iLENS uses LLMs for interpretable Alzheimer's disease risk prediction

Alzheimer's disease affects millions worldwide, and predicting when patients will convert from mild cognitive impairment to full-blown AD is a critical challenge. Traditional survival models have limited interpretability and cannot perform natural language reasoning. To address this, a team from institutions including University of Pennsylvania and Indiana University developed iLENS—an interpretable large language model (LLM) guided mixture-of-experts (MoE) framework. iLENS takes both structured neuroimaging measurements (like volumetric MRI data) and unstructured clinical information, then uses the LLM to intelligently route different expert models for the most accurate survival prediction. The MoE architecture allows specialized sub-models to focus on different patient subgroups, improving both accuracy and interpretability.

The iLENS framework demonstrates competitive predictive performance on Alzheimer's disease conversion benchmarks, and importantly, it enables patient subtyping—identifying distinct progression patterns among individuals. Beyond raw performance, iLENS provides transparent rationales for its routing decisions, explaining why certain experts were chosen based on biologically meaningful features. This bridges a long-standing gap between high-performance survival analysis and clinically interpretable decision support. The paper, published on arXiv, represents a promising step toward AI systems that doctors can trust for prognosis, combining the reasoning power of LLMs with the precision of specialized survival models. The approach could extend beyond Alzheimer's to other neurodegenerative diseases where interpretable risk prediction is essential.