arxiv: 2512.01116 · v3 · submitted 2025-11-30 · 💻 cs.CV

Structural Prognostic Event Modeling for Multimodal Cancer Survival Analysis

Yilan Zhang , Li Nanbo , Changchun Yang , J\"urgen Schmidhuber , Xin Gao This is my paper

Pith reviewed 2026-05-17 02:13 UTC · model grok-4.3

classification 💻 cs.CV

keywords slot attentionmultimodal cancer survivalprognostic eventshistology imagesgene profilessurvival analysisinterpretabilitymissing data robustness

0 comments

The pith

A slot attention model compresses multimodal cancer data into distinct prognostic events to improve survival predictions.

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

The paper sets out to show that high-dimensional histology images and gene profiles can be effectively summarized by a small set of compact, mutually distinctive slots, each standing in for a sparse prognostic event that drives patient outcomes. These slots are learned separately for each modality and then used to model interactions while allowing biological knowledge to be added directly. A reader would care because the approach claims to deliver both higher prediction accuracy and clearer explanations of why a given patient faces a certain risk, even when some genomic measurements are absent.

Core claim

By applying slot attention to compress each patient's multimodal inputs into compact, modality-specific sets of mutually distinctive slots, the framework encodes prognostic events in a way that supports efficient intra- and inter-modal interaction modeling and the direct inclusion of biological priors.

What carries the argument

Slot attention producing modality-specific sets of mutually distinctive slots that serve as encodings for sparse, patient-specific prognostic events.

If this is right

The model outperforms prior methods in eight of ten cancer cohorts.
It delivers an overall 2.9 percent gain in survival prediction performance.
Performance stays stable when genomic data is missing.
Interpretability improves through explicit decomposition into structured prognostic events.

Where Pith is reading between the lines

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

The slot decomposition may make it easier to connect model outputs to specific clinical or biological hypotheses for targeted follow-up.
The same compression strategy could be tested on other multimodal medical tasks such as treatment-response prediction.
Adding richer pathway-level priors during slot learning might increase the biological relevance of the discovered events.

Load-bearing premise

The learned slots genuinely correspond to meaningful sparse prognostic events rather than functioning only as a convenient data compression that improves scores.

What would settle it

A follow-up study that checks whether the extracted slots align with independently verified histologic patterns or known gene pathway activations in the same patient samples.

Figures

Figures reproduced from arXiv: 2512.01116 by Changchun Yang, J\"urgen Schmidhuber, Li Nanbo, Xin Gao, Yilan Zhang.

**Figure 1.** Figure 1: Framework of SlotSPE. Histology and gene features are extracted into bag structures, then compressed into slots via slot attention. Selective slot activation enforces sparsity and mutual competition, while a biologically guided cross-modal reconstruction aligns modalities. Finally, slot interactions are modeled using self- and cross-attention to predict survival. This allows us to compress the large input … view at source ↗

**Figure 2.** Figure 2: Detailed component structure of SlotSPE. (a) Selective slot activation: a Mixture-of [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Kaplan–Meier curves of predicted high-risk and low-risk groups. A p-value < 0.05 at the top indicates statistically significant separation between groups. The restricted mean survival time (RMST) up to 60 months is also reported, with values shown as ∆ (High–Low), and ratio (High/Low). (Zoom in to view details.) 4.3 ROBUSTNESS Since collecting genomic data is substantially more costly than acquiring histol… view at source ↗

**Figure 4.** Figure 4: Performance vs (Inference) Memory/Runtime Trade-off [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Interpretability of slots. (A) Original WSI, assignment map of histology-derived slots, [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗

**Figure 6.** Figure 6: Kaplan–Meier curves of predicted high-risk and low-risk groups. A p-value < 0.05 at the top indicates statistically significant separation between groups. The restricted mean survival time (RMST) up to 60 months is also reported, with values shown as ∆ (High–Low), and RMST ratio (High/Low). (Zoom in to view details.) 24 [PITH_FULL_IMAGE:figures/full_fig_p024_6.png] view at source ↗

**Figure 7.** Figure 7: Effect of different foundation models used as histology encoders. Performance (C-index) [PITH_FULL_IMAGE:figures/full_fig_p027_7.png] view at source ↗

**Figure 8.** Figure 8: Ablation on slot numbers. Boxplots show C-index across cohorts as the number of histol [PITH_FULL_IMAGE:figures/full_fig_p027_8.png] view at source ↗

**Figure 9.** Figure 9: Ablation study of hyperparameters. (a) Effect of different iteration number [PITH_FULL_IMAGE:figures/full_fig_p028_9.png] view at source ↗

**Figure 10.** Figure 10: Efficiency analysis. (a) Performance vs. inference memory/runtime trade-off across [PITH_FULL_IMAGE:figures/full_fig_p031_10.png] view at source ↗

**Figure 11.** Figure 11: Calibration curve, decision curve analysis, and Kaplan–Meier risk stratification for [PITH_FULL_IMAGE:figures/full_fig_p032_11.png] view at source ↗

**Figure 12.** Figure 12: Case study of patient-level interpretability in BRCA. (A) From left to right: original [PITH_FULL_IMAGE:figures/full_fig_p034_12.png] view at source ↗

**Figure 13.** Figure 13: Case study of patient-level interpretability in UCEC. (A) From left to right: original [PITH_FULL_IMAGE:figures/full_fig_p035_13.png] view at source ↗

**Figure 14.** Figure 14: Cohort-level pathway analysis for BRCA and UCEC. Patients are stratified into four [PITH_FULL_IMAGE:figures/full_fig_p036_14.png] view at source ↗

read the original abstract

The integration of histology images and gene profiles has shown great promise for improving survival prediction in cancer. However, current approaches often struggle to model intra- and inter-modal interactions efficiently and effectively due to the high dimensionality and complexity of the inputs. A major challenge is capturing critical prognostic events that, though few, underlie the complexity of the observed inputs and largely determine patient outcomes. These events, manifested as high-level structural signals such as spatial histologic patterns or pathway co-activations, are typically sparse, patient-specific, and unannotated, making them inherently difficult to uncover. To address this, we propose SlotSPE, a slot-based framework for structural prognostic event modeling. Specifically, inspired by the principle of factorial coding, we compress each patient's multimodal inputs into compact, modality-specific sets of mutually distinctive slots using slot attention. By leveraging these slot representations as encodings for prognostic events, our framework enables both efficient and effective modeling of complex intra- and inter-modal interactions, while also facilitating seamless incorporation of biological priors that enhance prognostic relevance. Extensive experiments on ten cancer benchmarks show that SlotSPE outperforms existing methods in 8 out of 10 cohorts, achieving an overall improvement of 2.9%. It remains robust under missing genomic data and delivers markedly improved interpretability through structured event decomposition.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SlotSPE gets small gains on ten cancer cohorts with slot attention but the claim that slots capture real prognostic events stays unverified.

read the letter

SlotSPE compresses histology images and gene profiles into modality-specific slots via slot attention, then treats those slots as encodings of sparse prognostic events to model interactions and improve survival prediction. The paper reports outperformance on eight of ten cohorts with a 2.9% overall lift and some robustness when genomic data is missing. That is the core result on offer. The framing around structural prognostic events is a fresh way to present the slot machinery for this task, and running the method across multiple real cancer benchmarks with a missing-data test is useful practical work. Those elements give the paper a clear technical baseline to build from. The soft spot is the interpretability story. Nothing in the reported experiments directly checks whether the learned slots line up with actual histologic patterns, pathway activations, or other biologically meaningful structures rather than just serving as an effective compression. Without that check the performance edge could come from standard representation learning benefits alone. Statistical details on the 2.9% figure and multiple-testing handling are also light in the summary, which leaves the strength of the gains harder to judge. This is for groups already working on multimodal survival models who want to test a slot-based architecture on oncology data. A reader focused on representation learning or missing-data handling could extract value from the experiments. It deserves peer review because the setup is concrete and the results are on real cohorts, even though the event-correspondence claim will need stronger support in revision.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes SlotSPE, a slot attention-based framework for structural prognostic event modeling in multimodal cancer survival analysis. It compresses high-dimensional histology images and gene profiles into compact, modality-specific sets of mutually distinctive slots that serve as encodings of sparse, patient-specific, unannotated prognostic events (e.g., spatial histologic patterns or pathway co-activations). These slots enable efficient intra- and inter-modal interaction modeling and incorporation of biological priors. Experiments on ten cancer cohorts report outperformance versus existing methods in 8/10 cases with an overall 2.9% improvement, robustness to missing genomic data, and improved interpretability via structured event decomposition.

Significance. If the slots can be validated as capturing biologically meaningful prognostic events rather than generic compressed features, the framework offers a principled way to decompose multimodal inputs for both predictive performance and interpretability in survival analysis. The scale of evaluation across ten cohorts and the reported robustness to missing data are strengths that would support practical adoption if the central correspondence claim is substantiated.

major comments (2)

[Abstract / Results] Abstract and Results: The claim of outperforming existing methods in 8 out of 10 cohorts with an 'overall improvement of 2.9%' provides no definition of the aggregate metric (e.g., mean C-index difference), no statistical significance tests, no error bars or confidence intervals, and no mention of multiple-testing correction across cohorts. This directly weakens the central empirical claim of superiority.
[Abstract / Methods] Abstract and Methods: The framework treats learned slots as encodings of 'sparse, patient-specific, and unannotated prognostic events' to justify both interaction modeling and 'markedly improved interpretability through structured event decomposition.' No quantitative validation (e.g., alignment with known histologic patterns, pathway markers, or supervised event labels) or ablation enforcing biological priors is described to establish this correspondence over generic representation learning. This assumption is load-bearing for the title, novelty, and interpretability claims.

minor comments (2)

[Methods] The exact formulation of the slot attention updates and the loss terms used to encourage mutual distinctiveness of slots should be stated explicitly with equations rather than high-level description.
[Figures] Figure captions and legends should clarify what the visualized slots represent (e.g., attention maps overlaid on histology) and how they are selected for display.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment point by point below, agreeing where revisions are needed to strengthen statistical reporting and providing clarifications on the interpretability claims while proposing targeted additions.

read point-by-point responses

Referee: [Abstract / Results] Abstract and Results: The claim of outperforming existing methods in 8 out of 10 cohorts with an 'overall improvement of 2.9%' provides no definition of the aggregate metric (e.g., mean C-index difference), no statistical significance tests, no error bars or confidence intervals, and no mention of multiple-testing correction across cohorts. This directly weakens the central empirical claim of superiority.

Authors: We agree that the aggregate metric requires explicit definition and statistical support for rigor. The reported 2.9% overall improvement is the mean C-index difference across the eight cohorts where SlotSPE outperformed baselines. In the revised manuscript, we will define this metric clearly as the average per-cohort C-index improvement, report individual cohort results with standard errors from cross-validation folds, include paired statistical significance tests (e.g., Wilcoxon signed-rank or DeLong's test), and apply Bonferroni correction for multiple comparisons. These details will be added to the Results section with updated tables and summarized concisely in the Abstract. revision: yes
Referee: [Abstract / Methods] Abstract and Methods: The framework treats learned slots as encodings of 'sparse, patient-specific, and unannotated prognostic events' to justify both interaction modeling and 'markedly improved interpretability through structured event decomposition.' No quantitative validation (e.g., alignment with known histologic patterns, pathway markers, or supervised event labels) or ablation enforcing biological priors is described to establish this correspondence over generic representation learning. This assumption is load-bearing for the title, novelty, and interpretability claims.

Authors: The slot attention design, inspired by factorial coding, produces mutually distinctive representations, and biological priors are incorporated to enhance prognostic relevance as described in the Methods. Interpretability is supported by qualitative analyses of slot specialization and survival associations in the current experiments. We acknowledge the absence of direct quantitative alignment metrics against annotated events, which would require additional labeled data not present in the public cohorts. In revision, we will add an ablation comparing performance and slot distinctiveness with versus without prior incorporation, plus expanded qualitative examples linking slots to known histologic or pathway features. This will better substantiate the structured decomposition claim while maintaining that the framework goes beyond generic compression through its event-oriented design. revision: partial

Circularity Check

0 steps flagged

No significant circularity; framework is an independent modeling proposal validated experimentally

full rationale

The paper introduces SlotSPE as a new slot-attention-based compression of multimodal inputs into representations that are then interpreted as encodings of sparse prognostic events. Performance improvements (2.9% overall on 8/10 cohorts) and robustness claims are presented as outcomes of direct benchmark experiments rather than any derivation that reduces by construction to fitted parameters or prior self-citations. No equations or steps in the provided text exhibit self-definitional loops, fitted-input-as-prediction, or load-bearing self-citation chains. The interpretability claim rests on the modeling choice itself and is therefore subject to external validation rather than internal tautology.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. The central modeling choice (mutually distinctive slots via slot attention) is presented as a design decision rather than a derived quantity.

pith-pipeline@v0.9.0 · 5533 in / 1114 out tokens · 26107 ms · 2026-05-17T02:13:09.230341+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Inspired by factorial coding (Schmidhuber, 1992; Higgins et al., 2017; Greff et al., 2020), our method compresses high-dimensional multimodal inputs into a compact set of semantic slots using a slot attention module (Locatello et al., 2020), where each slot corresponds to a latent prognostic event.
IndisputableMonolith/Foundation/BranchSelection.lean branch_selection unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Selective slot activation through a Mixture-of-Experts–style decoder that activates only the most predictive slots for each patient.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages · 4 internal anchors

[1]

Deep Variational Information Bottleneck

Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck.arXiv preprint arXiv:1612.00410,

work page internal anchor Pith review arXiv
[2]

The reactome pathway knowledgebase 2022.Nucleic acids research, 50(D1):D687–D692,

Marc Gillespie, Bijay Jassal, Ralf Stephan, Marija Milacic, Karen Rothfels, Andrea Senff-Ribeiro, Johannes Griss, Cristoffer Sevilla, Lisa Matthews, Chuqiao Gong, et al. The reactome pathway knowledgebase 2022.Nucleic acids research, 50(D1):D687–D692,

work page 2022
[3]

On the binding problem in artificial neural networks.arXiv preprint arXiv:2012.05208,

Klaus Greff, Sjoerd Van Steenkiste, and J ¨urgen Schmidhuber. On the binding problem in artificial neural networks.arXiv preprint arXiv:2012.05208,

work page arXiv 2012
[4]

Categorical Reparameterization with Gumbel-Softmax

Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax.arXiv preprint arXiv:1611.01144,

work page internal anchor Pith review Pith/arXiv arXiv
[5]

Adam: A Method for Stochastic Optimization

Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

work page internal anchor Pith review Pith/arXiv arXiv
[6]

Adaptive prototype learning for multimodal cancer survival analysis.arXiv preprint arXiv:2503.04643, 2025a

Hong Liu, Haosen Yang, Federica Eduati, Josien PW Pluim, and Mitko Veta. Adaptive prototype learning for multimodal cancer survival analysis.arXiv preprint arXiv:2503.04643, 2025a. Junzhuo Liu, Markus Eckstein, Zhixiang Wang, Friedrich Feuerhake, and Dorit Merhof. Spatial transcriptomics expression prediction from histopathology based on cross-modal mask ...

work page arXiv 2025
[7]

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

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

work page arXiv
[8]

The information bottleneck method

Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method.arXiv preprint physics/0004057,

work page internal anchor Pith review Pith/arXiv arXiv
[9]

Prediction of recurrence risk in endometrial cancer with multimodal deep learning.Nature medicine, 30(7): 1962–1973,

Sarah V olinsky-Fremond, Nanda Horeweg, Sonali Andani, Jurriaan Barkey Wolf, Maxime W La- farge, Cor D de Kroon, Gitte Ørtoft, Estrid Høgdall, Jouke Dijkstra, Jan J Jobsen, et al. Prediction of recurrence risk in endometrial cancer with multimodal deep learning.Nature medicine, 30(7): 1962–1973,

work page 1962
[10]

Adamhf: Adaptive multimodal hierarchical fusion for survival prediction.arXiv preprint arXiv:2503.21124,

Shuaiyu Zhang, Xun Lin, Rongxiang Zhang, Yu Bai, Yong Xu, Tao Tan, Xunbin Zheng, and Zitong Yu. Adamhf: Adaptive multimodal hierarchical fusion for survival prediction.arXiv preprint arXiv:2503.21124,

work page arXiv
[11]

Prototypical informa- tion bottlenecking and disentangling for multimodal cancer survival prediction.arXiv preprint arXiv:2401.01646,

Yilan Zhang, Yingxue Xu, Jianqi Chen, Fengying Xie, and Hao Chen. Prototypical informa- tion bottlenecking and disentangling for multimodal cancer survival prediction.arXiv preprint arXiv:2401.01646,

work page arXiv
[12]

19 B.2 Details of Selective Slot Activation

17 APPENDIXTABLE OFCONTENTS A The Use of Large Language Models (LLMs) 19 B Method Details 19 B.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 B.2 Details of Selective Slot Activation . . . . . . . . . . . . . . . . . . . . . . . . . . 19 B.3 Details of Reconstructions . . . . . . . . . . . . . . . . . . . . . . ...

work page 2024
[13]

(8) B.2 DETAILS OFSELECTIVESLOTACTIVATION To achieve sparse yet differentiable slot selection, we adopt the Gumbel-Top-K(Gumbel, 1954; Maddison et al., 2014; Kool et al.,

are: h(i) t =P(T=t|T≥t,z (i)),S (i) t = tY k=1 1−h (i) k , Lsurv({z(i), t(i), c(i)}ND i=1) =− NDX i=1 h c(i) logS (i) t(i) + (1−c (i)) logS (i) t(i)−1 + (1−c (i)) logh (i) t(i) i . (8) B.2 DETAILS OFSELECTIVESLOTACTIVATION To achieve sparse yet differentiable slot selection, we adopt the Gumbel-Top-K(Gumbel, 1954; Maddison et al., 2014; Kool et al.,

work page 1954
[14]

Given slot scoresr∈R S, we add i.i.d

trick combined with a Straight-Through (ST) estima- tor (Jang et al., 2016). Given slot scoresr∈R S, we add i.i.d. Gumbel noiseg k ∼Gumbel(0,1): ˜rk =r k +g k.(9) This perturbation transforms deterministic scores into a stochastic sampling process where largerrk values are more likely to be selected. To obtain a differentiable relaxation, we then apply so...

work page 2016
[15]

For genomics, slot embeddingsS g are decoded to approximate the original pathway embeddingsX g, guided by the positional embeddings Qg (Eq

as the reconstruction head. For genomics, slot embeddingsS g are decoded to approximate the original pathway embeddingsX g, guided by the positional embeddings Qg (Eq. 5), with reconstruction optimized via MSE loss: ˆXg =R(Q g,S g),L g recon =∥ ˆXg −X g∥2 2.(12) For histopathology, reconstructing entire WSIs is infea1sible due to their scale and random pa...

work page 2023
[16]

Histological data include all diagnostic WSIs, while transcriptomic profiles with DSS labels are obtained from cBioPortal

Following (Jaume et al., 2024), we predict disease-specific survival (DSS), a more precise indicator of patient status than overall survival. Histological data include all diagnostic WSIs, while transcriptomic profiles with DSS labels are obtained from cBioPortal

work page 2024
[17]

Pathway gene sets are curated from Hallmarks (Subra- manian et al., 2005; Liberzon et al.,

work page 2005
[18]

Evaluation

and Reactome (Gillespie et al., 2022), with genes absent in cBioPortal removed, yielding 330 pathways. Evaluation. We adopt 5-fold cross-validation and report the concordance index (C-index) (Har- rell Jr et al.,

work page 2022
[19]

In addition, we compute the restricted mean survival time (RMST) (Irwin, 1949; Karrison, 1986)

to evaluate global differences in sur- vival distributions. In addition, we compute the restricted mean survival time (RMST) (Irwin, 1949; Karrison, 1986). RMST, defined as the area under the estimated survival curve up to a clinically meaningful truncation time (60 months in our experiments), provides an interpretable summary of the average survival time...

work page 1949
[20]

In addition, we also report results using 8 slots per modality andT= 3iterations in slot attention, with further ablations on the number of slots provided in Appendix F.3.1 and on the number of iterations in Appendix F.3.2. The results demonstrate that even with a very small number of slots for both histopathology and genomics, SlotSPE maintains strong pe...

work page 2005
[21]

For completeness, we also include two strong methods (MOTCAT and CMTA)

While most baselines are not explicitly designed to handle missing modalities, they can still be evaluated under missing-genomics settings by supplying a neu- tral placeholder input in place of genomic features (Zhang et al., 2025). For completeness, we also include two strong methods (MOTCAT and CMTA). The results show that when these approaches rely sol...

work page 2025
[22]

F.2 HISTOLOGYENCODER ABLATIONS To assess the robustness of SlotSPE to the choice of visual encoder, we first replace the histology en- coder with a ResNet50 (Srivastava et al., 2015; He et al.,

work page 2015
[23]

This setting tests whether SlotSPE 25 Table 8: Ablation of model components reported as C-index (mean±std) across ten cancer datasets

pretrained only on ImageNet (Deng et al., 2009), rather than a pathology-specific foundation model. This setting tests whether SlotSPE 25 Table 8: Ablation of model components reported as C-index (mean±std) across ten cancer datasets. Best and second-best results are inboldand underline . Variants BRCA (N=1046) COADREAD (N=573) KIRC (N=488) UCEC (N=488) L...

work page 2009
[24]

As summarized in Table 9, SlotSPE continues to outperform all baselines under this weaker configuration, suggesting that its performance is not tightly coupled to a specialized encoder. We further extend this analysis by examining SlotSPE with recent pathology foundation models, includ- ing CONCH (Lu et al., 2024), CONCH v1.5 (used in TITAN (Ding et al., ...

work page 2024
[25]

LD-CV AE attains the second-best performance, yet its reliance on a variational autoencoder introduces sub- stantial memory and runtime overhead

as an approximation to full self-attention, but its capacity remains insufficient to capture the sparse, patient-specific prognostic signals. LD-CV AE attains the second-best performance, yet its reliance on a variational autoencoder introduces sub- stantial memory and runtime overhead. G.2 TRAININGRUNTIME ANDMEMORY We also compare training-time runtime a...

work page 2019
[26]

treat-all

using both clinical covariates and our model’s predicted risk scores. For each cohort, we merged the risk prediction of SlotSPE with these clinical variables on a per-patient basis. Importantly, the risk scores were obtained from the validation folds, ensuring that the model had no access to these samples during training. We then fitted Cox models across ...

work page 2000