Can large language models approximate human perceptions of disease severity? An evaluation using Global Burden of Disease 2010 disability weights

Background: Disability weights (DWs) quantify the severity of health loss and are essential for estimating disability-adjusted life years in the Global Burden of Disease (GBD) framework. Conventional DW estimation relies on resource-intensive population surveys that are difficult to update or adapt to emerging health states. Large language models (LLMs) may offer a scalable alternative by approximating human perceptions of disease severity through structured judgment tasks. Methods: This exploratory study evaluated the alignment between LLM-derived and human-derived DW rankings using 222 health states from GBD 2010. All possible pairwise comparisons (24,531 pairs, each repeated three times) were conducted across four LLMs (GPT-5 mini, GPT-5, Claude Haiku 4.5, and Claude Sonnet 4.5). DWs were estimated via probit regression and evaluated using Spearman's rank correlation and Steiger's z test. The effects of prompt language (English vs. Korean), cultural role prompting, and medical specialist role prompting on alignment were examined. Additionally, the Binomial-Logit Indifference-Point (BLIP) estimator was proposed and validated through leave-one-out cross-validation for estimating DWs for health states without established values. Results: All four LLMs showed high rank correlation with GBD 2010 DWs (Spearman's {rho} = 0.893 to 0.909), with no significant inter-model differences. Korean-language prompting significantly improved alignment with Korean DWs ({rho} = 0.756 vs. 0.715, p = 0.011), and Korean cultural role prompting improved alignment with both GBD 2010 DWs ({rho} = 0.922 vs. 0.909, p = 0.002) and Korean DWs ({rho} = 0.738 vs. 0.715, p = 0.001). Medical specialist role prompting significantly reduced alignment with GBD 2010 DWs ({rho} = 0.895 vs. 0.909, p = 0.001). BLIP demonstrated strong agreement with GBD 2010 DWs (Pearson's r = 0.862, MAE = 0.066) and produced plausible estimates for Long COVID (mild: 0.020, moderate: 0.298, severe: 0.529). Conclusions: LLMs can approximate human perceptions of disease severity with high rank-order consistency. Prompt language and role framing significantly influenced alignment, with culturally grounded lay prompting enhancing and specialist prompting reducing correspondence with population-based DWs. BLIP provides a practical framework for generating provisional DW estimates for emerging or underrepresented health states when conventional surveys are infeasible.