LipoAssist: a structured GPT-4–based clinical workflow for preliminary lipedema assessment

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Objective Lipedema is a chronic adipose tissue disorder characterized by bilateral and symmetrical subcutaneous fat accumulation, predominantly affecting women. Because it is frequently confused with obesity and lymphedema, diagnosis may be delayed. This proof-of-concept study aimed to evaluate the feasibility of LipoAssist, a structured GPT-4–based clinical workflow designed for the preliminary assessment of lipedema under simulated conditions. Methods Ten simulated clinical scenarios representing lipedema and relevant differential diagnoses were evaluated using LipoAssist. The workflow was designed to obtain a structured medical history, assess clinically relevant symptoms, and generate a standardized case summary. Three board-certified Physical Medicine and Rehabilitation specialists independently evaluated the AI-generated outputs using a 5-point Likert scale across six criteria. A total of 180 ratings were analyzed. Inter-rater agreement was assessed using the intraclass correlation coefficient and Kendall’s coefficient of concordance. Results The overall mean performance score was 3.55 ± 0.64. The highest scores were observed for correct understanding of the clinical condition (4.63 ± 0.49) and identification of the most likely diagnosis (4.53 ± 0.51). Lower scores were recorded for recommendations regarding further diagnostic evaluation (2.17 ± 0.83) and assessment of surgical necessity (1.93 ± 0.64). Case-summary clarity and adequacy of history taking received mean scores of 3.87 ± 0.63 and 3.73 ± 0.64, respectively. Inter-rater agreement was good (ICC = 0.82; 95% CI: 0.68–0.91), and Kendall’s W was 0.79. Conclusion LipoAssist demonstrated promising performance in structured history taking, organization of clinical information, and identification of the most likely diagnosis in simulated lipedema scenarios. However, its performance was limited in advanced diagnostic recommendations and surgical decision-making. These findings support the feasibility of a structured GPT-4–based workflow under simulated conditions but do not establish clinical validity, diagnostic accuracy, or readiness for routine implementation.