文献速递妇科产科

Emily与Lakisha:量化AI生成卵巢癌诊疗方案中的社会经济与结构性差异

Emily vs. Lakisha: Quantifying socioeconomic and structural disparities in AI-generated ovarian cancer care plans.
PubMed 文献 · Gynecol Oncol · 2026-02-24

该横断面随机算法审计纳入320例,评估四种大语言模型在2×2析因设计下对晚期卵巢癌转诊方案的结构性偏见。模型临床指南依从性高,但对黑人/农村交叉表型存在结构性无能,且财务固定由种族驱动,对Lakisha引入未提示费用警告的OR为16.80(p<0.001)。

要点

  • LLM生成方案临床准确但存在结构性偏见,形成「医学-结构分离」。
  • 对黑人/农村交叉表型结构性无能最严重,未优先提供物流解决方案。
  • 种族信号驱动财务固定,Lakisha比Emily更易触发未提示费用警告。

核心结论

  • 当前大语言模型临床准确但存在结构性偏见,可能加剧健康不平等并导致算法红线。
  • 安全部署需从生物医学保真转向结构对齐,使AI缓解而非固化系统差异。

研究设计与样本

  • 横断面随机算法审计,N=320,评估四种大语言模型(ChatGPT 5.2、Gemini 3、Grok 4.1、Claude 4.5)。
  • 采用2×2析因 vignette 设计,分离姓名种族代理(Emily vs. Lakisha)与地理(Boston vs. Indianola)效应。
  • 使用经过验证的自动结构能力评分管道,涵盖23项指标评估临床一致性、优先级和语言语气。

主要发现

  • 模型表现出医学-结构分离:无论人口学特征如何,临床指南依从性均高。
  • 算法对黑人/农村交叉表型显示严重结构无能。
  • 敏感性分析显示财务固定几乎完全由种族驱动(p<0.001),对Lakisha引入未提示费用警告的OR为16.80,与地点无关。

关键数值与推荐

  • 对Lakisha相比Emily引入未提示费用警告的比值比OR为16.80,p<0.001,与地点无关。
  • 评估涵盖23项指标,包括临床一致性、优先级和语言语气。

临床怎么用

  • 临床使用LLM生成诊疗方案时需警惕基于姓名和地理的隐性结构性偏见。
  • 应推动AI系统实现结构对齐,作为认知假体缓解而非固化系统差异。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

BACKGROUND: Rapid integration of Large Language Models (LLMs) into oncology workflows mandates scrutiny of implicit biases that may exacerbate health inequities. This study quantifies structural biases in LLM-generated referral plans for Advanced Ovarian Cancer. METHODS: We conducted a cross-sectional, randomized algorithmic audit (N = 320) of four state-of-the-art LLMs (ChatGPT 5.2, Gemini 3, Grok 4.1, Claude 4.5) using a 2 × 2 factorial vignette design to disentangle the effects of onomastic Race proxy (Emily vs. Lakisha) from Geography (Boston vs. Indianola). Responses were evaluated using the validated "Automated Structural Competence Scoring (ASCS) Pipeline" across 23 metrics assessing clinical concordance, prioritization, and linguistic tone. RESULTS: Models exhibited a critical "Medical-Structural Dissociation." While maintaining high clinical guideline adherence regardless of demographics, algorithms displayed severe structural incompetence for the intersectional phenotype (Black/Rural). Sensitivity analysis revealed that 'Financial Fixation' was driven almost exclusively by Race (p < 0.001), with models being 16.8 times more likely (OR: 16.80) to introduce unprompted cost warnings for 'Lakisha' compared to 'Emily,' regardless of location. CONCLUSION: Current LLMs are clinically accurate but structurally biased. By hallucinating financial scarcity based on onomastic signals and failing to prioritize logistical solutions for rural patients, these models risk imposing a cognitive burden on physicians and inducing "algorithmic redlining." Safe deployment in primary care requires moving beyond biomedical fidelity toward "Structural Alignment," ensuring AI functions as a cognitive prosthetic that mitigates, rather than entrenches, systemic disparities.

来源:Gynecol Oncol,查看原文。版权归原作者所有。

引用

  1. İnan M, Akar İnan S, Aypak C, et al. Emily vs. Lakisha: Quantifying socioeconomic and structural disparities in AI-generated ovarian cancer care plans. Gynecol Oncol. 2026. PMID: 41740457. doi:10.1016/j.ygyno.2026.02.011

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。