Leveraging LLMs for Early Alzheimer's Prediction

TL;DR

The paper tackles early Alzheimer's prediction by converting dynamic fMRI connectomics into prompts for a frozen large language model, enabling effective learning from limited clinical data. It introduces RevRIN-enabled temporal patching and a patch-to-token reprogramming scheme with semantic anchors to bridge connectome features to LLM inputs, followed by a lightweight prediction head. On the OASIS-3 dataset (677 subjects), the method achieves MAE for the MCI+IMP group well below the minimal clinically important difference, and global distance metrics outperform local ones, with consistent results across LLaMA, GPT-2, and BERT backbones. Ablation results confirm the necessity of RevRIN and carefully chosen architecture, demonstrating the feasibility and clinical potential of frozen-LLM pipelines for early neurodegenerative detection.

Abstract

We present a connectome-informed LLM framework that encodes dynamic fMRI connectivity as temporal sequences, applies robust normalization, and maps these data into a representation suitable for a frozen pre-trained LLM for clinical prediction. Applied to early Alzheimer's detection, our method achieves sensitive prediction with error rates well below clinically recognized margins, with implications for timely Alzheimer's intervention.

Figures (1)

• Figure 1: Overview of connectome-LLM framework.