Large language models are zero-shot clinical information extractors. arxiv, 2022. arXiv preprint arXiv:2205.12689 (2023)

arXiv 2023
[3]

Journal of medical systems41, 1–8 (2017)

Bush, R.A., Kuelbs, C., Ryu, J., Jiang, W., Chiang, G.: Structured data entry in the electronic medical record: perspectives of pediatric specialty physicians and surgeons. Journal of medical systems41, 1–8 (2017)

work page 2017
[4]

arXiv preprint arXiv:2412.18925 (2024)

Chen, J., Cai, Z., Ji, K., Wang, X., Liu, W., Wang, R., Hou, J., Wang, B.: Huatuogpt-o1, towards medical complex reasoning with llms. arXiv preprint arXiv:2412.18925 (2024)

arXiv 2024
[5]

Cancer cell40(8), 865–878 (2022)

Chen, R.J., Lu, M.Y., Williamson, D.F., Chen, T.Y., Lipkova, J., Noor, Z., Shaban, M., Shady, M., Williams, M., Joo, B., et al.: Pan-cancer integrative histology- genomic analysis via multimodal deep learning. Cancer cell40(8), 865–878 (2022)

2022
[6]

arXiv preprint arXiv:2412.05271 (2024)

Chen, Z., Wang, W., Cao, Y., Liu, Y., Gao, Z., Cui, E., Zhu, J., Ye, S., Tian, H., Liu, Z., et al.: Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271 (2024)

arXiv 2024
[7]

Journal of the Royal Statistical So- ciety: Series B (Methodological)34(2), 187–202 (1972)

Cox, D.R.: Regression models and life-tables. Journal of the Royal Statistical So- ciety: Series B (Methodological)34(2), 187–202 (1972)

1972
[8]

arXiv preprint arXiv:2407.21783 (2024)

Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al.: The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

arXiv 2024

Show all 27 references

[9]

arXiv preprint arXiv:2306.00024 (2023)

Gero, Z., Singh, C., Cheng, H., Naumann, T., Galley, M., Gao, J., Poon, H.: Self-verification improves few-shot clinical information extraction. arXiv preprint arXiv:2306.00024 (2023)

2023 arXiv
[10]

medRxiv pp

Han, C., Kim, D.W., Kim, S., You, S.C., Bae, S., Yoon, D.: Large-language-model- based 10-year risk prediction of cardiovascular disease: insight from the uk biobank data. medRxiv pp. 2023–05 (2023)

2023
[11]

ICLR1(2), 3 (2022)

Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. ICLR1(2), 3 (2022)

2022
[12]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700–4708 (2017)

2017
[13]

In: International conference on machine learning

Ilse,M.,Tomczak,J.,Welling,M.:Attention-baseddeepmultipleinstancelearning. In: International conference on machine learning. pp. 2127–2136. PMLR (2018)

2018
[14]

In: AAAI 2024 Spring Symposium on Clinical Foundation Models (2024)

Jeanselme, V., Agarwal, N., Wang, C.: Review of language models for survival analysis. In: AAAI 2024 Spring Symposium on Clinical Foundation Models (2024)

2024
[15]

NPJ precision oncology8(1), 4 (2024)

Jiang, L., Xu, C., Bai, Y., Liu, A., Gong, Y., Wang, Y.P., Deng, H.W.: Autosurv: interpretable deep learning framework for cancer survival analysis incorporating clinical and multi-omics data. NPJ precision oncology8(1), 4 (2024)

2024
[16]

Jiang, S., Gan, Z., Cai, L., Wang, Y., Zhang, Y.: Multimodal cross-task interaction forsurvivalanalysisinwholeslidepathologicalimages.In:InternationalConference on Medical Image Computing and Computer-Assisted Intervention. pp. 329–339. Springer (2024) 10 S. Rui et al

2024
[17]

In: Challenges of Trustable AI and Added-Value on Health, pp

Kaka, H., Michalopoulos, G., Subendran, S., Decker, K., Lambert, P., Pitz, M., Singh, H., Chen, H.: Pretrained neural networks accurately identify cancer recur- rence in medical record. In: Challenges of Trustable AI and Added-Value on Health, pp. 93–97. IOS Press (2022)

2022
[18]

In: International Conference on Medical Image Computing and Computer- Assisted Intervention

Kim, K., Lee, Y., Park, D., Eo, T., Youn, D., Lee, H., Hwang, D.: Llm-guided multi-modal multiple instance learning for 5-year overall survival prediction of lung cancer. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 239–249. ...

2024
[19]

Frontiers in Oncology11, 747250 (2021)

Kim, S., Lee, C.k., Choi, Y., Baek, E.S., Choi, J.E., Lim, J.S., Kang, J., Shin, S.J.: Deep-learning-based natural language processing of serial free-text radiologi- cal reports for predicting rectal cancer patient survival. Frontiers in Oncology11, 747250 (2021)

2021
[20]

In: Proceedings of the conference

Lee, H.G., Sholle, E., Beecy, A., Al’Aref, S., Peng, Y.: Leveraging deep represen- tations of radiology reports in survival analysis for predicting heart failure patient mortality. In: Proceedings of the conference. Association for Computational Lin- guistics. North American C...

2021
[21]

In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention

Qu, L., Huang, D., Zhang, S., Wang, X.: Multi-modal data binding for survival analysis modeling with incomplete data and annotations. In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention. pp. 501–510. Springer (2024)

2024
[22]

Journal of machine learning research21(140), 1–67 (2020)

Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research21(140), 1–67 (2020)

2020
[23]

The Journal of Pathology262(3), 310–319 (2024)

Truhn, D., Loeffler, C.M., Müller-Franzes, G., Nebelung, S., Hewitt, K.J., Brand- ner,S.,Bressem,K.K.,Foersch,S.,Kather,J.N.:Extractingstructuredinformation from unstructured histopathology reports using generative pre-trained transformer 4 (gpt-4). The Journal of Pathology262...

2024
[24]

arXiv preprint arXiv:2409.12191 (2024)

Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., Fan, Y., Dang, K., Du, M., Ren, X., Men, R., Liu, D., Zhou, C., Zhou, J., Lin, J.: Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint...

2024 arXiv
[25]

arXiv e-prints pp

Wei, X., Cui, X., Cheng, N., Wang, X., Zhang, X., Huang, S., Xie, P., Xu, J., Chen, Y., Zhang, M., et al.: Zero-shot information extraction via chatting with chatgpt. arXiv e-prints pp. arXiv–2302 (2023)

2023
[26]

Artificial Intelligence Review57(3), 65 (2024)

Wiegrebe, S., Kopper, P., Sonabend, R., Bischl, B., Bender, A.: Deep learning for survival analysis: a review. Artificial Intelligence Review57(3), 65 (2024)

2024
[27]

In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)

Zheng, Y., Zhang, R., Zhang, J., Ye, Y., Luo, Z., Feng, Z., Ma, Y.: Llamafac- tory: Unified efficient fine-tuning of 100+ language models. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations). Association...

2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

arXiv preprint arXiv:2303.08774 (2023)

Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

arXiv 2023

[2] [2]

arxiv, 2022

Agrawal, M., Hegselmann, S., Lang, H., Kim, Y., Sontag, D.: Large language models are zero-shot clinical information extractors. arxiv, 2022. arXiv preprint arXiv:2205.12689 (2023)

arXiv 2023

[3] [3]

Journal of medical systems41, 1–8 (2017)

Bush, R.A., Kuelbs, C., Ryu, J., Jiang, W., Chiang, G.: Structured data entry in the electronic medical record: perspectives of pediatric specialty physicians and surgeons. Journal of medical systems41, 1–8 (2017)

work page 2017

[4] [4]

arXiv preprint arXiv:2412.18925 (2024)

Chen, J., Cai, Z., Ji, K., Wang, X., Liu, W., Wang, R., Hou, J., Wang, B.: Huatuogpt-o1, towards medical complex reasoning with llms. arXiv preprint arXiv:2412.18925 (2024)

arXiv 2024

[5] [5]

Cancer cell40(8), 865–878 (2022)

Chen, R.J., Lu, M.Y., Williamson, D.F., Chen, T.Y., Lipkova, J., Noor, Z., Shaban, M., Shady, M., Williams, M., Joo, B., et al.: Pan-cancer integrative histology- genomic analysis via multimodal deep learning. Cancer cell40(8), 865–878 (2022)

2022

[6] [6]

arXiv preprint arXiv:2412.05271 (2024)

Chen, Z., Wang, W., Cao, Y., Liu, Y., Gao, Z., Cui, E., Zhu, J., Ye, S., Tian, H., Liu, Z., et al.: Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271 (2024)

arXiv 2024

[7] [7]

Journal of the Royal Statistical So- ciety: Series B (Methodological)34(2), 187–202 (1972)

Cox, D.R.: Regression models and life-tables. Journal of the Royal Statistical So- ciety: Series B (Methodological)34(2), 187–202 (1972)

1972

[8] [8]

arXiv preprint arXiv:2407.21783 (2024)

Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al.: The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

arXiv 2024