Pith. sign in

REVIEW 2 major objections 38 references

iLENS uses a language model to route Alzheimer's patients to specialized survival experts, matching risk prediction while improving subtype separation and explaining the routing in clinical language.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-13 07:35 UTC pith:JN5M347X

load-bearing objection Solid engineering synthesis of LLM routing + MoE survival clustering with real interpretability gains; the Table 1 ranking is real but rests on a single split, so treat the LogRank wins as provisional. the 2 major comments →

arxiv 2607.08778 v1 pith:JN5M347X submitted 2026-06-12 cs.LG cs.AI

iLENS: Interpretable LLM-Guided Mixture-of-Experts for Neuroimaging Survival Analysis

Farica Zhuang , Seong Woo Han , Zixuan Wen , Shu Yang , Yize Zhao , Li Shen This is my paper
classification cs.LG cs.AI
keywords survival analysismixture-of-expertslarge language modelsAlzheimer's diseaseneuroimaginginterpretabilitypatient subtypingMCI-to-AD conversion
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Survival models for Alzheimer's conversion usually output a risk score from a single shared encoder and give little readable reason for how patients were grouped. This paper argues that a large language model can act as the router inside a mixture-of-experts survival model: it first defines a small set of clinically separable phenotypes from imaging feature names and clinical-note fields, then assigns each patient sparse weights to those experts from mesoscale brain summaries plus medical history. The weighted expert representation feeds a latent mixture of Weibull survival curves that both predicts individual risk and discovers low- versus high-risk subtypes. On ADNI MCI-to-AD data across structural MRI, FDG PET, and amyloid PET, the method keeps competitive concordance while often lifting LogRank separation relative to prior clustering survival models, and it attaches a short biologically grounded rationale to each routing decision. A sympathetic reader cares because the same system that produces a survival curve also states, in plain language, which imaging-and-history pattern drove the assignment.

Core claim

The authors establish that LLM semantic routing over structured neuroimaging measurements and unstructured clinical notes can drive a mixture-of-experts survival model for MCI-to-AD conversion, yielding competitive C-index and improved or competitive LogRank subtype separation versus SCA, VaDeSC, NSC, and DCSM across VBM, FDG, and AV45 for two and three clusters, while producing natural-language rationales for expert assignment.

What carries the argument

LLM-guided semantic routing for mixture-of-experts survival: an LLM initializes a fixed bank of phenotype experts and, at inference, emits sparse softmax weights plus a clinical rationale from mesoscale ROI summaries and notes; those frozen weights combine expert embeddings into a representation that parameterizes a latent Weibull mixture of survival subtypes.

Load-bearing premise

The whole pipeline assumes one frozen commercial language model can map brain-region summaries and clinical notes onto four fixed, clinically separable AD phenotypes without systematic bias or hallucination that would corrupt the expert weights and the Weibull subtypes.

What would settle it

On a held-out multi-site MCI cohort, replace the LLM router with random sparse weights (or with a different LLM family) and check whether LogRank subtype separation and the agreement of routing rationales with independent clinician review both collapse relative to the reported ADNI gains; if they do, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Individual survival curves can be paired with a short, imaging-and-history grounded rationale for why a patient was sent to particular experts.
  • AD conversion subtypes can be anchored in modality-specific clinical phenotypes rather than a single shared encoder.
  • Sparse semantic routing that includes clinical notes can improve population-level subtype separation over non-semantic sparse baselines while remaining competitive on concordance.
  • The design offers a concrete baseline for interpretable survival analysis of prodromal Alzheimer's conversion.

Where Pith is reading between the lines

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

  • The same frozen-LLM-router plus Weibull-mixture pattern is a natural candidate for other heterogeneous time-to-event clinical problems where notes and structured biomarkers coexist.
  • Because routing weights are frozen, gains in interpretability may trade off against adaptive specialization that a learnable gate could still provide if paired with LLM rationales only for audit.
  • Stronger LogRank lifts on amyloid and metabolism than on structure suggest the method's value is largest when spatial biomarker patterns already have well-known clinical language counterparts.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 0 minor

Summary. The paper proposes iLENS, an LLM-guided mixture-of-experts framework for MCI-to-AD conversion survival analysis on ADNI. An LLM (GPT-4.1-mini) first defines E=4 phenotype experts from feature names and clinical narrative fields, then assigns sparse routing weights (TOP_K=2) from mesoscale ROI summaries plus clinical notes; expert MLPs process only structured imaging, and a latent mixture of K Weibull components yields individualized survival curves and soft subtype memberships. On a single 60/10/30 split of 398 MCI subjects, Table 1 reports competitive C-index and often higher LogRank versus SCA, VaDeSC, NSC and DCSM across VBM/FDG/AV45 and K=2/3; Figure 2 and Appendix ablations (Tables 8–10) support clearer subtype separation and a benefit of semantic routing. Natural-language routing rationales are provided as a second interpretability layer.

Significance. If the empirical ranking holds under stronger validation, the work would be a useful contribution at the intersection of survival clustering and clinical LLM routing: it couples competitive discrimination with improved subtype separation and human-readable, biologically framed expert assignments, addressing a real gap left by shared-encoder models such as DCSM/NSC. Strengths include explicit phenotype initialization, frozen LLM routing that avoids label leakage into the gate, modality-specific expert design, and ablations that isolate semantic versus non-semantic sparse routing. The single-split evaluation and proprietary LLM dependence currently limit how far the claim can be taken, but the architecture is a clear, falsifiable baseline for interpretable AD survival subtyping.

major comments (2)
  1. Table 1 and Appendix A.1–A.2: All C-index and LogRank numbers (and the Kaplan–Meier curves in Figure 2) come from one fixed 60/10/30 split of 398 subjects with a single Optuna run per objective. No cross-validation, bootstrap, multi-seed, or standard-error estimates are reported. LogRank is especially sensitive to the size of the high-risk arm (e.g., AV45 K=2: n=27 vs 88). Without multi-split stability the headline claim that iLENS improves subtype separation over DCSM/NSC is not yet statistically secured and could reverse under a different partition.
  2. Sections 3.1, Limitations, Ethical Considerations, and Tables 8–10: Routing weights are produced by a frozen proprietary LLM (GPT-4.1-mini) whose phenotype definitions and sparse weights are not validated against clinician agreement, inter-run stability, or alternative open models. Ablations show that semantic routing helps LogRank relative to random/uniform sparse baselines, but they inherit the same single-split design and do not quantify hallucination or bias risk that the authors themselves flag. A minimal multi-LLM or multi-seed routing stability check is needed before the “biologically grounded rationales” claim can be treated as load-bearing.

Circularity Check

1 steps flagged

No load-bearing circularity; only a minor non-forcing self-citation of the authors' prior DCSM Weibull-mixture formulation, while LLM routing and reported metrics remain independent of labels and of that citation.

specific steps
  1. self citation load bearing [Section 3.2, Eq. (3)-(4)]
    "Following the generative approach of previous clustering survival models such as DCSM (Hou et al., 2024), the survival function for a patient with representation h(x) is defined as S(t|x)=\sum_k \pi_k(x) S_k(t), … Each subtype survival function S_k(t) is modeled using a Weibull distribution …"

    The Weibull-mixture generative form is taken directly from the authors' own prior DCSM paper (overlapping co-authors Wen, Yang, Shen). The citation supplies the parametric backbone rather than an independently verified external theorem. However the citation is not load-bearing for the paper's central claims: LLM expert initialization, frozen semantic routing, and the empirical C-index/LogRank improvements over DCSM itself remain independent of that reference.

full rationale

iLENS is an empirical ML methods paper, not a first-principles derivation. Expert phenotypes are produced by an LLM from feature-name lists and task description alone (Sec. 3.1, App. A.4); routing weights are pre-computed by the same LLM from mesoscale imaging summaries plus clinical notes with explicit exclusion of survival labels, conversion status, and time-to-event (App. A.6); those weights are then frozen. Only the expert MLPs and the global Weibull parameters are trained on structured imaging features. C-index and LogRank are measured on a held-out test partition against external baselines (SCA, VaDeSC, NSC, DCSM). The single self-citation of DCSM (Hou et al. 2024, overlapping co-authors) merely adopts a previously published generative mixture-of-Weibulls construction; it does not force the LLM-routing weights, the expert specialization, or the numerical gains in Table 1. Ablations (Tables 8-10) further isolate the contribution of semantic routing. No equation reduces to its own inputs by construction, no uniqueness theorem is imported to forbid alternatives, and no fitted parameter is re-labeled a prediction. The architecture is therefore self-contained against external benchmarks; the residual self-citation is ordinary and non-circular.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 2 invented entities

The central empirical claim rests on a small set of modeling choices (fixed expert count, Weibull mixture, frozen LLM router) and on the unproven reliability of a commercial LLM as a clinical phenotype router. No new physical entities are postulated; the free parameters are architectural and optimization knobs fitted on the validation split.

free parameters (5)
  • E (number of experts) = 4
    Hard-coded to 4 by prompt; controls the capacity and specialization of the MoE bank.
  • K (latent survival subtypes) = 2 or 3
    Set to 2 or 3 for low/high-risk or moderate-granularity clustering; directly affects LogRank and C-index reporting.
  • TOP_K (sparse routing) = 2
    Fixed at 2 to enforce sparsity; changes the effective mixture.
  • Optuna-tuned LR, discount, dropout, batch size = modality- and objective-specific (e.g., LR 4.76e-4 for VBM K=2 C-index)
    Fifteen TPE trials maximize validation C-index or LogRank separately per modality and K; values listed in Table 2.
  • Weibull scale λ_k and shape ρ_k = learned per run
    Global parameters of each latent subtype, learned during training rather than fixed a priori.
axioms (3)
  • domain assumption A mixture of K Weibull distributions with soft memberships derived from an expert-weighted representation adequately models MCI-to-AD survival.
    Inherited from DCSM-style clustering survival models (Section 3.2, Eq. 3–4).
  • ad hoc to paper An off-the-shelf LLM can map mesoscale ROI summaries and free-text notes onto clinically separable AD phenotypes and produce calibrated sparse routing weights.
    Core design choice of Sections 3.1 and A.4–A.6; no independent clinical validation of the generated phenotypes is supplied.
  • standard math Softmax of LLM logits yields a valid probability distribution over experts that can be frozen during subsequent gradient training.
    Standard MoE gating construction (Eq. 1).
invented entities (2)
  • LLM-initialized phenotype experts (e.g., “Diffuse high amyloid with multiple comorbidities”) no independent evidence
    purpose: Provide semantically specialized sub-networks whose mixture yields both risk prediction and interpretable routing text.
    Generated on-the-fly by the LLM from feature names; no external clinical cohort confirms these four phenotypes as the natural partition of AD progression.
  • iLENS routing controller (LLM Strategist + LLM Router) no independent evidence
    purpose: Replace a learned neural gate with frozen semantic weights and natural-language rationales.
    Architectural novelty of the paper; existence and utility are demonstrated only inside this study.

pith-pipeline@v1.1.0-grok45 · 18509 in / 2911 out tokens · 70769 ms · 2026-07-13T07:35:04.190308+00:00 · methodology

0 comments
read the original abstract

Alzheimer's Disease (AD) is a complex neurodegenerative disorder that continues to impact millions of people worldwide. Predicting AD conversion during the prodromal stage remains critical for disease understanding and patient care. As such, survival models are widely used for AD risk prediction, yet they are typically static predictors with limited interpretability and no capacity for natural language reasoning. In this work, we propose iLENS, an interpretable large language model (LLM) guided framework based on mixture-of-experts (MoE) for survival prediction in AD conversion. Our approach uses LLM to synthesize structured neuroimaging measurements and unstructured information to guide expert routing. Our framework demonstrates competitive predictive performance and capability in patient subtyping. Furthermore, our framework provides transparent, biologically grounded rationales for its routing decisions, bridging the gap between high-performance survival analysis and interpretable clinical decision support.

Figures

Figures reproduced from arXiv: 2607.08778 by Farica Zhuang, Li Shen, Seong Woo Han, Shu Yang, Yize Zhao, Zixuan Wen.

Figure 1
Figure 1. Figure 1: Overview of iLENS. The system processes multimodal inputs of structured MRI features and clinical notes. An LLM Strategist defines the routing configuration, while an LLM Router assigns weights to specialized experts within a Mixture-of-Experts bank. This process generates an interpretable clinical reasoning rationale and a mixed Weibull distribution to produce individualized survival analysis curves for A… view at source ↗
Figure 2
Figure 2. Figure 2: Subtype Separation. Kaplan–Meier survival curves for AV45 with K = 2 across clustering methods. Blue and red curves indicate lower- and higher-risk subgroups respectively according to the mean predicted survival risk. Gray curve indicates that the test samples collapsed to a single cluster. namely the VBM, FDG, and AV45 for brain vol￾ume and biomarker measurements. For clinical notes, we extracted the init… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 8 linked inside Pith

  1. [1]

    Blood cancer journal , volume=

    A multistate model of survival prediction and event monitoring in prefibrotic myelofibrosis , author=. Blood cancer journal , volume=. 2020 , publisher=

    2020

  2. [2]

    Bmj , volume=

    A systematic review of physicians' survival predictions in terminally ill cancer patients , author=. Bmj , volume=. 2003 , publisher=

    2003

  3. [3]

    BMC bioinformatics , volume=

    Allowing for mandatory covariates in boosting estimation of sparse high-dimensional survival models , author=. BMC bioinformatics , volume=. 2008 , publisher=

    2008

  4. [4]

    Frontiers in Artificial Intelligence , volume=

    Survival prediction landscape: an in-depth systematic literature review on activities, methods, tools, diseases, and databases , author=. Frontiers in Artificial Intelligence , volume=. 2024 , publisher=

    2024

  5. [5]

    BMC medical research methodology , volume=

    DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network , author=. BMC medical research methodology , volume=. 2018 , publisher=

    2018

  6. [6]

    An R and S-PLUS companion to applied regression , volume=

    Cox proportional-hazards regression for survival data , author=. An R and S-PLUS companion to applied regression , volume=

  7. [7]

    IEEE Journal of Biomedical and Health Informatics , volume=

    Deep survival machines: Fully parametric survival regression and representation learning for censored data with competing risks , author=. IEEE Journal of Biomedical and Health Informatics , volume=. 2021 , publisher=

    2021

  8. [8]

    Medical image analysis , volume=

    Interpretable deep clustering survival machines for Alzheimer’s disease subtype discovery , author=. Medical image analysis , volume=. 2024 , publisher=

    2024

  9. [9]

    2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI) , pages=

    Deep clustering survival machines with interpretable expert distributions , author=. 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI) , pages=. 2023 , organization=

    2023

  10. [10]

    Multi-Modal Deep Clustering Survival Machines for Alzheimer's Disease Subtype Discovery , author=. Proc. of the IEEE/CVF Int. Conf. on Computer Vision , pages=

  11. [11]

    Conference on Health, Inference, and Learning , pages=

    Neural Survival Clustering: Non-parametric mixture of neural networks for survival clustering , author=. Conference on Health, Inference, and Learning , pages=. 2022 , organization=

    2022

  12. [12]

    arXiv preprint arXiv:2106.05763 , year=

    A deep variational approach to clustering survival data , author=. arXiv preprint arXiv:2106.05763 , year=

    Pith/arXiv arXiv

  13. [13]

    Proceedings of the ACM Conference on Health, Inference, and Learning , pages=

    Survival cluster analysis , author=. Proceedings of the ACM Conference on Health, Inference, and Learning , pages=

  14. [14]

    Journal of Machine Learning Research , volume=

    Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity , author=. Journal of Machine Learning Research , volume=

  15. [15]

    arXiv preprint arXiv:2505.19190 , year=

    I2moe: Interpretable multimodal interaction-aware mixture-of-experts , author=. arXiv preprint arXiv:2505.19190 , year=

    Pith/arXiv arXiv

  16. [16]

    arXiv preprint arXiv:2512.10966 , year=

    Multimodal Fusion of Regional Brain Experts for Interpretable Alzheimer's Disease Diagnosis , author=. arXiv preprint arXiv:2512.10966 , year=

    Pith/arXiv arXiv

  17. [17]

    Advances in Neural Information Processing Systems , volume=

    Flex-moe: Modeling arbitrary modality combination via the flexible mixture-of-experts , author=. Advances in Neural Information Processing Systems , volume=

  18. [18]

    Information Fusion , pages=

    Integrating images and genomics for multi-modal cancer survival analysis via mixture of experts , author=. Information Fusion , pages=. 2025 , publisher=

    2025

  19. [19]

    Advances in Neural Information Processing Systems , volume=

    Uni-med: a unified medical generalist foundation model for multi-task learning via connector-MoE , author=. Advances in Neural Information Processing Systems , volume=

  20. [20]

    IEEE Transactions on Multimedia , year=

    Moe-llava: Mixture of experts for large vision-language models , author=. IEEE Transactions on Multimedia , year=

  21. [21]

    arXiv preprint arXiv:2401.04088 , year=

    Mixtral of experts , author=. arXiv preprint arXiv:2401.04088 , year=

    Pith/arXiv arXiv

  22. [22]

    Advances in neural information processing systems , volume=

    Mome: Mixture of multimodal experts for generalist multimodal large language models , author=. Advances in neural information processing systems , volume=

  23. [23]

    International Workshop on Agentic AI for Medicine , pages=

    Adagent: Llm agent for alzheimer’s disease analysis with collaborative coordinator , author=. International Workshop on Agentic AI for Medicine , pages=. 2025 , organization=

    2025

  24. [24]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    Medagents: Large language models as collaborators for zero-shot medical reasoning , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    2024

  25. [25]

    arXiv preprint arXiv:2503.18968 , year=

    Medagent-pro: Towards evidence-based multi-modal medical diagnosis via reasoning agentic workflow , author=. arXiv preprint arXiv:2503.18968 , year=

    Pith/arXiv arXiv

  26. [26]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    Mmedagent: Learning to use medical tools with multi-modal agent , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    2024

  27. [27]

    Advances in Neural Information Processing Systems , volume=

    Mdagents: An adaptive collaboration of llms for medical decision-making , author=. Advances in Neural Information Processing Systems , volume=

  28. [28]

    arXiv preprint arXiv:2406.18665 , year=

    Routellm: Learning to route llms with preference data , author=. arXiv preprint arXiv:2406.18665 , year=

    Pith/arXiv arXiv

  29. [29]

    arXiv preprint arXiv:2501.09636 , year=

    LLM-Based Routing in Mixture of Experts: A Novel Framework for Trading , author=. arXiv preprint arXiv:2501.09636 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2403.12031 , year=

    Routerbench: A benchmark for multi-llm routing system , author=. arXiv preprint arXiv:2403.12031 , year=

    Pith/arXiv arXiv

  31. [31]

    Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

    Optuna: A next-generation hyperparameter optimization framework , author=. Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

  32. [32]

    Neurology , volume=

    Alzheimer's disease Neuroimaging Initiative (ADNI) clinical characterization , author=. Neurology , volume=. 2010 , publisher=

    2010

  33. [33]

    Neuroimage , volume=

    Automated anatomical labelling atlas 3 , author=. Neuroimage , volume=. 2020 , publisher=

    2020

  34. [34]

    Neuroimage , volume=

    Voxel-based morphometry—the methods , author=. Neuroimage , volume=. 2000 , publisher=

    2000

  35. [35]

    Brain communications , volume=

    Prediction of conversion to Alzheimer’s disease using deep survival analysis of MRI images , author=. Brain communications , volume=. 2020 , publisher=

    2020

  36. [36]

    Alzheimer's & dementia , volume=

    Toward defining the preclinical stages of Alzheimer’s disease: Recommendations from the National Institute on Aging-Alzheimer's Association workgroups on diagnostic guidelines for Alzheimer's disease , author=. Alzheimer's & dementia , volume=. 2011 , publisher=

    2011

  37. [37]

    Neurobiology of aging , volume=

    Predicting time-to-conversion for dementia of Alzheimer's type using multi-modal deep survival analysis , author=. Neurobiology of aging , volume=. 2023 , publisher=

    2023

  38. [38]

    Alzheimer's & Dementia: Diagnosis, Assessment & Disease Monitoring , volume=

    Characterizing heterogeneity in the progression of Alzheimer's disease using longitudinal clinical and neuroimaging biomarkers , author=. Alzheimer's & Dementia: Diagnosis, Assessment & Disease Monitoring , volume=. 2018 , publisher=

    2018