Sex-Specific Prompt Engineering for OSA Severity Classification Using Large Language Models
2
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
Obstructive Sleep Apnea (OSA) is a prevalent but underdiagnosed sleep disorder associated with serious health risks. Althoughpolysomnography (PSG) is the diagnostic gold standard, its cost and clinical burden limit scalability. This study explores the use of LargeLanguage Models (LLMs), particularly the lightweight GPT-4o-mini, to classify OSA severity (Normal, Moderate, Severe) based solely onnon-invasive demographic and sleep questionnaire data, without access to PSG features. We develop a prompt engineering pipeline thatincorporates sex-specific feature importance extracted from Random Forest classifiers to design three types of prompts: Basic, Role-Based,and Chain-of-Thought (CoT). Experimental results show that Chain-of-Thought (CoT) prompts with feature emphasis tailored to maleand female subgroups achieve F1 scores of up to 0.60–0.62, outperforming non-sex-specific (generic) prompts and matching theperformance of Random Forest baselines. These findings highlight the effectiveness of sex-aware, few-shot prompting in improvingdiagnostic performance. Our results suggest that prompt-based LLMs can serve as interpretable, low-resource alternatives to traditionalsupervised models, with potential for scalable and equitable OSA screening.
Abstract
Obstructive Sleep Apnea (OSA) is a prevalent but underdiagnosed sleep disorder associated with serious health risks. Althoughpolysomnography (PSG) is the diagnostic gold standard, its cost and clinical burden limit scalability. This study explores the use of LargeLanguage Models (LLMs), particularly the lightweight GPT-4o-mini, to classify OSA severity (Normal, Moderate, Severe) based solely onnon-invasive demographic and sleep questionnaire data, without access to PSG features. We develop a prompt engineering pipeline thatincorporates sex-specific feature importance extracted from Random Forest classifiers to design three types of prompts: Basic, Role-Based,and Chain-of-Thought (CoT). Experimental results show that Chain-of-Thought (CoT) prompts with feature emphasis tailored to maleand female subgroups achieve F1 scores of up to 0.60–0.62, outperforming non-sex-specific (generic) prompts and matching theperformance of Random Forest baselines. These findings highlight the effectiveness of sex-aware, few-shot prompting in improvingdiagnostic performance. Our results suggest that prompt-based LLMs can serve as interpretable, low-resource alternatives to traditionalsupervised models, with potential for scalable and equitable OSA screening.