Pith. sign in

REVIEW 4 major objections 5 minor 25 references

CIGTSurv: Clinical Information Guided Tri-modal Survival Prediction with Local Prototype Association and Global Feature Alignment

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read CIGTSurv claims that rewriting clinical records as text and aligning pathology and genomic features around them improves cancer survival prediction, reaching a mean C-index of 0.788 across five cohorts.

desk verdict Sensible tri-modal fusion with clinical anchors, but the SOTA claim is unverified without the MMSurv comparison and significance tests. read the letter →

arxiv 2608.03247 v1 pith:FLNYBKOE submitted 2026-08-04 cs.CV cs.CL

Jing Dai , Qibin Zhang , Weiwei Zhou , Mingde Xu , Jingsong Liu , Jingdong Zhang , Hongming Xu This is my paper
classification cs.CVcs.CL
keywords survivalpredictionmultimodallearningclinicaltextencodingcross-attentionmaximummeandiscrepancyprototypecomputationalpathologycancerprognosis
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

CIGTSurv sets out to show that clinical information, normally a handful of categorical variables, can be turned into a useful third modality for survival prediction by rewriting it as a holistic sentence and embedding it with a biomedical language model. The framework anchors pathology images and genomic data to that clinical text through two complementary mechanisms: local prototype association, a cross-attention step in which clinical tokens query compact prototypes distilled from whole-slide images and gene sets, and global feature alignment, an MMD loss that pulls the modality distributions together. Across five cancer cohorts with 5-fold cross-validation, the paper reports a mean C-index of 0.788, higher than every uni-, dual-, and tri-modal baseline it compares against, and its ablations attribute the gain to the text embedding, the prototypes, and both interaction modules. If the result is right, low-dimensional clinical tables are not a weak afterthought in multimodal oncology models but a workable semantic anchor around which high-dimensional image and genomic features can be aligned.

What carries the argument

The load-bearing machinery is the clinical-anchored dual-level interaction: LPA is cross-attention in which clinical embeddings provide the queries and prototype-based pathology or genomic embeddings provide keys and values, producing clinic-aware fused sequences; GFA is a squared Maximum Mean Discrepancy between each fused sequence and the clinical anchor in a Gaussian RKHS, added to the survival loss as $\lambda L_{\text{align}}$. Prototypes make the attention tractable: the WSI is reduced from thousands of patches to $K=50$ clusters with 10 sampled patches each (500 tokens), and gene expression is grouped into $n_g=6$ biological function sets, so the cross-attention operates on a few hund

What would settle it

Run CIGTSurv and the strongest compared baselines on identical cohorts, folds, early-stopping rules, and random seeds. If the mean C-index no longer exceeds the best dual-modal baseline (reported as 0.775 for MOTCat) or the ablation gaps in Table 2 shrink to noise, the SOTA and component claims are falsified. A targeted test of the text-template claim: replace the biomedical language model embeddings with one-hot clinical features at equal parameter count; if C-index does not drop, the semantic encoding is not the source of gain.

Watch

Extended reading notes

Core claim

The central claim is that a tri-modal survival model with clinical text as the anchor and prototypes as the attention tokens can surpass established multimodal predictors. Concretely, the paper constructs a text template from age, race, gender, pathologic stage, and TNM status, encodes it with a pretrained biomedical language model, and uses the resulting clinical tokens as queries in cross-attention over two prototype sets: K-means clusters of WSI patch features and six predefined biological gene-function sets. The aligned pathology and genomic features are then pulled toward the clinical distribution by a squared MMD loss, and the three pooled embeddings are concatenated and fed to a linea

Load-bearing premise

The central claim assumes the baselines were run on exactly the same preprocessed patient cohorts with fair, comparable tuning; Section 3.1 says incomplete cases were excluded and 5-fold cross-validation was used, but not enough per-baseline detail is given to verify this.

Editorial extensions

If this is right

  • If CIGTSurv's result holds, clinical information can be treated as a third learned modality rather than as a few categorical inputs: the holistic text template captures dependencies among age, race, gender, stage, and TNM status that one-hot vectors miss.
  • Prototype-based attention means gigapixel WSIs can be distilled to a few hundred representative patches and gene expression to a handful of biological function sets without losing the cross-modal signal needed for survival prediction.
  • The dual-level mechanism, explicit cross-attention for local correspondences plus MMD for global distribution alignment, offers a recipe for other medical multimodal tasks in which one modality is low-dimensional and the others are high-dimensional.
  • The reported Kaplan-Meier separation of high- and low-risk groups across all five cohorts supports the model's potential for patient stratification in clinical decision support.

Reading between the lines

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

  • The paper does not report per-baseline tuning details or the exact patient subset after excluding incomplete cases, so the SOTA claim is only as strong as the comparability assumption; a shared benchmark with fixed cohorts, folds, and seeds would settle it.
  • If the gain is really from semantic text embeddings, then a randomized text template (the same sentence with shuffled variables) should lose performance; if it does not, the benefit may be structural rather than semantic.
  • MMD alignment to the clinical anchor is essentially unsupervised domain adaptation, so the framework may support inference when one modality is missing, for example pathology absent, by aligning remaining features to clinical text; this is an untested consequence.
  • Because patch selection samples randomly within each K-means cluster, the reported C-index likely varies with the sampling seed; reporting that variance would clarify how much of the improvement is attributable to prototype composition.
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.

Referee Report

4 major / 5 minor

Summary. CIGTSurv proposes a tri-modal survival prediction framework that converts clinical tabular data into a holistic text description encoded by BiomedBERT, and uses the clinical embedding as an anchor for two interaction mechanisms: a local prototype association (LPA) module based on cross-attention over pathology and genomic prototypes, and a global feature alignment (GFA) loss based on MMD. The fused features are pooled and used with an NLL survival loss. Experiments on five TCGA cohorts (BLCA, BRCA, COADREAD, LIHC, KIRP) report mean C-index 0.788, outperforming uni-, dual-, and tri-modal baselines. The source code is publicly available.

Significance. If the performance claims hold, the work makes a useful contribution by showing how structured clinical variables can be embedded as text and used as a query/anchor for cross-modal fusion, and by combining explicit token-level association with distribution-level alignment. Strengths include a clearly specified architecture, consistent ablations (each component removal lowers mean C-index), KM stratification with log-rank tests, and public code. I find no circularity: survival NLL is label-based, and LPA/GFA are label-free regularizers. However, the headline SOTA conclusion is currently under-supported because the most relevant tri-modal baseline is omitted and because the reported gains over strong baselines are small and not significance-tested.

major comments (4)
  1. [Sec. 3.2, Table 1] MMSurv (Ref. [13], Briefings in Bioinformatics 2025), which the Introduction cites as evidence for the importance of clinical information, integrates pathological images, clinical information, and sequencing data — exactly the same tri-modal setting as CIGTSurv. It is not included in any comparison row of Table 1. Without this baseline, the abstract's claim of achieving 'state-of-the-art survival prediction performance' cannot be evaluated; the claim is conditional on an untested competitor. Please add MMSurv to the comparison (or, if access is impossible, clearly delimit the claim to the compared methods).
  2. [Sec. 3.2, Table 1] The headline improvement is modest: mean C-index 0.788 vs 0.775 for MOTCat and 0.762 for SurvPGC, and the paper reports no significance testing. With only five datasets and 5-fold cross-validation, the reported differences may be within fold/split variability. I recommend paired bootstrap or permutation tests over patient-level predictions (or per-dataset confidence intervals) to support 'surpassing all other comparative methods.' Without this, the SOTA claim is not statistically grounded.
  3. [Sec. 3.1] The comparability of baselines is underspecified. The text says 'incomplete cases were excluded' and that all methods used 5-fold cross-validation, but it does not state whether each baseline used the identical patient subset, the same patch/feature preprocessing (e.g., same UNI features and K-means prototype settings), the same hyperparameter search budget, early-stopping criterion, or random seeds. Since several baselines' published results may have been obtained under different preprocessing, a direct comparison requires either a single unified evaluation protocol with per-baseline tuning described, or released scripts/configs that reproduce each baseline row of Table 1.
  4. [Sec. 3.1, Eq. (7)] The alignment weight λ is selected per dataset from {0.01, 0.05, 0.1, 0.5, 1} with dataset-specific values. It is not stated whether this selection is performed on held-out validation folds within each cross-validation split or on the test folds. If the latter, the reported C-indices are optimistically biased by test-set selection, and comparisons with baselines that did not receive equivalent test-set tuning are invalid. Please clarify the validation protocol and, ideally, report sensitivity to λ.
minor comments (5)
  1. [Abstract/Introduction] Minor typos: 'patient' s' in the Abstract and 'Tri-timodal' in the Introduction should be 'patient's' and 'tri-modal'.
  2. [Sec. 2.1] 'to from a fused feature representation' appears to be a typo for 'to form'.
  3. [Sec. 2.5] 'feature tockens' should be 'feature tokens'.
  4. [Table 1] The caption states 'The best and the second best results are highlighted,' but in the plain text no highlighting is visible. Please mark best/second-best entries with bold/underline consistently.
  5. [Sec. 2.5, Eq. (5)] The MMD estimator is written with a single N for both samples. Since the two feature sets may differ in size (or for clarity), it would be cleaner to use N_x and N_y for the two empirical means.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the survival objective is label-based; all modality interaction terms are label-free regularizers.

full rationale

The paper's central claim is an empirical performance comparison, not a derivation from a premise that contains the conclusion. The survival prediction is trained with a Negative Log-Likelihood loss computed from true survival labels (Sec. 2.6, Eq. 7), and the two proposed interaction mechanisms—LPA (cross-attention) and GFA (MMD)—are label-free regularizers applied to feature distributions (Secs. 2.4-2.5, Eqs. 2-6). No equation defines the target outcome in terms of the input by construction: clinical text is built from age, race, gender, stage, and TNM, all of which are input covariates, not derived from survival labels. The reported C-index is evaluated against external TCGA cohorts and previously published baselines (Sec. 3.2, Table 1), so the comparison is not forced by a fitted parameter renamed as a prediction. The paper contains no load-bearing self-citation: the cited prior works (e.g., MCAT, SurvPath, SurvPGC, UNI, BiomedBERT) are external methods or pretrained models, and no 'uniqueness theorem' from the current authors is invoked. The lambda hyperparameter is tuned on validation data, which is standard practice and does not reduce the central claim to a fit. The skeptic's point about the omitted MMSurv baseline is a comparison completeness concern, not a circularity step: an omitted baseline weakens the strength of the SOTA claim but does not make the derivation self-referential. Overall, the derivation chain is self-contained and no circular step can be exhibited from the paper's equations or citation structure.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

No new physical or mathematical entities are introduced. The central claim rests on one fitted hyperparameter (lambda, per dataset), several hand-chosen prototype sizes, and standard domain assumptions about pretrained encoders and biological gene groupings. The most consequential free parameter is lambda because it is tuned per dataset and directly weights the alignment loss in every reported result.

free parameters (5)
  • Alignment weight lambda = BLCA 0.01, BRCA 1, COADREAD 0.1, LIHC 0.01, KIRP 0.5
    Selected per dataset from {0.01, 0.05, 0.1, 0.5, 1} (Section 3.1) to weight L_align in Eq. 7; selection procedure is not described and directly affects all results.
  • Number of pathology clusters K and sampled patches per cluster S = K=50, S=10, n_p=500
    Chosen in Section 2.3 following prior work [19]; hand-selected prototype sizes that determine input sequence length and influence cross-attention.
  • Number of genomic prototype sets n_g = 6
    Pre-defined biological function sets from [8] (Section 2.3); the grouping into 6 sets is a design choice that determines the genomic token count.
  • Embedding dimension d = 256
    Set in Section 2.2; architectural choice for all modality projections.
  • Gaussian kernel bandwidth sigma = median heuristic
    Computed by median heuristic (Section 3.1); a data-dependent choice that sets the MMD kernel scale.
assumptions (7)
  • domain assumption The clinical text template with age, race, gender, stage, and TNM captures survival-relevant context.
    Section 2.2 constructs the template from these five variables and uses it as the alignment anchor; omitted variables such as treatment or comorbidities could weaken the anchor.
  • domain assumption Pretrained UNI and BiomedBERT representations transfer to TCGA pathology and clinical text without domain adaptation.
    Sections 2.2 and 2.3 use frozen or projected pretrained embeddings; if these features are not informative for TCGA, the pipeline loses signal.
  • domain assumption The six pre-defined biological function sets from [8] are an appropriate grouping for survival-relevant gene expression.
    Section 2.3 creates six prototypes from these sets; the grouping is taken from prior literature and is not validated in this paper.
  • domain assumption Gaussian-kernel MMD with median-heuristic bandwidth measures cross-modal distribution mismatch in a way that improves survival prediction.
    Section 2.5 introduces L_align; the choice of kernel and bandwidth is heuristic.
  • domain assumption K-means into 50 clusters and sampling 10 patches per cluster preserves the prognostic content of the WSI.
    Section 2.3 reduces N_p patches to 500; if the sampled patches miss informative regions, performance suffers. Adopted from [19].
  • domain assumption Excluding cases with incomplete clinical variables yields an unbiased cohort for comparing methods.
    Section 3.1 excludes incomplete cases; if exclusion is correlated with outcome or with method behavior, the comparison is biased.
  • standard math Survival NLL correctly handles right-censored time-to-event data.
    Section 2.6 adopts NLL loss from prior survival models [8,10]; standard in survival analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CIGTSurv: Clinical Information Guided Tri-modal Survival Prediction with Local Prototype Association and Global Feature Alignment." pith.science (2026). https://pith.science/paper/FLNYBKOE

@misc{pith2026260803247,
  author       = {Pith},
  title        = {Pith review of: CIGTSurv: Clinical Information Guided Tri-modal Survival Prediction with Local Prototype Association and Global Feature Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FLNYBKOE}},
  note         = {Machine review of arXiv:2608.03247}
}
read the original abstract

Multimodal learning has significantly advanced survival prediction by integrating pathology images with genomic data. However, clinical information, despite its critical role in reflecting a patient' s overall health, remains underutilized due to its discrete, sparse, and low-dimensional nature. Furthermore, the inherent heterogeneity across these modalities pose significant challenges in modeling cross-modal interactions. In this paper, we propose CIGTSurv, a Clinical Information Guided Tri-modal framework for Survival prediction. Specifically, we first design a holistic text template and use pretrained foundation models to transform clinical tabular data into high-dimensional tokenized embeddings. Using clinical information as an anchor, we then introduce a dual-level interaction mechanism: 1) a local prototype association (LPA) module based on cross-attention to explicitly learn token-level correspondences between different modalities, and 2) a global feature alignment (GFA) loss based on Maximum Mean Discrepancy (MMD) to implicitly enhance cross-modal distribution consistency. Extensive experiments on five TCGA cancer cohorts demonstrate that CIGTSurv achieves state-of-the-art (SOTA) survival prediction performance. Our source code is publicly available at https://github.com/Daijing-ai/CIGT-Surv.git.

Figures

Figures reproduced from arXiv: 2608.03247 by the authors.

Figure 1
Figure 1. Overview of our proposed CIGTSurv framework featured by (a) text encoding from clinical information, (b) local prototype association and (c) global feature align￾ment. 2 Methods 2.1 Overview The overall architecture of the proposed CIGTSurv is shown in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Kaplan-Meier (KM) curves of the proposed model on five cancer datasets, where patient stratification of low risk (blue) and high risk (orange) are presented. presents the comparative results, where our proposed CIGTSurv achieves an av￾erage C-index of 78.8%, surpassing all other comparative methods. The superior performance is mainly attributed to its dual-level mechanism, which effectively bridges the modality gap … view at source ↗
Figure 3
Figure 3. T-SNE visualization of the feature distributions. (a) Feature distributions in the baseline. (b) Feature distributions after incorporating LPA. (c) Feature distributions by our proposed CIGTSurv framework. 3.3 Ablation Studies We conduct comprehensive ablation studies from four aspects to evaluate our proposed method, as shown in [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 20 canonical work pages

  1. [13]

    Hailong Yang, Jia Wang, Wenyan Wang, Shufang Shi, Lijing Liu, Yuhua Yao, Geng Tian, Peizhen Wang, and Jialiang Yang. Mmsurv: a multimodal multi-instance multi-cancer survival prediction model integrating pathological images, clinical in- formation, and sequencing data.Briefings in Bioinformatics, 26(3):bbaf209, 2025

    work page 2025

  2. [1]

    Multimodal data integration for precision oncology: Challenges and future directions.arXiv preprint arXiv:2406.19611, 2024

    Huajun Zhou, Fengtao Zhou, Chenyu Zhao, Yingxue Xu, Luyang Luo, and Hao Chen. Multimodal data integration for precision oncology: Challenges and future directions.arXiv preprint arXiv:2406.19611, 2024

    arXiv 2024

  3. [2]

    Attention-based deep multiple instance learning

    Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. InInternational conference on machine learning, pages 2127–

  4. [3]

    Data-efficient and weakly supervised computational pathology on whole-slide images.Nature biomedical engineering, 5(6):555–570, 2021

    Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Bar- bieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images.Nature biomedical engineering, 5(6):555–570, 2021

    work page 2021

  5. [4]

    Transmil: Transformer based correlated multiple instance learning for whole slide image classification.Advances in neural information processing systems, 34:2136–2147, 2021

    Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification.Advances in neural information processing systems, 34:2136–2147, 2021. 10 J. Dai et al

    work page 2021

  6. [5]

    Feature re-embedding: Towards foundation model-level performance in computa- tional pathology

    Wenhao Tang, Fengtao Zhou, Sheng Huang, Xiang Zhu, Yi Zhang, and Bo Liu. Feature re-embedding: Towards foundation model-level performance in computa- tional pathology. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11343–11352, 2024

    work page 2024

  7. [6]

    Multiple instance classification: Review, taxonomy and compara- tive study.Artificial intelligence, 201:81–105, 2013

    Jaume Amores. Multiple instance classification: Review, taxonomy and compara- tive study.Artificial intelligence, 201:81–105, 2013

    work page 2013

  8. [7]

    Pan-cancer integrative histology-genomic analysis via multimodal deep learning

    Richard J Chen, Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Jana Lipkova, Zahra Noor, Muhammad Shaban, Maha Shady, Mane Williams, Bumjin Joo, et al. Pan-cancer integrative histology-genomic analysis via multimodal deep learning. Cancer cell, 40(8):865–878, 2022

    work page 2022

Show all 25 references
  1. [8]

    Multimodal co- attention transformer for survival prediction in gigapixel whole slide images

    Richard J Chen, Ming Y Lu, Wei-Hung Weng, Tiffany Y Chen, Drew FK Williamson, Trevor Manz, Maha Shady, and Faisal Mahmood. Multimodal co- attention transformer for survival prediction in gigapixel whole slide images. In Proceedings of the IEEE/CVF international conference on c...

    2021

  2. [9]

    Multimodal optimal transport-based co-attention transformer with global structure consistency for survival prediction

    Yingxue Xu and Hao Chen. Multimodal optimal transport-based co-attention transformer with global structure consistency for survival prediction. InProceed- ings of the IEEE/CVF international conference on computer vision, pages 21241– 21251, 2023

    2023

  3. [10]

    Modeling dense multimodal interactions between biological pathways and histology for survival prediction

    Guillaume Jaume, Anurag Vaidya, Richard J Chen, Drew FK Williamson, Paul Pu Liang, and Faisal Mahmood. Modeling dense multimodal interactions between biological pathways and histology for survival prediction. InProceedings of the IEEE/CVF Conference on Computer Vision and Patt...

    2024

  4. [11]

    Multimodal prototyping for cancer survival prediction.arXiv preprint arXiv:2407.00224, 2024

    Andrew H Song, Richard J Chen, Guillaume Jaume, Anurag J Vaidya, Alexan- der S Baras, and Faisal Mahmood. Multimodal prototyping for cancer survival prediction.arXiv preprint arXiv:2407.00224, 2024

    2024 arXiv

  5. [12]

    Cohort-individual cooperative learn- ing for multimodal cancer survival analysis.IEEE Transactions on Medical Imag- ing, 44(2):656–667, 2024

    Huajun Zhou, Fengtao Zhou, and Hao Chen. Cohort-individual cooperative learn- ing for multimodal cancer survival analysis.IEEE Transactions on Medical Imag- ing, 44(2):656–667, 2024

    2024

  6. [14]

    Multimodal data fusion for cancer biomarker discovery with deep learning.Nature machine intelli- gence, 5(4):351–362, 2023

    Sandra Steyaert, Marija Pizurica, Divya Nagaraj, Priya Khandelwal, Tina Hernandez-Boussard, Andrew J Gentles, and Olivier Gevaert. Multimodal data fusion for cancer biomarker discovery with deep learning.Nature machine intelli- gence, 5(4):351–362, 2023

    2023

  7. [15]

    A two-stage modeling approach for breast cancer survivability prediction.International journal of medical informatics, 149:104438, 2021

    Zahra Sedighi-Maman and Alexa Mondello. A two-stage modeling approach for breast cancer survivability prediction.International journal of medical informatics, 149:104438, 2021

    2021

  8. [16]

    Multimodal deep learning for cancer prognosis prediction with clinical information prompts integration.npj Digital Medicine, 2025

    Jiaxin Hou, Ranran Zhang, Yaoqin Xie, Chao Li, and Wenjian Qin. Multimodal deep learning for cancer prognosis prediction with clinical information prompts integration.npj Digital Medicine, 2025

    2025

  9. [17]

    Domain-specific language model pretraining for biomedical natural language processing.ACM Transactions on Computing for Healthcare (HEALTH), 3(1):1–23, 2021

    Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. Domain-specific language model pretraining for biomedical natural language processing.ACM Transactions on Computing for Healthcare (HEALTH), 3(1):1–23, 2021

    2021

  10. [18]

    Towards a general-purpose foundation model for computational pathology

    Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, Title Suppressed Due to Excessive Length 11 et al. Towards a general-purpose foundation model for computational pathology. Nature m...

    2024

  11. [19]

    Rui Yan, Xueyuan Zhang, Zihang Jiang, Baizhi Wang, Xiuwu Bian, Fei Ren, and S Kevin Zhou. Pathway-aware multimodal transformer (pamt): Integrating patho- logical image and gene expression for interpretable cancer survival analysis.IEEE Transactions on Pattern Analysis and Mach...

    2025

  12. [20]

    Self-normalizing neural networks.Advances in neural information processing sys- tems, 30, 2017

    Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing neural networks.Advances in neural information processing sys- tems, 30, 2017

    2017

  13. [21]

    A kernel two-sample test.The journal of machine learning research, 13(1):723–773, 2012

    Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bernhard Schölkopf, and Alexander Smola. A kernel two-sample test.The journal of machine learning research, 13(1):723–773, 2012

    2012

  14. [22]

    Analysis of survival data under the proportional hazards model.International Statistical Review/Revue Internationale de Statistique, pages 45–57, 1975

    Norman E Breslow. Analysis of survival data under the proportional hazards model.International Statistical Review/Revue Internationale de Statistique, pages 45–57, 1975

    1975

  15. [23]

    Pathomic fusion: an integrated framework for fusing histopathology and genomic features for cancer diagnosis and prognosis

    Richard J Chen, Ming Y Lu, Jingwen Wang, Drew FK Williamson, Scott J Rodig, Neal I Lindeman, and Faisal Mahmood. Pathomic fusion: an integrated framework for fusing histopathology and genomic features for cancer diagnosis and prognosis. IEEE Transactions on Medical Imaging, 41...

    2020

  16. [24]

    Multi-omics deep learning improves fdg pet-ct-based long-term prognostication of breast cancer.npj Precision Oncology, 2026

    Xinglong Liang, Tianyu Zhang, Miguel Braga, Luyi Han, Maarten Donswijk, Jiaju Huang, Jinhong Song, Chunyao Lu, Xin Wang, Yuan Gao, et al. Multi-omics deep learning improves fdg pet-ct-based long-term prognostication of breast cancer.npj Precision Oncology, 2026

    2026

  17. [25]

    Visualizing data using t-sne.Jour- nal of machine learning research, 9(11), 2008

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne.Jour- nal of machine learning research, 9(11), 2008

    2008

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.