CLONE: Synthetic Guideline-based Clinical Reasoning with LLMs for Early Diagnosis of Mild Cognitive Impairment

Abstract

Early diagnosis of mild cognitive impairment (MCI) is essential to prevent its progression to Alzheimer's disease. Human expert-driven diagnosis provides interpretable rationales but is time-consuming, while machine learning-based approaches offer efficiency but lack human-readable rationales. To address these limitations, we propose CLONE (Clinical Reasoning via Neuropsychologist Emulation), a three-stage framework leveraging large language models (LLMs) for MCI diagnosis: (1) emulating experts through role-playing, (2) synthesizing step-by-step diagnostic guidelines, and (3) performing clinical reasoning using the guideline. CLONE was evaluated on a real-world dataset of 65 subjects, achieving 89.23% diagnostic accuracy and outperforming the few-shot chain-of-thought (CoT) baseline by 6.15%, with specificity improving by 10.71%. Moreover, the synthesized guideline enhanced rationale quality, making rationales more consistent, correct, specific, helpful, and human-like compared to baselines. These findings highlight CLONE's potential to enable accurate diagnosis and reliable clinical reasoning, addressing challenges in the field of MCI diagnosis.

Related Projects

• VR-EP-EEG-MRI 디지털 바이오마커 기초연구실

  2021.06 — 2024.02