Location
https://www.kennesaw.edu/ccse/events/computing-showcase/sp26-cday-program.php
Document Type
Event
Start Date
22-4-2026 4:00 PM
Description
This study evaluates how speech disfluencies and prompting strategies impact LLM-based Alzheimer’s Disease (AD) detection. We compared transcripts with preserved disfluencies (ADReSS) against clean transcripts (ADReSSo) using four state-of-the-art LLMs. Key Discovery: Complex prompts induce a "mirror-image" classification bias, where DeepSeek models severely over-classify AD, and GPT-5.2 over-classifies Cognitively Normal (CN) individuals. Optimization Fix: Applying DSPy MIPROv2 effectively mitigated bias in simpler prompts, while TextGrad successfully optimized complex, multi-step prompts.
Included in
GRM-094-176 Influence of Speech Disfluencies and Prompt Optimization on LLM-Based Alzheimer's Detection
https://www.kennesaw.edu/ccse/events/computing-showcase/sp26-cday-program.php
This study evaluates how speech disfluencies and prompting strategies impact LLM-based Alzheimer’s Disease (AD) detection. We compared transcripts with preserved disfluencies (ADReSS) against clean transcripts (ADReSSo) using four state-of-the-art LLMs. Key Discovery: Complex prompts induce a "mirror-image" classification bias, where DeepSeek models severely over-classify AD, and GPT-5.2 over-classifies Cognitively Normal (CN) individuals. Optimization Fix: Applying DSPy MIPROv2 effectively mitigated bias in simpler prompts, while TextGrad successfully optimized complex, multi-step prompts.