甲状腺癌人工智能模型的代表性与偏倚:系统综述
该研究系统检索EMBASE、PubMed和Google Scholar至2024年1月,纳入197项涉及甲状腺癌管理且报告人口学细节的AI模型研究。多数研究聚焦诊断(n=133)和预测/预后(n=47),参与者主要来自中国(n=124)和美国(n=26)。结果显示,东亚人群在数据中代表性不足,白人和黑人参与者相对过代表,社会经济因素、婚姻状况和种族/民族较少被纳入模型。
要点
- 甲状腺癌AI模型以诊断类为主(133项),预测/预后类47项。
- 东亚人群代表性不足(14.6% vs 全球患病率18.7%),白人和黑人相对过代表。
- 模型较少考虑社会经济因素、婚姻状况和种族/民族等变量。
核心结论
- 甲状腺癌AI模型在数据多样性和代表性方面存在显著缺口,模型虽与流行病学趋势一致,但缺乏全面的人口学纳入。
- 未来需开发并验证更公平的AI模型,以改善不同人群的甲状腺癌诊疗。
研究设计与样本
- 系统综述,已在PROSPERO注册(ID: CRD42024519238),遵循Cochrane手册和PRISMA指南。
- 检索EMBASE、PubMed和Google Scholar至2024年1月,纳入涉及甲状腺癌管理AI模型并提供人口学细节的研究。
- 数据提取和偏倚风险评估由两名独立评审员完成,最终纳入197项研究。
主要发现
- 197项研究中,诊断类133项,预测/预后类47项。
- 参与者主要来自中国(n=124)和美国(n=26),女性(n=12,410)多于男性(n=4222)。
- 197项研究共248,896名参与者的族裔数据显示,东亚人群占14.6%,低于全球甲状腺癌患病率18.7%。
- 白人(26.8%)和黑人(26.8%)参与者相对全球患病率(20.7%和11.3%)过代表。
- 社会经济因素、婚姻状况和种族/民族在模型中较少被考虑。
临床怎么用
- 选用甲状腺癌AI模型时应关注其训练数据的人群构成与目标人群是否匹配。
- 对东亚等代表性不足人群应用AI输出时需谨慎解读,警惕诊断与治疗建议的差异。
局限性
- 检索截止2024年1月,未涵盖此后发表的模型研究。
- 仅纳入提供人口学细节的研究,可能遗漏未报告人口学信息的AI模型。
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: There has been growing interest in the application of artificial intelligence (AI) in thyroid cancer care, given its potential to enhance diagnostic accuracy, predict patient outcomes, and personalize treatment plans. However, bias introduced during the development of AI algorithms used for thyroid cancer care poses a significant challenge, as biased datasets can lead to disparities in diagnosis and treatment recommendations, particularly in underrepresented populations. This systematic review evaluates the current landscape of AI models for thyroid cancer, focusing on demographic representation and potential biases. Methods: This systematic review was registered on PROSPERO (ID: CRD42024519238) and conducted in accordance with the Cochrane handbook and reported in accordance with Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines. A literature search was performed on EMBASE, PubMed, and Google Scholar up to January 2024. Studies were included if they involved AI models for thyroid cancer management and provided demographic details. Data extraction and risk-of-bias assessments were conducted by two independent reviewers. Results: A total of 197 studies were included in the review, with the majority focusing on diagnosis (n = 133) and prediction/prognosis (n = 47). Most studies predominantly involved participants from China (n = 124) and the United States (n = 26), with more female participants (n = 12,410) than males (n = 4222). Ethnicity data from 197 studies (248,896 participants) revealed a significant underrepresentation of East Asians (14.6%) compared with their global thyroid cancer prevalence (18.7%), while White (26.8%) and Black participants (26.8%) were overrepresented relative to their global prevalence (20.7% and 11.3%, respectively). Socioeconomic factors, marital status, and race/ethnicity were less frequently considered in the models. Conclusion: The findings highlight significant gaps in the diversity and representativeness of data used in thyroid cancer AI models. Current models align with epidemiological trends but lack comprehensive demographic inclusion. As such, more representative AI models are required that account for all aspects of a patient's demographics and sociocultural background. Future research should focus on developing and validating more equitable AI models to improve thyroid cancer care across diverse populations.
引用
- Ramchandani R, Guo E, Biglou SG, et al. Representation and Bias in Artificial Intelligence Models for Thyroid Cancer: A Systematic Review. Thyroid. 2025. PMID: 40865545. doi:10.1177/10507256251372175
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。