Andrii Bidochko

• Updated: July 22, 2026
• 7 min read

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

Direct Answer

iLENS is a novel, interpretable survival‑analysis system that combines a large language model (LLM) with a mixture‑of‑experts (MoE) architecture to predict conversion from mild cognitive impairment to Alzheimer’s disease (AD). By letting the LLM translate raw neuroimaging measurements and clinical notes into routing cues for specialized expert networks, iLENS delivers competitive accuracy while exposing biologically meaningful rationales for each prediction.

Background: Why This Problem Is Hard

Alzheimer’s disease unfolds over years, and early identification of patients who will progress from prodromal stages to full‑blown dementia is a cornerstone of both therapeutic research and patient care. Traditional survival models—Cox proportional hazards, random‑survival forests, or deep survival networks—have two critical blind spots:

• Static feature handling: They treat imaging biomarkers (e.g., hippocampal volume, cortical thickness) as fixed vectors, ignoring the rich textual context found in radiology reports, physician notes, and longitudinal narratives.
• Lack of interpretability: Even when performance is high, clinicians receive a black‑box risk score with no clear link to the underlying pathology, limiting trust and regulatory acceptance.

Moreover, neuroimaging data are high‑dimensional and heterogeneous. Combining MRI‑derived tensors with sparse clinical descriptors often forces researchers to hand‑craft feature pipelines, a process that is both time‑consuming and error‑prone. The field therefore needs a method that can (a) ingest structured and unstructured data, (b) adapt its reasoning to patient sub‑populations, and (c) surface transparent, disease‑relevant explanations.

What the Researchers Propose

The iLENS framework tackles the above gaps by weaving together three core ideas:

1. LLM‑guided synthesis: A pre‑trained large language model reads free‑text clinical notes and neuroimaging summaries, then produces a concise “routing token” that captures the most salient disease cues for a given patient.
2. Mixture‑of‑Experts (MoE) survival network: A collection of expert subnetworks—each trained on a distinct phenotypic slice of the data (e.g., amyloid‑positive vs. amyloid‑negative cohorts)—receives the routing token and decides whether to specialize or share computation.
3. Neuroimaging feature integration: Structured imaging metrics (volumes, texture features, connectivity scores) are fed into each expert alongside the LLM token, allowing the model to align visual biomarkers with linguistic insights.

In essence, iLENS lets a language model act as a “clinical triage officer,” directing each patient’s data to the most appropriate expert, while the MoE backbone preserves the statistical rigor of survival analysis.

How It Works in Practice

Conceptual workflow

1. Data ingestion: Raw MRI scans are processed through a standard neuroimaging pipeline (e.g., FreeSurfer) to extract quantitative measures. Simultaneously, electronic health records (EHR) and radiology narratives are collected.
2. LLM reasoning: The LLM receives the unstructured text, performs entity extraction (e.g., “hippocampal atrophy,” “APOE ε4”), and outputs a low‑dimensional embedding that encodes the clinician’s narrative judgment.
3. Routing decision: The embedding is fed into a lightweight gating network. The gate computes a probability distribution over the set of experts, effectively selecting which subnetworks will process the case.
4. Expert survival computation: Each chosen expert runs a deep survival model (e.g., a Weibull‑parameterized network) on the concatenated structured imaging vector and the LLM token, producing a hazard function over time.
5. Aggregation & output: The weighted hazards from all active experts are summed to yield a final survival curve, from which conversion risk at any future horizon (e.g., 2‑year, 5‑year) can be extracted.
6. Interpretability layer: Because the gate’s decision is driven by the LLM token, the system can back‑track to the original textual cues that triggered a particular expert, offering clinicians a transparent rationale.

What makes iLENS different

• Dynamic expert allocation: Unlike static MoE models that assign data based on pre‑defined clusters, iLENS adapts routing per patient using natural‑language context.
• Unified multimodal fusion: The LLM token and imaging vector are merged early, allowing the survival experts to learn cross‑modal interactions without hand‑crafted feature engineering.
• Built‑in interpretability: The gating mechanism is directly traceable to human‑readable language, satisfying clinical audit requirements.

Evaluation & Results

Experimental setup

The authors evaluated iLENS on a longitudinal cohort drawn from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). The dataset comprised:

• 3,200 participants with baseline MRI and up to 8 years of follow‑up.
• Structured imaging features (volumes of hippocampus, ventricles, cortical thickness maps).
• Unstructured clinical notes (baseline neuropsychological assessments, radiology reports).

Four baselines were compared:

1. Cox proportional hazards with handcrafted imaging covariates.
2. DeepSurv (a deep neural survival model) using only structured data.
3. Transformer‑based survival model that ingests text but lacks expert routing.
4. Standard MoE without LLM guidance.

Key findings

• Predictive performance: iLENS achieved a concordance index (C‑index) of 0.82, surpassing the best baseline (DeepSurv at 0.76) by 8 percentage points.
• Patient sub‑typing: The gating network consistently routed APOE ε4 carriers to a specialist expert that emphasized amyloid‑related imaging markers, demonstrating biologically plausible stratification.
• Interpretability audit: In 92 % of cases, the top‑ranked textual cue matched a known AD risk factor (e.g., “memory decline,” “temporal lobe atrophy”), providing clinicians with a clear decision trail.
• Computational efficiency: Because only a subset of experts is activated per patient, inference time remained comparable to a single‑expert model, despite the added LLM step.

These results indicate that iLENS not only improves risk prediction but also bridges the gap between high‑performance AI and actionable clinical insight.

Why This Matters for AI Systems and Agents

iLENS exemplifies a new class of AI agents that combine language understanding with domain‑specific numeric reasoning. For practitioners building health‑focused agents, the framework offers several practical takeaways:

• Modular design: The separation of LLM reasoning, gating, and expert survival modules mirrors the micro‑service architecture favored in enterprise AI platforms, making it easier to swap components (e.g., replace the LLM with a newer model).
• Explainable decision pipelines: By exposing the LLM‑generated routing token, developers can embed “why‑this‑prediction” widgets directly into dashboards, satisfying regulatory demands for transparency.
• Scalable expert orchestration: The gating mechanism can be mapped onto existing workflow orchestration tools—such as the Workflow automation studio—to dynamically allocate compute resources based on patient complexity.
• Cross‑modal agent capabilities: iLENS shows that an agent can ingest both free‑text and high‑dimensional imaging data without manual feature alignment, a pattern that can be replicated for other multimodal domains (e.g., pathology slides + lab reports).

In short, iLENS provides a blueprint for building AI agents that are both high‑performing and clinically trustworthy, a combination that has been elusive in most current health‑AI deployments.

What Comes Next

While iLENS marks a significant step forward, several avenues remain open for exploration:

• Generalization to other diseases: Extending the MoE‑LLM paradigm to Parkinson’s disease, multiple sclerosis, or oncology could validate its broader applicability.
• Real‑time deployment: Integrating iLENS into hospital information systems will require latency‑optimised LLM inference (e.g., via quantisation or on‑device models).
• Richer multimodal inputs: Adding PET scans, CSF biomarkers, or wearable sensor streams could further improve risk stratification.
• Regulatory pathways: Formalizing the interpretability audit into a certified “explainability module” would accelerate FDA or EMA clearance.
• Open‑source tooling: Publishing a plug‑and‑play version of the gating network could foster community contributions and accelerate adoption.

Organizations interested in prototyping such pipelines can start by leveraging the UBOS platform overview, which offers pre‑built connectors for LLM APIs and data stores, or explore the AI marketing agents for inspiration on orchestrating multiple expert services.

References

iLENS: Interpretable LLM‑Guided Mixture‑of‑Experts for Neuroimaging Survival Analysis (arXiv)

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.