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

# Interpretable LLM-Guided Mixture-of-Experts for Neuroimaging Survival Analysis
Source: [https://arxiv.org/html/2607.08778](https://arxiv.org/html/2607.08778)
Farica Zhuang1,Seong Woo Han1,Zixuan Wen1,Shu Yang1, Yize Zhao2,Li Shen1, 1University of Pennsylvania,2Yale University

###### 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\.

iLENS:InterpretableLLM\-Guided Mixture\-of\-Experts forNeuroimagingSurvival Analysis

Farica Zhuang1, Seong Woo Han1, Zixuan Wen1, Shu Yang1,Yize Zhao2,Li Shen1,1University of Pennsylvania,2Yale University

## 1Introduction

Survival analysis, or time\-to\-event prediction, is an important component of precision medicine\. In neurodegenerative disorders, such as Alzheimer’s Disease \(AD\), predicting risk scores for conversion to AD and identifying specific risk profiles are essential for early intervention and resource allocationSperlinget al\.\([2011](https://arxiv.org/html/2607.08778#bib.bib36)\); Nakagawaet al\.\([2020](https://arxiv.org/html/2607.08778#bib.bib35)\); Mirabnahrazamet al\.\([2023](https://arxiv.org/html/2607.08778#bib.bib37)\)\. Classical survival methods and modern deep learning models have improved predictive accuracy, yet they remain primarily as predictors for numerical outcome risk scoresFox and Weisberg \([2002](https://arxiv.org/html/2607.08778#bib.bib6)\); Nagpalet al\.\([2021](https://arxiv.org/html/2607.08778#bib.bib7)\); Katzmanet al\.\([2018](https://arxiv.org/html/2607.08778#bib.bib5)\)\. In clinical settings, however, the ability to identify patient subgroups or clusters based on risk profiles is valuable, as such subtypes can reveal heterogeneous disease progression patterns and support more interpretable prognostic decision\-making\. This is particularly important in AD, where patients often exhibit diverse trajectories of neurodegeneration and biomarker burdenGoyalet al\.\([2018](https://arxiv.org/html/2607.08778#bib.bib38)\)\. To bridge this gap, we introduce iLENS, an interpretable LLM\-guided MoE for neuroimaging survival analysis\. Moving beyond static encoders, we investigate whether language models can serve as semantic routing controllers within an expert\-based survival clustering framework\. This design provides two\-layer interpretability\. First, we provide natural language interpretability based on each patient’s structured neuroimaging biomarker measurements to justify the model’s clinically\-grounded expert routing\. Second, we maintain clustering interpretability and uncover the underlying survival subtypes\. We evaluate our framework using real\-world longitudinal neuroimaging data from the Alzheimer’s Disease Neuroimaging Initiative \(ADNI\) datasetPetersen and others \([2010](https://arxiv.org/html/2607.08778#bib.bib32)\)to predict and subtype AD conversion\.

![Refer to caption](https://arxiv.org/html/2607.08778v1/x1.png)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 Alzheimer’s disease progression\.
## 2Related Work

Survival Prediction and Subtype Clustering\.Traditional survival models often focus on risk score predictions for a given outcomeFox and Weisberg \([2002](https://arxiv.org/html/2607.08778#bib.bib6)\); Katzmanet al\.\([2018](https://arxiv.org/html/2607.08778#bib.bib5)\); Nagpalet al\.\([2021](https://arxiv.org/html/2607.08778#bib.bib7)\)\. However, a crucial goal of survival prediction in medical applications is also to identify patient subgroups with similar survival patterns and risk profiles, known as subtyping \(e\.g\., high risk v\.s\. low risk\)Abbasiet al\.\([2024](https://arxiv.org/html/2607.08778#bib.bib4)\)\. This helps with treatment assignment, resource allocation, and disease understandingCarobbioet al\.\([2020](https://arxiv.org/html/2607.08778#bib.bib1)\); Glareet al\.\([2003](https://arxiv.org/html/2607.08778#bib.bib2)\); Binder and Schumacher \([2008](https://arxiv.org/html/2607.08778#bib.bib3)\)\. Hence, neural survival clustering \(NSC\) models were introduced as a new class of survival models that not only predict individualized risk, but also perform explicit subtype discoveryJeanselmeet al\.\([2022](https://arxiv.org/html/2607.08778#bib.bib11)\); Houet al\.\([2023](https://arxiv.org/html/2607.08778#bib.bib9),[2024](https://arxiv.org/html/2607.08778#bib.bib8)\); Wenet al\.\([2025](https://arxiv.org/html/2607.08778#bib.bib10)\)\. Yet, these models rely on a single shared encoder, which often fails to capture distinct feature patterns in heterogeneous populations and disease structures, commonly seen in complex diseases, where AD patient profiles vary in atrophy and other biomarkers\.

MoE in Clinical Modeling\.MoE architectures demonstrate improved ability to capture heterogeneous data distribution through specialized sub\-networks, known as expertsFeduset al\.\([2022](https://arxiv.org/html/2607.08778#bib.bib14)\); Zhuet al\.\([2024](https://arxiv.org/html/2607.08778#bib.bib19)\); Shenet al\.\([2024](https://arxiv.org/html/2607.08778#bib.bib22)\)\. The routing logic of the experts is learned by a gating network, typically a neural network\. Standalone MoE architectures have since been applied to clinical tasks, showing strengths in handling data heterogeneity, interpretability, and missing modalities in clinical scenariosZhanget al\.\([2025](https://arxiv.org/html/2607.08778#bib.bib18)\); Yunet al\.\([2024](https://arxiv.org/html/2607.08778#bib.bib17)\); Xinet al\.\([2025](https://arxiv.org/html/2607.08778#bib.bib15)\); Zhuanget al\.\([2025](https://arxiv.org/html/2607.08778#bib.bib16)\)\. However, MoE interpretability is typically limited to the expert weights from the gating network, leaving the routing logic needing to be analyzed post\-hoc and reducing transparency in clinical setting\.

LLM for Clinical Decision Support\.LLM is increasingly utilized to perform various tasks, plan, and make predictions in single\- or multi\-agent frameworks for clinical tasksTanget al\.\([2024](https://arxiv.org/html/2607.08778#bib.bib24)\); Wanget al\.\([2025](https://arxiv.org/html/2607.08778#bib.bib25)\); Houet al\.\([2025](https://arxiv.org/html/2607.08778#bib.bib23)\); Liet al\.\([2024](https://arxiv.org/html/2607.08778#bib.bib26)\); Kimet al\.\([2024](https://arxiv.org/html/2607.08778#bib.bib27)\)\. Given these capabilities, recent advances further demonstrate that LLMs can serve as a router, acting as the initial processor of information before making a decision where a task should goLiu and Lo \([2025](https://arxiv.org/html/2607.08778#bib.bib29)\); Onget al\.\([2024](https://arxiv.org/html/2607.08778#bib.bib28)\); Huet al\.\([2024](https://arxiv.org/html/2607.08778#bib.bib30)\)\. Unlike traditional classifiers, LLM routers can ingest natural language instructions and reasoning to improve routing decisions beyond structured data alone in clinical applications\.

## 3Method

### 3\.1LLM\-Guided Routing

Unlike traditional MoE models that rely on a gating network, our framework utilizes the LLM semantic reasoning to define and guide the expert specialization process in two stages as shown in Figure[1](https://arxiv.org/html/2607.08778#S1.F1)\. We use GPT 4\.1\-mini as the LLM for the two stages\.

Phenotype\-Driven Expert Initialization\.We utilize an LLM to map the high\-dimensional structured feature space from the list of neuroimaging measurement regions and the list of unstructured clinical notes into a set ofEEdistinct physiological phenotypes \(Appendix[A\.4](https://arxiv.org/html/2607.08778#A1.SS4), Table[3](https://arxiv.org/html/2607.08778#A1.T3)\)\. Given the set of feature namesFFand the prognostic task, the LLM generates an architectural specificationS=\{e1,e2,…,eE\}S=\\\{e\_\{1\},e\_\{2\},\.\.\.,e\_\{E\}\\\}, where each expert definitioneie\_\{i\}contains information such as the expert name and description \(Table[4](https://arxiv.org/html/2607.08778#A1.T4)\)\. In our setting, we useE=4E=4\. By definingEEwith the prompt, we ensure the MoE backbone is anchored in a fixed number of clinically separable AD progression patterns\.

Semantic Routing\.For a given patient with structured neuroimaging featuresxn​e​u​r​ox\_\{neuro\}\(i\.e\., measured brain regions of interest\) and unstructured clinical notesxn​o​t​e​sx\_\{notes\}\(i\.e\., initial assessment and medical history\), the routing weightsg​\(x\)g\(x\)are assigned by the LLM based on the patient’s alignment with the initialized phenotypesSS\. ROI features are aggregated into mesoscale brain region summaries to improve semantic interpretability \(Appendix[A\.5](https://arxiv.org/html/2607.08778#A1.SS5), Table[5](https://arxiv.org/html/2607.08778#A1.T5)\)\. The LLM serves as the semantic gating function:

g​\(x\)=Softmax​\(LLM​\(xn​e​u​r​o,xn​o​t​e​s,S\)\)\.g\(x\)=\\mbox\{Softmax\}\(\\mbox\{LLM\}\(x\_\{neuro\},x\_\{notes\},S\)\)\.\(1\)An example of the prompt and routing output are in Appendix[A\.6](https://arxiv.org/html/2607.08778#A1.SS6), Table[6](https://arxiv.org/html/2607.08778#A1.T6)\. The LLM outputs a probability distribution and clinical reasoning for the topkkexperts, where we definek=2k=2to mimic sparsity in MoE\. The reasoning provides human\-readable interpretation of the routing decisions\.

Table 1:Performance Comparison\.Comparison across brain imaging modalities \(VBM, FDG, AV45\) and cluster counts \(KK\)\. iLENS demonstrates improved or competitive prognosis \(C\-index\) and subtype separation \(LogRank\) across nearly all settings while maintaining competitive prognostic discrimination\. For each metric, the best\-performing model is highlighted inboldand the second\-best isunderlined\.−\-: LogRank undefined due to assignment of test samples to a single cluster\.ClusterCountModelVBM \(MRI\)FDG \(Metabolism\)AV45 \(Amyloid\)C\-Idx↑\\uparrowLog\-R↑\\uparrowC\-Idx↑\\uparrowLog\-R↑\\uparrowC\-Idx↑\\uparrowLog\-R↑\\uparrowKK=2SCA0\.5021\.630\.5881\.780\.5610\.79VaDeSC0\.5482\.450\.525−\-0\.61828\.73NSC0\.6044\.250\.6796\.560\.728−\-DCSM0\.638−\-0\.654−\-0\.78616\.81\\cellcolorgray\!15iLENS \(Ours\)\\cellcolorgray\!150\.670\\cellcolorgray\!152\.92\\cellcolorgray\!150\.775\\cellcolorgray\!1511\.22\\cellcolorgray\!150\.759\\cellcolorgray\!1536\.34KK=3SCA0\.4400\.600\.4503\.960\.5274\.46VaDeSC0\.4831\.550\.5620\.740\.56514\.81NSC0\.5290\.720\.69015\.500\.5083\.49DCSM0\.5957\.070\.8039\.310\.78320\.24\\cellcolorgray\!15iLENS \(Ours\)\\cellcolorgray\!150\.634\\cellcolorgray\!152\.54\\cellcolorgray\!150\.722\\cellcolorgray\!1525\.97\\cellcolorgray\!150\.767\\cellcolorgray\!1529\.01
### 3\.2Expert\-Routed Survival Clustering

Given the expert routing weightsg​\(x\)∈ℝEg\(x\)\\in\\mathbb\{R\}^\{E\}, each patient representationh​\(x\)h\(x\)is obtained as a weighted combination of expert outputs,

h​\(x\)=∑e=1Ege​\(x\)​fe​\(x\),h\(x\)=\\sum\_\{e=1\}^\{E\}g\_\{e\}\(x\)f\_\{e\}\(x\),\(2\)wherefe​\(x\)f\_\{e\}\(x\)denotes the output of experteeandge​\(x\)g\_\{e\}\(x\)is its routing weight\. To identify clinically meaningful subgroups, we model the survival process as a latent mixture ofKKdistinct survival subtypes\. Following the generative approach of previous clustering survival models such as DCSMHouet al\.\([2024](https://arxiv.org/html/2607.08778#bib.bib8)\), the survival function for a patient with representationh​\(𝐱\)h\(\\mathbf\{x\}\)is defined as

S​\(t∣𝐱\)=∑k=1Kπk​\(𝐱\)​Sk​\(t\),S\(t\\mid\\mathbf\{x\}\)=\\sum\_\{k=1\}^\{K\}\\pi\_\{k\}\(\\mathbf\{x\}\)S\_\{k\}\(t\),\(3\)whereπk​\(𝐱\)\\pi\_\{k\}\(\\mathbf\{x\}\)represents the probability of the patient belonging to thekk\-th survival subtype\. In our implementation, we setK=2K=2corresponding to a high\-risk and low\-risk stratification\. Each subtype survival functionSk​\(t\)S\_\{k\}\(t\)is modeled using a Weibull distribution, chosen for its flexibility and closed\-form expression,

Sk​\(t\)=exp⁡\(−\(tλk\)ρk\),S\_\{k\}\(t\)=\\exp\\left\(\{\-\\left\(\\frac\{t\}\{\\lambda\_\{k\}\}\\right\)^\{\\rho\_\{k\}\}\}\\right\),\(4\)whereλk\\lambda\_\{k\}andρk\\rho\_\{k\}denote the scale and shape parameters, respectively\. Unlike traditional models that estimate these parameters per\-patient, we treat them as global parameters that define the shared survival trajectory of each latent subtype\. The assignment weightsπk​\(𝐱\)\\pi\_\{k\}\(\\mathbf\{x\}\)are estimated from the LLM\-guided expert\-enhanced representationh​\(𝐱\)h\(\\mathbf\{x\}\)via a softmax layer:

πk​\(𝐱\)=exp⁡\(wkT​h​\(𝐱\)\)∑j=1Kexp⁡\(wjT​h​\(𝐱\)\)\.\\pi\_\{k\}\(\\mathbf\{x\}\)=\\frac\{\\exp\(w\_\{k\}^\{T\}h\(\\mathbf\{x\}\)\)\}\{\\sum\_\{j=1\}^\{K\}\\exp\(w\_\{j\}^\{T\}h\(\\mathbf\{x\}\)\)\}\.\(5\)These weights act as soft cluster memberships, allowing the model to provide an individualized survival prediction while maintaining interpretablility to population\-level risk profiles\. Details of the LLM\-guided survival modeling framework are provided in Appendix[A\.7](https://arxiv.org/html/2607.08778#A1.SS7)\.

## 4Experiments

### 4\.1Dataset

We evaluate iLENS on the Alzheimer’s Disease Neuroimaging Initiative \(ADNI\)Petersen and others \([2010](https://arxiv.org/html/2607.08778#bib.bib32)\)dataset, a longitudinal multimodal study that includes clinical, imaging, genetic, and biochemical biomarkers\. In this work, we obtained three structured imaging measurements from ADNI, namely the VBM, FDG, and AV45 for brain volume and biomarker measurements\. For clinical notes, we extracted the initial assessment and recent medical history \(Appendix[A\.1](https://arxiv.org/html/2607.08778#A1.SS1)\)\.

![Refer to caption](https://arxiv.org/html/2607.08778v1/x2.png)
![Refer to caption](https://arxiv.org/html/2607.08778v1/x3.png)
![Refer to caption](https://arxiv.org/html/2607.08778v1/x4.png)
![Refer to caption](https://arxiv.org/html/2607.08778v1/x5.png)
![Refer to caption](https://arxiv.org/html/2607.08778v1/x6.png)

Figure 2:Subtype Separation\.Kaplan–Meier survival curves for AV45 withK=2K=2across 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\.
### 4\.2Baselines

We compare iLENS with state\-of\-the\-art clustering survival models, including SCAChapfuwaet al\.\([2020](https://arxiv.org/html/2607.08778#bib.bib13)\), VaDeSCManduchiet al\.\([2021](https://arxiv.org/html/2607.08778#bib.bib12)\), NSCJeanselmeet al\.\([2022](https://arxiv.org/html/2607.08778#bib.bib11)\), and Deep Clustering Survival Machine \(DCSM\)Houet al\.\([2024](https://arxiv.org/html/2607.08778#bib.bib8)\)\. All models are hyperparameter\-tuned using OptunaAkibaet al\.\([2019](https://arxiv.org/html/2607.08778#bib.bib31)\)\(Appendix[A\.2](https://arxiv.org/html/2607.08778#A1.SS2), Table[2](https://arxiv.org/html/2607.08778#A1.T2)\)\.

### 4\.3Interpretable Prognosis and Subtyping

We first compare our framework against baselines using Concordance Index \(C\-index\) for prognostic performance and LogRank statistic to measure separation between discovered subtypes\. In practice, there is often a competitive trade\-off between individual prognostic accuracy and population\-level subtype separation, as seen in other baselines commonly sacrificing one for the other\. However, our results show that iLENS maintains competitive performance for both C\-index and LogRank \(Table[1](https://arxiv.org/html/2607.08778#S3.T1)\) across modalities and cluster counts, suggesting that LLM\-guided expert routing better captures heterogeneous AD progression patterns compared to a shared encoder\. Figure[2](https://arxiv.org/html/2607.08778#S4.F2)illustrates that iLENS produces more distinct survival subtype clusters compared to all baselines, reflecting clinically differentiable risk trajectories \(Appendix[A\.3](https://arxiv.org/html/2607.08778#A1.SS3)\) Importantly, iLENS also produces direct natural language rationale for expert assignments, which are grounded in each patient’s brain imaging measurements and medical history \(Table[7](https://arxiv.org/html/2607.08778#A1.T7)\)\. We provide qualitative examples of interpretable semantic routing in Appendix[A\.8](https://arxiv.org/html/2607.08778#A1.SS8)\. Additionally, to show the effectiveness of semantic LLM\-guided routing, we conduct ablation studies against non\-semantic and non\-specialized routing strategies, where we see that sparse LLM routing outperforms other sparse routing settings \(Appendix[A\.9](https://arxiv.org/html/2607.08778#A1.SS9), Tables[8](https://arxiv.org/html/2607.08778#A1.T8),[9](https://arxiv.org/html/2607.08778#A1.T9), and[10](https://arxiv.org/html/2607.08778#A1.T10)\)\.

## 5Conclusion

We introduced iLENS, a new LLM\-guided MoE framework for survival prediction\. By integrating the semantic reasoning of LLM with the capability of MoE in handling inherently heterogeneous clinical data, iLENS offers clinically\-grounded, interpretable survival predictions and improved subtyping\. Specifically, the results on the ADNI cohort show that our framework incorporates unstructured clinical notes alongside structured neuroimaging data to guide interpretable expert routing, whereas traditional survival models rely on voxel\-level correlations from the structured neuroimaging data\. This is demonstrated by improved subtype separation over all baselines measured by LogRank, including a two\-fold improvement over a standard MoE\-based clustering framework, establishing iLENS as a strong baseline for interpretable AD conversion survival analysis\.

## Limitations

We also identify several limitations of the current framework, including reliance on LLM\-based routing and expert initialization, as well as evaluation on a single cohort and disease setting\. In this work, we focus on subtypesK∈\{2,3\}K\\in\\\{2,3\\\}for clinically interpretable low/high\-risk stratification and moderate\-granularity subtype discovery\. Future work will assess robustness across different LLMs, survival outcomes, and subtyping settings\.

## Ethical Considerations

Our proposed framework is intended for research and exploratory purposes on survival modeling\. The generated expert assignments and natural language interpretability are based on LLM, which needs to be further evaluated for accuracy and hallucinations, especially when LLM\-generated explanations may still appear clinically plausible even when uncertain or partially correct\. At this stage, human clinical oversight remains necessary for any downstream clinical interpretation of patient assignments and rationales\.

## Potential Risks

Our proposed framework has several potential risks to be considered\. First, the framework uses semantic routing that relies on LLM, which may inherit biases present in pretrained language models\. Second, the generated expert descriptions also rely on semantic interpretations by the LLM on brain measurement features based on prior knowledge in the LLM’s training data\. If prior knowledge contains incorrect information or LLM suffers from hallucination, it may lead to incorrect or unstable routing assignments that may disrupt prognosis and subtyping in certain cases\. Third, the proposed framework was evaluated on ADNI cohort, which may not fully represent broader patient demographics, imaging protocols, or healthcare systems\.

## References

- Survival prediction landscape: an in\-depth systematic literature review on activities, methods, tools, diseases, and databases\.Frontiers in Artificial Intelligence7,pp\. 1428501\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p1.1)\.
- T\. Akiba, S\. Sano, T\. Yanase, T\. Ohta, and M\. Koyama \(2019\)Optuna: a next\-generation hyperparameter optimization framework\.InProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining,pp\. 2623–2631\.Cited by:[§A\.2](https://arxiv.org/html/2607.08778#A1.SS2.p1.1),[§4\.2](https://arxiv.org/html/2607.08778#S4.SS2.p1.1)\.
- J\. Ashburner and K\. J\. Friston \(2000\)Voxel\-based morphometry—the methods\.Neuroimage11\(6\),pp\. 805–821\.Cited by:[§A\.1](https://arxiv.org/html/2607.08778#A1.SS1.p1.2)\.
- H\. Binder and M\. Schumacher \(2008\)Allowing for mandatory covariates in boosting estimation of sparse high\-dimensional survival models\.BMC bioinformatics9\(1\),pp\. 14\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p1.1)\.
- A\. Carobbio, P\. Guglielmelli, E\. Rumi, C\. Cavalloni, V\. De Stefano, S\. Betti, A\. Rambaldi, M\. C\. Finazzi, J\. Thiele, A\. M\. Vannucchi,et al\.\(2020\)A multistate model of survival prediction and event monitoring in prefibrotic myelofibrosis\.Blood cancer journal10\(10\),pp\. 100\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p1.1)\.
- P\. Chapfuwa, C\. Li, N\. Mehta, L\. Carin, and R\. Henao \(2020\)Survival cluster analysis\.InProceedings of the ACM Conference on Health, Inference, and Learning,pp\. 60–68\.Cited by:[§4\.2](https://arxiv.org/html/2607.08778#S4.SS2.p1.1)\.
- W\. Fedus, B\. Zoph, and N\. Shazeer \(2022\)Switch transformers: scaling to trillion parameter models with simple and efficient sparsity\.Journal of Machine Learning Research23\(120\),pp\. 1–39\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p2.1)\.
- J\. Fox and S\. Weisberg \(2002\)Cox proportional\-hazards regression for survival data\.An R and S\-PLUS companion to applied regression2002,pp\. 7\.Cited by:[§1](https://arxiv.org/html/2607.08778#S1.p1.1),[§2](https://arxiv.org/html/2607.08778#S2.p1.1)\.
- P\. Glare, K\. Virik, M\. Jones, M\. Hudson, S\. Eychmuller, J\. Simes, and N\. Christakis \(2003\)A systematic review of physicians’ survival predictions in terminally ill cancer patients\.Bmj327\(7408\),pp\. 195\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p1.1)\.
- D\. Goyal, D\. Tjandra, R\. Q\. Migrino, B\. Giordani, Z\. Syed, J\. Wiens, A\. D\. N\. Initiative,et al\.\(2018\)Characterizing heterogeneity in the progression of alzheimer’s disease using longitudinal clinical and neuroimaging biomarkers\.Alzheimer’s & Dementia: Diagnosis, Assessment & Disease Monitoring10,pp\. 629–637\.Cited by:[§1](https://arxiv.org/html/2607.08778#S1.p1.1)\.
- B\. Hou, H\. Li,et al\.\(2023\)Deep clustering survival machines with interpretable expert distributions\.In2023 IEEE 20th International Symposium on Biomedical Imaging \(ISBI\),pp\. 1–4\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p1.1)\.
- B\. Hou, Z\. Wen, J\. Bao, R\. Zhang, B\. Tong, S\. Yang, J\. Wen, Y\. Cui, J\. H\. Moore, A\. J\. Saykin,et al\.\(2024\)Interpretable deep clustering survival machines for alzheimer’s disease subtype discovery\.Medical image analysis97,pp\. 103231\.Cited by:[§A\.7](https://arxiv.org/html/2607.08778#A1.SS7.p3.2),[§2](https://arxiv.org/html/2607.08778#S2.p1.1),[§3\.2](https://arxiv.org/html/2607.08778#S3.SS2.p1.7),[§4\.2](https://arxiv.org/html/2607.08778#S4.SS2.p1.1)\.
- W\. Hou, G\. Yang, Y\. Du, Y\. Lau, L\. Liu, J\. He, L\. Long, and S\. Wang \(2025\)Adagent: llm agent for alzheimer’s disease analysis with collaborative coordinator\.InInternational Workshop on Agentic AI for Medicine,pp\. 23–32\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p3.1)\.
- Q\. J\. Hu, J\. Bieker, X\. Li, N\. Jiang, B\. Keigwin, G\. Ranganath, K\. Keutzer, and S\. K\. Upadhyay \(2024\)Routerbench: a benchmark for multi\-llm routing system\.arXiv preprint arXiv:2403\.12031\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p3.1)\.
- V\. Jeanselme, B\. Tom, and J\. Barrett \(2022\)Neural survival clustering: non\-parametric mixture of neural networks for survival clustering\.InConference on Health, Inference, and Learning,pp\. 92–102\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p1.1),[§4\.2](https://arxiv.org/html/2607.08778#S4.SS2.p1.1)\.
- J\. L\. Katzman, U\. Shaham, A\. Cloninger, J\. Bates, T\. Jiang, and Y\. Kluger \(2018\)DeepSurv: personalized treatment recommender system using a cox proportional hazards deep neural network\.BMC medical research methodology18\(1\),pp\. 24\.Cited by:[§1](https://arxiv.org/html/2607.08778#S1.p1.1),[§2](https://arxiv.org/html/2607.08778#S2.p1.1)\.
- Y\. Kim, C\. Park, H\. Jeong, Y\. S\. Chan, X\. Xu, D\. McDuff, H\. Lee, M\. Ghassemi, C\. Breazeal, and H\. W\. Park \(2024\)Mdagents: an adaptive collaboration of llms for medical decision\-making\.Advances in Neural Information Processing Systems37,pp\. 79410–79452\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p3.1)\.
- B\. Li, T\. Yan, Y\. Pan, J\. Luo, R\. Ji, J\. Ding, Z\. Xu, S\. Liu, H\. Dong, Z\. Lin,et al\.\(2024\)Mmedagent: learning to use medical tools with multi\-modal agent\.InFindings of the Association for Computational Linguistics: EMNLP 2024,pp\. 8745–8760\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p3.1)\.
- K\. Liu and M\. Lo \(2025\)LLM\-based routing in mixture of experts: a novel framework for trading\.arXiv preprint arXiv:2501\.09636\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p3.1)\.
- L\. Manduchi, R\. Marcinkevičs, M\. C\. Massi, T\. Weikert, A\. Sauter, V\. Gotta, T\. Müller, F\. Vasella, M\. C\. Neidert, M\. Pfister,et al\.\(2021\)A deep variational approach to clustering survival data\.arXiv preprint arXiv:2106\.05763\.Cited by:[§4\.2](https://arxiv.org/html/2607.08778#S4.SS2.p1.1)\.
- G\. Mirabnahrazam, D\. Ma, C\. Beaulac, S\. Lee, K\. Popuri, H\. Lee, J\. Cao, J\. E\. Galvin, L\. Wang, M\. F\. Beg,et al\.\(2023\)Predicting time\-to\-conversion for dementia of alzheimer’s type using multi\-modal deep survival analysis\.Neurobiology of aging121,pp\. 139–156\.Cited by:[§1](https://arxiv.org/html/2607.08778#S1.p1.1)\.
- C\. Nagpal, X\. Li, and A\. Dubrawski \(2021\)Deep survival machines: fully parametric survival regression and representation learning for censored data with competing risks\.IEEE Journal of Biomedical and Health Informatics25\(8\),pp\. 3163–3175\.Cited by:[§1](https://arxiv.org/html/2607.08778#S1.p1.1),[§2](https://arxiv.org/html/2607.08778#S2.p1.1)\.
- T\. Nakagawa, M\. Ishida, J\. Naito, A\. Nagai, S\. Yamaguchi, K\. Onoda, and A\. D\. N\. Initiative \(2020\)Prediction of conversion to alzheimer’s disease using deep survival analysis of mri images\.Brain communications2\(1\),pp\. fcaa057\.Cited by:[§1](https://arxiv.org/html/2607.08778#S1.p1.1)\.
- I\. Ong, A\. Almahairi, V\. Wu, W\. Chiang, T\. Wu, J\. E\. Gonzalez, M\. W\. Kadous, and I\. Stoica \(2024\)Routellm: learning to route llms with preference data\.arXiv preprint arXiv:2406\.18665\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p3.1)\.
- R\. C\. Petersenet al\.\(2010\)Alzheimer’s disease neuroimaging initiative \(adni\) clinical characterization\.Neurology74\(3\),pp\. 201–209\.Cited by:[§1](https://arxiv.org/html/2607.08778#S1.p1.1),[§4\.1](https://arxiv.org/html/2607.08778#S4.SS1.p1.1)\.
- E\. T\. Rolls, C\. Huang, C\. Lin, J\. Feng, and M\. Joliot \(2020\)Automated anatomical labelling atlas 3\.Neuroimage206,pp\. 116189\.Cited by:[§A\.1](https://arxiv.org/html/2607.08778#A1.SS1.p1.2)\.
- L\. Shen, G\. Chen, R\. Shao, W\. Guan, and L\. Nie \(2024\)Mome: mixture of multimodal experts for generalist multimodal large language models\.Advances in neural information processing systems37,pp\. 42048–42070\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p2.1)\.
- R\. A\. Sperling, P\. S\. Aisen, L\. A\. Beckett, D\. A\. Bennett, S\. Craft, A\. M\. Fagan, T\. Iwatsubo, C\. R\. Jack Jr, J\. Kaye, T\. J\. Montine,et al\.\(2011\)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\.Alzheimer’s & dementia7\(3\),pp\. 280–292\.Cited by:[§1](https://arxiv.org/html/2607.08778#S1.p1.1)\.
- X\. Tang, A\. Zou, Z\. Zhang, Z\. Li, Y\. Zhao, X\. Zhang, A\. Cohan, and M\. Gerstein \(2024\)Medagents: large language models as collaborators for zero\-shot medical reasoning\.InFindings of the Association for Computational Linguistics: ACL 2024,pp\. 599–621\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p3.1)\.
- Z\. Wang, J\. Wu, L\. Cai, C\. H\. Low, X\. Yang, Q\. Li, and Y\. Jin \(2025\)Medagent\-pro: towards evidence\-based multi\-modal medical diagnosis via reasoning agentic workflow\.arXiv preprint arXiv:2503\.18968\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p3.1)\.
- Z\. Wen, B\. Hou,et al\.\(2025\)Multi\-modal deep clustering survival machines for alzheimer’s disease subtype discovery\.InProc\. of the IEEE/CVF Int\. Conf\. on Computer Vision,pp\. 2264–2272\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p1.1)\.
- J\. Xin, S\. Yun, J\. Peng, I\. Choi, J\. L\. Ballard, T\. Chen, and Q\. Long \(2025\)I2moe: interpretable multimodal interaction\-aware mixture\-of\-experts\.arXiv preprint arXiv:2505\.19190\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p2.1)\.
- S\. Yun, I\. Choi, J\. Peng, Y\. Wu, J\. Bao, Q\. Zhang, J\. Xin, Q\. Long, and T\. Chen \(2024\)Flex\-moe: modeling arbitrary modality combination via the flexible mixture\-of\-experts\.Advances in Neural Information Processing Systems37,pp\. 98782–98805\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p2.1)\.
- W\. Zhang, W\. Xu, T\. Chen, C\. Sakal, and X\. Li \(2025\)Integrating images and genomics for multi\-modal cancer survival analysis via mixture of experts\.Information Fusion,pp\. 103521\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p2.1)\.
- X\. Zhu, Y\. Hu, F\. Mo, M\. Li, and J\. Wu \(2024\)Uni\-med: a unified medical generalist foundation model for multi\-task learning via connector\-moe\.Advances in Neural Information Processing Systems37,pp\. 81225–81256\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p2.1)\.
- F\. Zhuang, D\. Aliyeva, S\. Yang, Z\. Wen, D\. Duong\-Tran, C\. Davatzikos, T\. Chen, S\. Wang, and L\. Shen \(2025\)Multimodal fusion of regional brain experts for interpretable alzheimer’s disease diagnosis\.arXiv preprint arXiv:2512\.10966\.Cited by:[§2](https://arxiv.org/html/2607.08778#S2.p2.1)\.

## Appendix AAppendix

### A\.1Dataset Details

For our experiment, we focus on 398 unique participants who were diagnosed with Mild Cognitive Impairment \(MCI\) and remained MCI in their last visit \(censored\) or converted to AD diagnosis at any point in the study \(uncensored\)\. We include their brain imaging measurements in tabular format mapped using Anatomical Automatic Labeling \(AAL\)Rollset al\.\([2020](https://arxiv.org/html/2607.08778#bib.bib33)\)\. Measurements include structural magnetic resonance imaging \(MRI\), commonly used to detect atrophy in the brain, that were converted to Voxel\-Based Morphometry \(VBM\) mapAshburner and Friston \([2000](https://arxiv.org/html/2607.08778#bib.bib34)\), as well as Florbetapir \(AV45\) and fluorodeoxyglucose \(FDG\) positron emission tomography \(PET\) imaging, which are used to detect the amyloid\-β\\beta\(Aβ\\beta\) load and brain glucose metabolism in the brain\. Hence, the modalities are named VBM, FDG, and AV45 in the results table \(Table[1](https://arxiv.org/html/2607.08778#S3.T1)\) of this work\. This results in a table of quantitative measures for each region of interest \(ROI\) in the brain from the MRI image, where each value represents the relative gray matter volume within the specific brain structures\. Additionally, we include summary descriptions from each participant’s initial assessment and recent medical history\. These are left as natural language\. To construct the data for survival prediction, we labeled unique participants with whether the event did not occur \(censored\)\. For participants with multiple visits for brain imaging, we take the imaging information from the first MCI diagnosis\. For patients who converted to AD, we take the date of the first AD diagnosis to calculate the number of days from the first MCI diagnosis\. Using this data, we perform AD diagnosis for each participant into one of the three states\. Data is split into 60/10/30 for training, validation, and test datasets, respectively\.

The ADNI dataset was previously collected under institutional review board \(IRB\) approval and informed consent procedures\. All experiments were conducted on de\-identified research data\. The ADNI data we use in this work is publicly available but requires approval for access\. More information about the dataset and instructions on how to request for access are available through the ADNI website[https://adni\.loni\.usc\.edu/](https://adni.loni.usc.edu/)\. We use the ADNI dataset consistently with its intended use for research on Alzheimer’s disease\.

### A\.2Training, Hyperparameter Tuning, and Evaluation

For all methods, we run OptunaAkibaet al\.\([2019](https://arxiv.org/html/2607.08778#bib.bib31)\)for hyperparameter tuning\. We employ the Tree\-structured Parzen Estimator \(TPE\) sampler for a total of 15 trials per model\. This ensures a rigorous and fair comparison across all architectures by identifying optimal configurations for both baselines and our proposed framework\. We maximize the validation metric for each C\-index and LogRank across the number of clustersKKand imaging modalities \(VBM, FDG, AV45\)\. We report the resulting performance using the optimal hyperparameters on the test data in Table[1](https://arxiv.org/html/2607.08778#S3.T1)based on the single data split\. Optimal hyperparameters for iLENS are shown in Table[2](https://arxiv.org/html/2607.08778#A1.T2)

Table 2:Optimized hyperparameters for iLENS across modalities and latent subgroup settings\. Separate Optuna studies were conducted for C\-index and LogRank optimization objectives\.ModalityKKObjectiveLRDiscountDropoutBatchVBM2C\-index4\.76e\-40\.5120\.00164VBM3C\-index2\.46e\-30\.9800\.15564FDG2C\-index3\.10e\-40\.6340\.03316FDG3C\-index4\.58e\-50\.8110\.06232AV452C\-index2\.71e\-30\.5880\.15216AV453C\-index1\.20e\-30\.9670\.19816VBM2LogRank3\.71e\-40\.8450\.06164VBM3LogRank1\.14e\-50\.9800\.10964FDG2LogRank2\.52e\-30\.7990\.11632FDG3LogRank4\.81e\-40\.6870\.18616AV452LogRank1\.12e\-50\.5740\.13516AV453LogRank1\.66e\-30\.5040\.26832
### A\.3Subtyping

Figure[2](https://arxiv.org/html/2607.08778#S4.F2)shows the Kaplan–Meier survival curves for AV45 PET amyloid imaging withK=2K=2latent survival subtypes across clustering survival methods\. iLENS produces the clearest separation between high\-risk and low\-risk groups, resulting in the highest LogRank statistic reported in Table[1](https://arxiv.org/html/2607.08778#S3.T1)\. Other baseline methods show weaker separation or collapsed subtype assignments, such as for NSC where all samples are assigned to a single cluster, resulting in no subtyping\. DCSM and VaDeSC are able to identify more distinguished survival curves, but they show less separation compared to iLENS\. These results suggest that our LLM\-guided expert routing framework is able to improve the discovery of AD progression patterns based on amyloid PET imaging\.

### A\.4Expert Design

In the first step of our LLM\-routing framework, we performed expert design using the LLM\. We prompted the LLM with modality\-specific feature names, a description of the imaging modality and measurements, and representative clinical narrative fields, while constraining the output to four JSON\-formatted expert phenotypes\. The generated experts and descriptions allow the downstream LLM routing to interpret them in a consistent manner for the mixture\-of\-experts\. To encourage clinically meaningful specialization of the experts, we explicitly constrained the LLM to generate experts corresponding to distinct progression mechanisms and multimodal phenotypes rather than feature missingness patterns or arbitrary partitions\. The prompt additionally restricts the LLM to use only information available within the current modality\. As a result, the generated experts differ across modalities \(e\.g\., AV45 amyloid PET, FDG PET, and VBM MRI\) and reflect modality\-specific progression characteristics\.

Table[3](https://arxiv.org/html/2607.08778#A1.T3)shows an example the prompt template used for AV45 expert initialization, while Table[4](https://arxiv.org/html/2607.08778#A1.T4)shows the resulting AV45 expert phenotypes generated by the LLM\.

LLM Expert Initialization PromptSystem Prompt:You are an expert nuclear\-medicine neurologist preparing phenotype experts for survival analysis\.Design experts for MCI\-to\-AD conversion risk using amyloid PET \(Florbetapir AV\-45\) SUVR mesoscale patterns plus clinical text\.You do not diagnose or recommend treatment\.Hard constraints:\- Use AV\-45 regional SUVR\-style signals \(from av45\_\* features\) and clinical narratives \(initial assessment / recent medical history\)\.\- Do NOT assume FDG hypometabolism, structural MRI/VBM, CSF, or tau PET unless such columns appear in feature\_names\.\- Experts must be clinically separable mechanisms \(e\.g\. spatial amyloid distribution, comorbid context\), not missingness buckets\.Output JSON only, no markdown\.Rules:\- num\_experts must be exactly 4\.\- Expert ids must be 0\.\.3\.Task Prompt:Propose 4 AD progression phenotype experts using AV\-45 amyloid PET \(cortical SUVR\) patterns plus initial assessment/recent medical history narratives for subtype clustering and survival/log\-rank separation\.Table 3:Prompt template used for AV45 expert initialization\.IDNameRiskDescriptionE0Vascular comorbidity low\-amyloid phenotypeLowLow SUVR values across cortical regions \(especially frontal, parietal, temporal\) combined with clinical text indicating hypertension, stroke, or vascular disease\. Highlights patients whose cognitive impairment may be driven more by vascular pathology than amyloid burden, explaining slower AD progression risk\.E1Frontal amyloid with neuropsychiatric symptomsMid\-HighElevated SUVR in frontal superior, middle, and inferior gyri plus clinical notes describing agitation, apathy, or executive dysfunction\. Captures a phenotype with frontal lobe amyloid pathology linked to behavioral symptoms, associated with faster progression than low\-amyloid vascular phenotype\.E2Posterior cortical amyloid with memory impairmentMid\-HighHigh SUVR in precuneus, posterior cingulate, parietal, and occipital cortices combined with clinical text emphasizing memory loss and cognitive decline\. Represents the classical AD progression phenotype with posterior amyloid deposition and memory impairment, supporting established clinical\-pathological correlations\.E3Diffuse high amyloid with multiple comorbiditiesHighElevated SUVR in frontal, parietal, temporal, and occipital regions combined with clinical notes indicating rapid progression, multiple medical issues, or neuropsychiatric plus memory symptoms\. Identifies patients with extensive amyloid pathology and complex clinical profiles, associated with the highest risk of rapid MCI\-to\-AD conversion\.Table 4:LLM\-initialized AV45 expert phenotypes used for semantic routing\.
### A\.5Mesoscale Brain Region Summarization

The original ADNI imaging features consist of region\-of\-interest \(ROI\) measurements derived from the Automated Anatomical Labeling \(AAL\) atlas, where each anatomical structure is represented separately for the left and right hemispheres \(e\.g\.,Frontal\_Sup\_LandFrontal\_Sup\_R\)\. There are a total of 116 ROI measurements in each modality\. To reduce dimensionality and improve semantic interpretability for LLM routing, we aggregate anatomically related ROIs into mesoscale brain\-region groups\.

For each patient, we compute the mean across ROI\-level SUVR measurements within each mesoscale group to produce compact regional summaries \(e\.g\., Frontal\-Lobe, Cingulate, Parietal\-Lobe\)\. These summaries are then provided to the LLM router as interpretable imaging descriptors alongside the clinical narrative summaries for each user input\. The resulting mesoscale representation preserves broad spatial amyloid distribution patterns while substantially reducing prompt complexity\. Table[5](https://arxiv.org/html/2607.08778#A1.T5)shows representative mesoscale region groupings used\.

Mesoscale RegionAAL ROIs IncludedFrontal\-LobeFrontal\_Sup, Frontal\_Sup\_Orb, Frontal\_Mid, Frontal\_Mid\_Orb, Frontal\_Inf\_Oper, Frontal\_Inf\_Tri, Frontal\_Inf\_Orb, Rolandic\_Oper, Supp\_Motor\_Area, Frontal\_Sup\_Medial, Frontal\_Med\_Orb, RectusInsulaInsulaCingulateCingulum\_Ant, Cingulum\_Mid, Cingulum\_PostTemporal\-Lobe\-CorticalOlfactory, ParaHippocampal, Lingual, Fusiform, Heschl, Temporal\_Sup, Temporal\_Pole\_Sup, Temporal\_Mid, Temporal\_Pole\_Mid, Temporal\_InfTemporal\-SubcorticalHippocampus, AmygdalaOccipital\-LobeCalcarine, Cuneus, Occipital\_Sup, Occipital\_Mid, Occipital\_InfSensory\-Motor\-CortexPrecentral, Postcentral, Paracentral\_LobuleParietal\-LobeParietal\_Sup, Parietal\_Inf, SupraMarginal, Angular, PrecuneusStriatum/Basal\-GangliaCaudate, Putamen, PallidumThalamusThalamusCerebellumCerebelum\_Crus1, Cerebelum\_Crus2, Cerebelum\_3, Cerebelum\_4\_5, Cerebelum\_6, Cerebelum\_7b, Cerebelum\_8, Cerebelum\_9, Cerebelum\_10, Vermis\_1\_2, Vermis\_3, Vermis\_4\_5, Vermis\_6, Vermis\_7, Vermis\_8, Vermis\_9, Vermis\_10Table 5:Mesoscale region groupings derived from AAL atlas regions\. Each anatomical structure contains both left and right hemisphere measurements, which are aggregated using mean SUVR summaries within each mesoscale category\.
### A\.6LLM\-routing

During routing for the mixture\-of\-experts, the LLM receives only modality\-specific imaging summaries across mesoscale brain regions, clinical text, and the previously generated expert definitions\. We ensure that the router does not receive survival labels, conversion outcomes, censoring status, subtype assignments, or time\-to\-event information\. We provide the prompt template used for AV45 routing in Table[6](https://arxiv.org/html/2607.08778#A1.T6)as an example\. The prompt instructs the LLM to assign sparse mixture weights across phenotype experts using only imaging measurements and clinical notes while constraining routing to clinically meaningful progression patterns\. We set TOP\_K to 2 for each modality to employ a sparse mixture\-of\-experts\.

LLM Router Prompt Template for AV45 PETSystem Prompt:You are a routing controller for an Alzheimer disease \(AD\) mixture\-of\-experts survival model \(amyloid PET AV\-45 \+ clinical text\)\. Route patients to phenotype experts using imaging and history; you do not diagnose, stage, or prescribe\.Imaging input:Mesoscale summaries fromav45\_\*columns \(AAL atlas\)\. Each value is the nanmean across ROI\-level SUVR columns within the bucket\. These are per\-patient regional means on the merged SUVR scale in this run, not z\-scores versus a cognitively normal reference\.Interpretation guide – AV45:Values are SUVR\-like cortical amyloid uptake\. Higher regional values generally indicate greater fibrillar amyloid deposition in that compartment, subject to acquisition and reference\-region normalization\. Compare across mesoscale regions, such as precuneus/posterior cingulate, lateral frontal, and temporal neocortex, to identify spatial amyloid burden patterns\. Values near 0 may reflect low uptake, reference choice, or sparse data; interpret alongside the clinical narrative\. Use as probabilistic phenotype routing for progression risk, not as a clinical decision\.Clinical text:clinical\_narrative\_summaryis one paragraph\. It explainsINITHEALTHversusRECMHIST, gives calendar\-day offsets from the cohort anchor visit and AV45 PET acquisition, and quotes the intake and recent\-history narratives\.Do not request FDG, MRI/VBM, CSF, or raw ROI tables beyond what is given\. Assign nonnegative mixture weights summing to 1 for at mostTOP\_Kexperts\.Output format:```
{
  "weights": {"<expert_id>": <float>, ...},
  "rationale": "<string>"
}
```

Use 1–TOP\_Knonzero experts and renormalize if needed\.Patient\-specific input template:```
{
  "TOP_K": 2,
  "num_experts": 4,
  "expert_personas": [
    {
      "id": 0,
      "name": "Vascular comorbidity low-amyloid phenotype",
      "mechanism": "...",
      "focus": "...",
      "avoid_routing_when": "...",
      "expected_relative_risk": "lowest",
      "interpretability_value": "..."
    },
    { "id": 1, "...": "..." },
    { "id": 2, "...": "..." },
    { "id": 3, "...": "..." }
  ],
  "imaging_mesoscale_summary": {
    "summary_style": "regional_mean",
    "mesoscale_region_means": {
      "av45": {
        "Frontal-Lobe": 1.47,
        "Cingulate": 1.85,
        "...": "..."
      }
    }
  },
  "clinical_narrative_summary":
    "**Patient initial health assessment and medical history narrative** ..."
}
```

Table 6:Prompt template used for AV45 PET semantic routing\. The router receives imaging summaries, clinical narrative summaries, and expert definitions, but not survival outcomes or subtype labels\.
### A\.7LLM\-guided Survival Modeling

We train a survival clustering framework using the LLM\-routed sparse mixture\-of\-experts \(MoE\) representation layer\. The experts are initialized as identical neural networks\. During training, we obtain the LLM\-assigned weights during LLM routing\. The assigned experts then receive only structured modality\-specific imaging features as numeric inputs without information from clinical notes\. Clinical notes instead influences the model indirectly through the precomputed LLM routing weights generated during previous LLM routing stage\.

More specifically, the representation layer consists ofE=4E=4parallel expert multilayer perceptrons \(MLPs\), where each expert maps the imaging inputxxinto a latent embeddinghe​\(x\)∈ℝdh\_\{e\}\(x\)\\in\\mathbb\{R\}^\{d\}\. The final representation is constructed as a weighted combination of expert embeddingsΦ​\(x\)=∑e=1Ewe​he​\(x\)\\Phi\(x\)=\\sum\_\{e=1\}^\{E\}w\_\{e\}h\_\{e\}\(x\), where𝐰\\mathbf\{w\}denotes the sparse LLM\-generated routing weights\. Unlike conventional MoE architectures with learnable gating networks, the routing weights are frozen during optimization and retrieved from precomputed LLM routing outputs\. Only the expert subnetworks and downstream survival parameters are updated during training\.

After the MoE representation layer, the model learns a latent mixture ofKKWeibull survival components, whereK∈\{2,3\}K\\in\\\{2,3\\\}in our experiments\. These latent components define statistical survival subgroups used for survival prediction and subtype discoveryHouet al\.\([2024](https://arxiv.org/html/2607.08778#bib.bib8)\)\.

At inference time, the fitted Weibull mixture model produces individualized survival risk estimates\. For Kaplan–Meier and LogRank evaluation, patients are assigned to latent survival subgroups based on the highest predicted subgroup probability, and clusters are ordered from low to high risk according to mean predicted survival risk\.

### A\.8Qualitative Routing Examples

To further examine the interpretability of our proposed framework, we provide representative example of LLM\-guided semantic routing using AV45 amyloid PET imaging and clinical narrative summaries\. During routing, the LLM receives only modality\-specific imaging summaries, clinical text, and the previously generated expert definitions as presented in the previous subsection \(Appendix[A\.6](https://arxiv.org/html/2607.08778#A1.SS6)\)\. The prompt format for the LLM router is provided in Table[6](https://arxiv.org/html/2607.08778#A1.T6)

Table[7](https://arxiv.org/html/2607.08778#A1.T7)presents the output example from a representative patient who later converted from MCI to AD\. The example shows how the LLM combines spatial amyloid distribution patterns with clinical comorbidity information to produce semantically interpretable expert assignments\. In this case, the router assigns the highest weight to the diffuse high\-amyloid phenotype while also partially routing the patient to the posterior cortical amyloid phenotype, reflecting mixed progression characteristics observed from the imaging and clinical narrative context\.

Semantic Routing Output Example \(AV45 PET\)Patient Imaging Summary:Diffuse elevated AV\-45 SUVR values across frontal \(1\.47\), cingulate \(1\.85\), temporal \(1\.35\), parietal \(1\.39\), occipital \(1\.32\), striatal \(1\.56\), and thalamic \(1\.53\) regions\.Clinical Narrative Summary:History includes anxiety \(citalopram\), sleep apnea \(CPAP\), hypothyroidism, and hyperlipidemia\.LLM Routing Weights:E2 \(Posterior cortical amyloid\): 0\.30E3 \(Diffuse high amyloid\): 0\.70LLM Rationale:The patient shows diffuse high amyloid SUVR values across multiple cortical regions including frontal \(1\.47\), cingulate \(1\.85\), temporal \(1\.35\), parietal \(1\.39\), occipital \(1\.32\), and subcortical areas \(striatum 1\.56, thalamus 1\.53\)\. This widespread elevated amyloid burden aligns with the “Diffuse high amyloid with multiple comorbidities” phenotype \(expert 3\)\. Clinical text indicates multiple comorbidities such as mild anxiety, sleep apnea, hypothyroidism, and hyperlipidemia, supporting routing to expert 3\. A secondary weight is assigned to expert 2\. Overall, the patient fits best with expert 3 \(highest risk\) and to a lesser extent expert 2 \(posterior cortical amyloid with memory impairment\)\.Subtype Cluster:High riskOutcome \(hidden from router\):Converted to AD after 723 days\.Table 7:Representative semantic routing output example from the AV45 PET modality\. The router receives only imaging summaries, clinical narrative text, and expert definitions, without access to survival labels or conversion outcomes\.
### A\.9Routing Ablation Studies

We performed additional ablation experiments to evaluate the contribution of semantic LLM\-guided routing compared to non\-semantic or non\-specialized routing strategies\. Specifically, we compared against \(i\) uniform dense routing, where all experts receive equal weights, \(ii\) uniform sparse routing with equal sparse weights across selected experts, \(iii\) random sparse routing, and \(iv\) LLM routing without clinical narrative information, where routing is generated using imaging summaries only\.

Tables[8](https://arxiv.org/html/2607.08778#A1.T8),[9](https://arxiv.org/html/2607.08778#A1.T9), and[10](https://arxiv.org/html/2607.08778#A1.T10)summarize the results across AV45 PET, FDG PET, and VBM MRI modalities for bothK=2K=2andK=3K=3latent survival subgroup settings\.

Overall, semantic LLM\-guided routing consistently outperformed sparse non\-semantic routing variants, including random sparse routing and uniform sparse routing, particularly for LogRank to measure subtype separation\. The strongest improvements can be seen in AV45 PET and FDG PET, where semantic routing substantially improved survival subgroup stratification\. Removing clinical narrative information generally reduced LogRank performance, suggesting that clinical text contributes information beyond imaging measurements alone\.

Interestingly, uniform dense routing remains competitive in several ablation settings, indicating that averaging across all experts can provide stable representations even without sparse semantic specialization\. However, we note that the dense setting increases model complexity\. However, one downside with dense routing setting is that it removes patient\-specific expert selection and does not provide interpretable semantic routing assignments\. In contrast, our proposed LLM\-guided routing framework produces sparse, clinically interpretable expert allocations while achieving competitive or improved survival stratification performance across multiple modality and subgroup configurations\.

Routing VariantK=2K=2K=3K=3C\-index↑\\uparrowLogRank↑\\uparrowC\-index↑\\uparrowLogRank↑\\uparrowUniform Dense0\.78280\.00000\.782643\.2742Uniform Sparse0\.73973\.73510\.77088\.4878Random Sparse0\.70871\.17500\.750114\.5708LLM Routing w/o Clinical Notes0\.781023\.98170\.784240\.9760iLENS0\.759036\.34000\.767029\.0100Table 8:Routing ablation study for AV45 PET imaging\.Routing VariantK=2K=2K=3K=3C\-index↑\\uparrowLogRank↑\\uparrowC\-index↑\\uparrowLogRank↑\\uparrowUniform Dense0\.769521\.69030\.760913\.2011Uniform Sparse0\.73960\.55500\.51332\.3439Random Sparse0\.68011\.51890\.61080\.3452LLM Routing w/o Clinical Notes0\.780618\.06320\.74881\.9295iLENS0\.775011\.22000\.722025\.9700Table 9:Routing ablation study for FDG PET imaging\.Routing VariantK=2K=2K=3K=3C\-index↑\\uparrowLogRank↑\\uparrowC\-index↑\\uparrowLogRank↑\\uparrowUniform Dense0\.66671\.83220\.684410\.3793Uniform Sparse0\.65620\.22360\.64200\.9036Random Sparse0\.58900\.43130\.56622\.7410LLM Routing w/o Clinical Notes0\.67000\.00480\.66311\.8550iLENS0\.67002\.92000\.63402\.5400Table 10:Routing ablation study for VBM MRI imaging\.
### A\.10AI Assistance in Research

We used AI assistance \(e\.g\., Cursor, ChatGPT\) to support grammatical checks and writing revisions, coding frameworks and prompt design, and visualization generation\.

### A\.11Computational Resources\.

Experiments were conducted on Apple Silicon hardware using an Apple M4 processor with 16GB unified memory\.