In the study of Alzheimer's Disease (AD), predicting its conversion process is crucial. Traditional survival models, while widely used for AD risk prediction, are typically static and lack interpretability and natural language reasoning capabilities. To address this, we propose iLENS, an interpretable large language model (LLM) guided framework based on a mixture-of-experts (MoE) approach for survival prediction.
This method utilizes LLM to synthesize structured neuroimaging measurements and unstructured information, guiding expert routing decisions. Our framework demonstrates competitive predictive performance and effective patient subtyping. Furthermore, iLENS provides transparent, biologically grounded rationales for its routing decisions, bridging the gap between high-performance survival analysis and interpretable clinical decision support.
Blogger's Review: The innovation of iLENS lies in combining LLM with a mixture-of-experts model, significantly enhancing the interpretability and utility of survival analysis. This approach not only aids clinical decision-making but also opens new avenues for research in neuroimaging, making it a noteworthy development.