Locally deployed large language model for real-world lecanemab eligibility pre-screening.

INTRODUCTION

The introduction of disease-modifying Alzheimer's therapies requires complex, labor-intensive patient screening. Cloud-based large language models (LLMs) could support this task but are often unsuitable for routine care due to data protection constraints.

METHODS

We evaluated an on-premises, open-weights LLM (gpt-oss-120b) for automated extraction of therapy-relevant variables for lecanemab eligibility from German memory clinic reports. In a two-stage design, LLM-based extraction prompts, a deterministic rule-based extractor, and a shared downstream rule-based classifier were optimized on a development set (n = 97) and evaluated on an independent hold-out set (n = 99), with expert consensus as ground truth.

RESULTS

The LLM-based pipeline achieved 94% accuracy and a Cohen's kappa of 0.90 on the hold-out set, significantly surpassing the rule-based comparator (80% accuracy) and demonstrating performance comparable to human experts.

DISCUSSION: A locally deployed, on-premises LLM may assist eligibility screening as a triage support tool, potentially facilitating access to novel therapies without compromising patient data privacy.