Multi-Modal Tumor Survival Prediction via Graph-Guided Mixture of Experts
cs.LG · cs.MA
Quick verdict
ASU MS thesis: graph-guided MoE ensemble beats naive averaging on TCGA-LUAD (CI 0.703 vs 0.665), but evidence is thin: no significance testing, nonstandard c-index, no code.
Summary
This ASU master's thesis presents a graph-guided mixture-of-experts (MoE) framework for multimodal survival prediction in lung adenocarcinoma. Four existing models act as experts: Lasso on clinical data, HGCN on all modalities, and CTransPath and PSSLM on tissue slides; a graph-attention gating network computes softmax weights over the four predictions, which are combined by weighted averaging. The design is motivated by prompting experiments in which GPT-3.5 selected tools suboptimally in four qualitative examples. Evaluation uses 470 TCGA-LUAD cases with clinical, genomic, and slide data, the concordance index, and Kaplan-Meier stratification, in a stated 5-fold procedure whose protocol remains under-specified. The MoE reaches a c-index of 0.703±0.022 versus 0.665±0.020 for naive ensembling and 0.611–0.650 for individual models; ablations (Fig. 5.4b) show each expert contributes, with CI falling to 0.592 without Lasso. Limitations: the gate performs learnable weighted averaging, not discrete model selection; no significance test supports the ensemble gain; the c-index formula ignores censored pairs; no external cohort and no code release are reported.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A graph-attention gating network over four frozen expert models (Lasso, HGCN, CTransPath, PSSLM) improves TCGA-LUAD survival-prediction concordance to 0.703±0.022, above a naive ensemble (0.665) and every individual expert (0.611–0.650) (Table 5.1).
- GPT-3.5 fails to reliably route multimodal inputs to the correct expert tool in prompting experiments (§3.1), motivating a learned gate instead.
- Each expert contributes: removing Lasso, PSSLM, or HGCN costs more than removing CTransPath (CI 0.592/0.612/0.683 vs 0.683; Fig. 5.4b), and activation frequencies favor Lasso and PSSLM (Fig. 5.4a).
- The framework is 'flexible, capable of adding more experts or removing experts' due to the inductive gating model (§5.2), and MoE-based risk stratification achieves SOTA KM separation on LUAD (§5.1, §6.1).
Strengths
Pragmatic expert portfolio covers every input modality
The four experts (§3.3.2) span clinical data (Lasso), fused clinical+genomic+slide data (HGCN, Hou et al. 2023), and pathology slides (CTransPath, PSSLM), so the ensemble never depends on one modality being present.
Learned weighting clearly beats naive averaging
Table 5.1: Dense MoE 0.703±0.022 vs Naive ensemble 0.665±0.020 vs best individual (Lasso) 0.650±0.017 — the largest step in the table is gate vs. naive combination.
Honest negative motivation for LLM-based routing
Ablation and activation-frequency evidence is internally coherent
Weaknesses
'Graph-guided MoE' is a learnable weighted ensemble
No statistical support for the headline claim
Censoring mishandled in evaluation and training
'SOTA' claim lacks external baselines
Reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Review: Multi-Modal Tumor Survival Prediction via Graph-Guided Mixture of Experts (arXiv 2609.14072)
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does the graph-guided Mixture of Experts paper (arXiv 2609.14072) do?
It is a 2024 Arizona State University master's thesis (posted to arXiv in 2026) on predicting lung adenocarcinoma patient survival from clinical data, genomics, and pathology slides. It combines four existing prediction models — Lasso, HGCN, CTransPath, and PSSLM — whose outputs are weighted by a small graph-attention gating network. On 470 TCGA-LUAD cases it reports a c-index of 0.703 versus 0.665 for simple score averaging.
Is this a real Mixture of Experts model like GPT-4's MoE layers?
No. All four expert models run on every patient and their risk predictions are combined through softmax weights computed by a gating network — a learnable weighted ensemble. There is no sparse routing or conditional computation as in transformer MoE layers, and no LLM participates at inference; GPT-3.5 is only used in a motivating prompting experiment that failed to pick tools.
What results does the thesis report on TCGA-LUAD?
Concordance index (mean±SD over a 5-fold procedure): the proposed MoE 0.703±0.022, naive ensemble 0.665±0.020, and individual models Lasso 0.650, HGCN 0.633, PSSLM 0.642, CTransPath 0.611. Kaplan-Meier curves for MoE-defined high/low-risk groups are separated with a reported logrank p-value of about 5×10⁻¹⁹, though the risk-group threshold is not specified.
Is the code or the TCGA-LUAD dataset available?
The thesis mentions no code release. The underlying TCGA-LUAD data (clinical, genomic, and diagnostic slides) is publicly available from the Genomic Data Commons, but the paper's slide preprocessing and expert training details are too sparse to guarantee exact replication.
What are the main criticisms of this work?
Reviewers would note: no statistical test for the improvement over the naive ensemble; a c-index computed only on uncensored pairs (nonstandard); the 'dynamic expert selection' claim conflicts with dense softmax gating; no comparison to published multimodal survival methods; inconsistent dataset counts (522/585/470); and no released code.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- GPT-3.5 prompting examples (Fig. 3.1–3.2 panels) and KM curves are referenced but their content is not analyzable from the extracted text.
- Figures 5.1–5.4 are read from embedded axis labels and caption text only; exact KM group sizes and curves could not be inspected.
- The document is a master's thesis deposited on arXiv, not a peer-reviewed conference or journal paper.
- Some arithmetic inconsistencies in the source (522/585/470 patient counts; km axis units) could not be resolved from the text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.