List of Papers Browse by Subject Areas Author List
Abstract
Accurate survival prediction is essential for personalized treatment planning in head and neck cancer, yet remains challenging due to the heterogeneous and high-dimensional nature of multimodal clinical data. While deep survival models have improved predictive performance over classical statistical approaches, existing methods typically rely on static fusion strategies or temporally agnostic modeling, limiting their ability to capture structured clinical workflows.
In this work, we propose ChronoSurv, a heterogeneous hierarchical directed graph framework for multimodal survival analysis. ChronoSurv represents patient care as a progression-aware clinical trajectory using directed graphs aligned with key diagnostic steps. A hierarchical topology incorporates fine-grained, coarse, and global representations, further supporting flexible adaptation to missing modalities, while heterogeneous message passing models complex and asymmetric relationships across modalities and clinical steps.
Experimental results on two public datasets demonstrate that ChronoSurv achieves state-of-the-art discriminative performance while maintaining statistically reliable calibration. Comprehensive ablation studies further confirm the contribution of each architectural component, highlighting the potential of trajectory-aware graph modeling for multimodal survival prediction.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3722_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/MICS-Lab/ChronoSurv
Link to the Dataset(s)
HANCOCK: https://www.hancock.research.uni-erlangen.org/download
TCGA: https://portal.gdc.cancer.gov/
BibTex
@InProceedings{MicHug_ChronoSurv_MICCAI2026,
author = { Miccinilli, Hugo AND Di Piazza, Theo},
title = { { ChronoSurv: A Clinical Pathway-Guided Graph Framework for Multimodal Survival Analysis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16887},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
- This paper proposes a novel multimodal survival analysis model, ChronoSurv, which incorporates progression-aware clinical trajectories using a hierarchical directed graph.
- The main contributions of ChronoSurv include: (1) enabling progression-aware multimodal survival prediction, (2) supporting adaptable topology for robustness to missing modalities, and (3) improving performance over multiple benchmark methods on public head and neck cancer datasets.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- ChronoSurv introduces a novel clinical pathway-guided multimodal fusion strategy by explicitly modeling patient trajectories from patient background to cancer diagnosis and surgical intervention.
- The experimental setup is well designed, including a wide range of conventional non-deep learning and deep learning survival analysis methods, as well as state-of-the-art multimodal deep learning models, evaluated across multiple public head and neck cancer datasets.
- ChronoSurv enables progression-wise interpretation of the clinical trajectory, thereby allowing the model to capture stage-specific contributions of different modalities to survival prediction.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- This work aims to represent a progression-aware clinical trajectory using a hierarchical directed graph. However, the implementation details of this trajectory are insufficiently specified in the main text. In Section 2.2, the authors introduce three key clinical steps—(1) patient background, (2) initial cancer diagnosis, and (3) local surgery—but do not explicitly specify how individual sub-modalities (e. g. , clinical, history, pathology) are assigned to these steps. Clarifying how sub-modalities are assigned to each clinical step would improve the reproducibility of the proposed framework.
- The performance improvements reported in Table 2 appear to be relatively marginal compared to the benchmark methods. In particular, given that the evaluation is based on a single 5-fold cross-validation and small dataset sizes, it is difficult to consider the reproducible results. Providing additional evaluation, such as reporting statistical significance (e. g. , Wilcoxon rank-sum test), would strengthen the validity of the results.
- There is no mention or description of Fig. 2(c) (Example of prediction) in the main text.
- Several details necessary for reproducibility are missing:
- The architecture of the survival head is not described.
- In the component-wise ablation study, it is unclear how the model aggregates the Level 2 step nodes in the w/o level 3 setting. While it appears that a simple pooling operation may have been used, this is not explicitly stated. Such details are important for ensuring reproducibility.
- Corrections:
- In Table 2, the best C-index for the TCGA-HNSC dataset is achieved by HFBSurv (. 660).
- In Table 3, the best C-index for the TCGA-HNSC dataset is achieved by Cross-attention (. 656).
- In Table 4, the best IBS for the TCGA-HNSC dataset is achieved by the w/o heterogeneity setting (. 179).
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper proposes a novel clinical pathway-guided multimodal fusion framework based on a hierarchical directed graph, which enables explicit modeling of progression-aware patient trajectories. The proposed approach is computationally innovative and promising, not only for head and neck cancer survival prediction but also for broader applications across diverse malignancies, as relevant annotations become available in other cancer types. The experimental evaluation is comprehensive, including comparisons with state-of-the-art multimodal survival analysis models, alternative multimodal aggregation schemes, as well as detailed ablation studies. However, several issues remain regarding reproducibility and minor corrections. In addition, further statistical validation would be beneficial to support the validity of the reported results. Therefore, I recommend a weak accept.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have addressed several of the major concerns raised in the original review, and important points are now clarified. Overall, I believe the paper presents a potentially valuable framework for progression-aware multimodal survival analysis, and the rebuttal addressed most of the key concerns.
The following points have been clarified in the rebuttal:
- Clinical-step assignment
- Reproducibility details (survival head and ablation setting)
- Clarification of Fig. 2(c) and survival prediction formulation
- Minor table corrections
However, some concerns still remain:
- Although the authors acknowledge the limited cohort size and lack of statistical validation, additional statistical significance analysis would further strengthen the reliability of the reported performance gains.
Review #2
- Please describe the contribution of the paper
The paper proposed CronoSurv, a model for survival prediction based on multi-modal inputs (tabular/clinical, text, image, …) combined with multi-modal fusion based on a (directed) graph neural network with multiple hierarchies with cross-level message passing. They utilize the framework to predict discrete hazard rates for time-to-event data with feature-conditional independent right-censoring.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The authors tackle an important problem, namely the use of multiple modalities, tacking into account the temporal processes and dependencies characteristic for time-to-event data. The use of (directed) GNNs is conceptually useful in this context. The design of the model and its application follow current developments in the respective areas (i.e. heterogeneous convolutions rather than late fusion, etc.). Patient-specific graph topology is a good idea. The article is well written, easy to follow and for the most part technically sound. Results are presented clearly and the ablation study is well done.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
There are two main limitations/concerns for this paper: 1) As is typical for time-to-event data (especially for the type of analysis pursued here, i.e. multi-modal data with images) the availability of data is a recurring problem: There are few suitable data sets and the ones available have small sample sizes, and worse, small number of events, 763 * .72(censoring) * .15 (test set) = 82 events in the test set for the big data set, which hinders generalization. The results table is also a bit misleading. UniHNC looks like a separate data point but is in fact a weighted average of the other two, with the HANCOCK data being the bigger one and the only one where the model outperforms SurvPCG. So the overall evidence for improvement is not too overwhelming. Altough I should stress that I don’t care about showing that the model outperforms competitors. Null-results (more or less) are fine, but the presentation and honest result description is important.
2) Level two clinical steps seem to be an important driver of the performance, but this step is not explained clearly and is crucial to understand temporal dependencies and the nature of the problem. a) Why do some level one parameters connect to step 2, but not step 1 and 3 and vice versa? is this based on how these are applied accross time? b) there is also a directed progression from step1 -> step3 and step1->step2->step3. The authors need to explain the arrow of time better, i.e. at which time the respective information is available fro prediction. When does the observation start (time-origin) and at which time-points are the information about the individual steps given and what is the end-point (death?). My concern is that some of the steps are only known after baseline, for example t=0 -> start of observation, t=10 -> surgery, t=11 -> death. Using info about surgery at t=10 would imply that event cannot have happened beforehand (immortal time bias). Using such time-dependent features is a totally different task then predicting from baseline (without taking future information into account) and the method then should be compared to models with same amount of information (e.g. time-dependent Cox, dynamic DeepHit, …). Authors should clarify if that is the case here and reformulate to make clear that this is not the case otherwise.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
In Figure 2c) , please explain how you obtain time-to-event predictions. the model returns discrete hazards that can be used to estimate S(t). But in presence of censoring (50-70% in your data), the survival curve is improper. Estimating the expected value however requires integration over the survival function. Its not clear what the Lymph node/Primary tumor images have to to with the rest of the image). Panel c) of Figure 2 is not mentioned in 3.3.at all. Please provide more details regarding the calculation of C-index. Not sure how you obtained a one-number summary from the prediction of multiple hazards and which C-index was used (see C-hacking (https://academic.oup.com/bioinformatics/article/38/17/4178/6640155) and C-index multiverse papers (https://arxiv.org/abs/2508.14821)) You could make clearer that you are essentially using discrete-time-to-event analysis and reference related literature better (e.g. Gensheimer & Narasimhan (2019), https://peerj.com/articles/6257.pdf)
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Paper is well written and motivated. The general idea and techniques used are sound. Results are mostly presented clearly, which makes me lean towards accept. The presentation of main results in Table 2 is somewhat misleading (but could be fixed). My main concern is about the arrow of time and the models playing on level playing field (immortal time, time-dependent features). If the concern is justified, paper would need to be rejected in my opinion. Also, taken into account some inevitable researcher degrees of freedom and 2 small data sets, the generalization of the results is somewhat questionable (although I would weigh this point less strongly given the limitations regarding data availability in the field). The authors should acknowledge this limitation however.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
As acknowledged by the authors, the surgery data is from after the baseline where follow-up started (therefore arrow of time is violated). While the average time until surgery is short compared to entire follow-up, it could still leak information about immortal time between diagnosis and surgery (e.g. in surgery report/description). This is not clear atm. While authors claim all methods received the same info, but ChronoSurv might have specific advantages from that, and its not clear how exactly all models got the same info (e.g. how text data was used in Cox model) and the linked repo only contains ChronoSurv, not the full setup with all models. Also, some of the conclusions in the paper are to some extent compromised by the wrong reporting of best models, e.g. ablation study (as pointed out by other reviewers).
Review #3
- Please describe the contribution of the paper
The paper proposes ChronoSurv, a multimodal survival prediction framework that models each patient as a hierarchical directed heterogeneous graph aligned with the clinical care pathway. The approach encodes multimodal inputs (WSI, tabular, text) as leaf nodes, organizes them into clinically meaningful temporal steps, and aggregates them through multi-level heterogeneous message passing to obtain a patient-level representation for discrete-time survival prediction. A key aspect is the progression-aware graph topology, which explicitly captures temporal dependencies in clinical workflows while naturally handling missing modalities through adaptive graph construction.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Clinically grounded modeling: The explicit alignment of the graph structure with the clinical care pathway (background → diagnosis → surgery) introduces domain knowledge into the model design, improving interpretability and plausibility. Principled handling of missing modalities: Missing data are addressed by removing nodes and edges from the graph, avoiding imputation or masking heuristics and yielding a clean and adaptive formulation. Hierarchical representation learning: The multi-level graph enables structured aggregation from fine-grained to global representations, which is well-supported by ablation results. Heterogeneous message passing: Relation-specific transformations allow modeling asymmetric and semantically different interactions (e.g., modality-to-step vs. temporal dependencies). Comprehensive experimental validation: Evaluation includes multiple datasets (HANCOCK, TCGA-HNSC, UniHNC), cross-validation, and comparison with classical, deep, and multimodal baselines. Calibration analysis: Inclusion of calibration metrics and D-calibration testing strengthens the evaluation beyond standard discrimination metrics. Detailed ablation study: Systematic removal of graph components (steps, hierarchy, directionality, heterogeneity) provides strong evidence of the contribution of each design choice.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Hand-crafted graph topology: The clinical pathway (three steps) is predefined and fixed, which may limit scalability and adaptability to other diseases or clinical workflows. Simplified temporal modeling: The clinical trajectory is reduced to three discrete steps, which may not capture the complexity and variability of real-world patient pathways. Limited modality diversity: Although multimodal, the evaluation is restricted to specific combinations (WSI, tabular, text), with no experiments on broader imaging modalities (e.g., CT, MRI) or more diverse clinical settings. No analysis of computational complexity: The cost of graph construction and heterogeneous message passing is not discussed, especially compared to simpler fusion baselines. Limited treatment of missing modalities: Although the method adapts the graph topology by removing nodes corresponding to missing modalities, it does not explicitly model missingness mechanisms (e.g., MCAR/MAR/MNAR) nor evaluate robustness under different missing data patterns. This limits the understanding of its behavior in realistic clinical scenarios where missingness is often structured and non-random.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
Add analysis of different missingness patterns, particularly non-random missing modalities. Include computational cost analysis (training/inference complexity vs. baselines). Clarify the rationale behind the chosen clinical steps and discuss how the framework adapts to different workflows.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a well-motivated and clinically relevant approach to multimodal survival prediction, introducing a structured graph-based formulation aligned with clinical workflows. The hierarchical and progression-aware design is intuitive and supported by solid empirical results and ablation studies. Limitations in generalization, evaluation breadth, and analysis of missing data reduce the overall impact. The work is technically sound and valuable.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After reading the rebuttal, I maintain my positive assessment of the paper and support acceptance.
The authors successfully addressed the most important concern regarding the “arrow of time” and the potential risk of immortal time bias. The rebuttal clearly explains that ChronoSurv is formulated as a static prognosis model using information available within the initial clinical workup, rather than a longitudinal or dynamic survival prediction framework. Importantly, the authors clarify that all compared methods receive the exact same input features and preprocessing pipeline, and that no follow-up or future-event information is used as model input. The directed graph structure therefore models the clinical order of information acquisition rather than temporal evolution after baseline. This clarification substantially strengthens the validity and fairness of the experimental comparison.
The rebuttal also improves the reproducibility and interpretability of the work by explicitly detailing the assignment of sub-modalities to clinical steps, the survival head architecture, the aggregation strategy used in the ablation settings, and the computation of survival predictions and risk scores. In addition, the clarification regarding UniHNC as a combined multi-cohort benchmark rather than an independent dataset improves transparency of the evaluation protocol.
The paper still has some limitations. The clinical workflow graph remains manually designed, the temporal modeling is simplified into three coarse stages, and the datasets are relatively small with limited event counts, which may affect generalization. Furthermore, missingness mechanisms are not explicitly modeled. However, these limitations are acknowledged by the authors and are reasonable given the current availability of multimodal survival datasets in head and neck cancer.
Overall, I find the work technically sound, clinically motivated, and methodologically well justified. The progression-aware graph formulation is interesting, the experiments are comprehensive, and the rebuttal successfully resolved the major conceptual concerns.
Author Feedback
We sincerely thank the Reviewers [R] and the Meta-Reviewer [MR] for their valuable feedback. We address the main concerns below: [R2, MR] Arrow of time / fair comparison. We agree that the prediction setting must be clarified. ChronoSurv is not intended as a first-consultation model. It is a static prognosis model using information available from the index clinical workup. Importantly, ChronoSurv and all 11 competing methods in Table 2 receive the identical input feature set under the same preprocessing. In both HANCOCK and TCGA-HNSC, t=0 is the date of initial pathologic diagnosis, following each benchmark’s standard survival convention. Diagnostic workup-derived variables (e.g., primary tumor FFPE WSIs typically acquired at resection) enter as static workup observations rather than time-dependent covariates: a standard practice in multimodal oncologic survival analysis, mirroring the canonical treatment of post-surgical pTNM as static despite its post-resection acquisition. Surgery typically occurs soon after diagnosis (mean ~36 days in UniHNC vs. ~1215 days mean time to last information), making the workup window short relative to the observation period. This workup-window effect therefore reflects a limitation of the standard retrospective benchmark formulation, shared identically by all models, rather than future information uniquely exploited by ChronoSurv. Furthermore, the directed edges in ChronoSurv encode the clinical order of feature acquisition during workup, not follow-up evolution. No timestamp info (e.g., surgery date) is used as model input. Time-dependent follow-up features (e.g., recurrence, progression) are deliberately excluded. Extension to dynamic graphs for longitudinal survival analysis, with comparison against time-dependent models (e.g., time-dependent Cox), is left for future work. [R1, R2, R3] Clinical-step assignment. Sub-modality -> step mapping is deterministic (Fig. 1b). Step 1: Background (age, sex, smoking status). Step 2: Clinical (e.g., primary site, metastasis at diagnosis), Blood (38 analytes), History (text). Step 3: Pathology (e.g., post-surgical pTNM, grade, invasion, margins), WSIs (primary tumor + lymph-node), Surgery (reports + description). Each leaf is embedded (Sec. 2.1) and linked by directed edges (Sec. 2.2). Topology may be adapted to other diseases, but our contribution is the framework: a generalizable clinical pathway-guided survival graph. [R1, R2, MR] Data limitations & statistical validation. We acknowledge the limited cohort sizes and event counts in public HNC benchmarks. UniHNC is HANCOCK+TCGA-HNSC, not an independent dataset (will be clarified in Sec. 3). Despite these limitations, ChronoSurv achieves competitive performance across folds with per-fold std reported in Table 2: best-or-tied IBS, competitive C-index values, and consistent D-Calibration test (5/5 folds vs. SurvPCG 1/5). [R3, MR] Computational complexity. End-to-end inference (graph construction + heterogeneous message passing) takes ~30 ms/sample on an RTX 2080 Ti, enabling clinical deployment. The GNN forward pass adds negligible cost on top of the (frozen) feature extractors shared with baselines. [R1, R2] Survival predictions + risk score. Survival from hazards: S(t_j|X) = prod_{k=1..j} (1 - lambda(t_k|X)). The Fig. 2c estimate is the restricted mean T_hat = sum_{t=0..T_max} S_hat(t). Per-patient risk is -T_hat, with the Uno C-index [33]. We will reference Fig. 2c in Sec. 3.3 and cite Gensheimer & Narasimhan (2019). [R3] Missing modality analysis. Missingness is heterogeneous: curation gaps or selective analytes (e.g., CRP [15]). ChronoSurv does not explicitly model missingness mechanisms (MCAR/MAR/MNAR) but handles them via adaptive topology. [R1] Reproducibility. Survival head: 3-layer MLP + sigmoid producing per-bin hazards. w/o-level-3 ablation: step-2 nodes mean-pooled before the head. Table 2-4 highlights will be corrected.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Though the reviewer accepted this paper, they still raised concerns about reproducibility, statistical validation, and generalization due to small datasets and researcher degrees of freedom. One main issue is potential bias from the arrow of time and unfair model comparisons (e.g., immortal time, time-dependent features), which would warrant rejection if justified; they also note a missing analysis of computational complexity.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After reviewing the comments and the authors’ responses, I find that the main concerns have been addressed. Overall, the paper is in good shape for acceptance. I recommend accepting the submission, while encouraging the authors to incorporate the reviewers’ suggestions carefully when preparing the final version.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
This paper proposes a clinical pathway-guided graph framework for multimodal survival analysis. Reviewers appreciate the clinically motivated graph design. However, after rebuttal, key concern remains regarding the arrow of time in the survival prediction setting and whether post-baseline information may introduce bias or give the proposed model an unfair advantage. The rebuttal clarifies the intended static prognosis setting and addresses several reproducibility details, but it does not fully resolve the concern about temporal validity and fair comparison, especially given the small datasets and limited statistical validation.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The main concern is the unresolved arrow-of-time issue: some surgery/pathology features appear to be acquired after the baseline time origin, which may introduce immortal time bias or future-information leakage. This directly affects the validity of the survival prediction setting and the fairness of comparison. Although the framework is clinically interesting, the rebuttal does not fully eliminate this concern, and the evidence is not sufficient for acceptance.