Explicit Knowledge-Guided In-Context Learning for Early Detection of Alzheimer's Disease

TL;DR

This paper tackles Alzheimer’s disease detection from narrative transcripts using in-context learning, addressing poor task recognition and label-task misalignment in clinical, data-scarce, and OOD scenarios. It introduces EK-ICL, which injects explicit knowledge from three sources—SLM-based confidence ($S_{ ext{conf}}$), parsing-feature scores ($S_{ ext{feat}}$), and in-distribution label words (ID labels)—and complements them with parsing-based demo retrieval and ensemble majority voting. The approach demonstrates significant performance gains over fine-tuning and standard ICL baselines across three AD datasets, with analysis showing that careful label semantics alignment and task context are crucial for robust clinical reasoning. The findings highlight the value of explicit knowledge in guiding LLM reasoning under resource constraints and semantic-homogeneity challenges in medical transcripts.

Abstract

Detecting Alzheimer's Disease (AD) from narrative transcripts remains a challenging task for large language models (LLMs), particularly under out-of-distribution (OOD) and data-scarce conditions. While in-context learning (ICL) provides a parameter-efficient alternative to fine-tuning, existing ICL approaches often suffer from task recognition failure, suboptimal demonstration selection, and misalignment between label words and task objectives, issues that are amplified in clinical domains like AD detection. We propose Explicit Knowledge In-Context Learners (EK-ICL), a novel framework that integrates structured explicit knowledge to enhance reasoning stability and task alignment in ICL. EK-ICL incorporates three knowledge components: confidence scores derived from small language models (SLMs) to ground predictions in task-relevant patterns, parsing feature scores to capture structural differences and improve demo selection, and label word replacement to resolve semantic misalignment with LLM priors. In addition, EK-ICL employs a parsing-based retrieval strategy and ensemble prediction to mitigate the effects of semantic homogeneity in AD transcripts. Extensive experiments across three AD datasets demonstrate that EK-ICL significantly outperforms state-of-the-art fine-tuning and ICL baselines. Further analysis reveals that ICL performance in AD detection is highly sensitive to the alignment of label semantics and task-specific context, underscoring the importance of explicit knowledge in clinical reasoning under low-resource conditions.

Figures (5)

• Figure 1: Schematic illustration of reasoning failure in ICL under out-of-distribution (OOD) task scenarios.
• Figure 2: Parsing category distributions across datasets in AD detection. This figure compares the average frequency of six syntactic categories—Actions, Objects, Fillers, Locations, Subjects, and Pronouns—across the ADReSS challenge-Train, ADReSS challenge-Test, Lu, Pitt, and Total datasets.
• Figure 3: Overview of the proposed EK-ICL framework.
• Figure 4: Comparative performance of ICL baselines under OOD/ID label-word settings. (a) ICL$_{\text{Van}}$ and ICL$_{\text{Sem}}$ on the Test dataset. (b) ICL$_{\text{Log}}$ and ICL$_{\text{Ens}}$ on three datasets.
• Figure 5: Performance of EK-ICL on the Test set across 30 label word pairs categorized into three configurations: Aligned, Fixed Good, and Fixed Bad. Gray hatched bars indicate values that are not applicable due to division by zero.