文献速递医学人工智能与信息学流行病学与生物统计基础与转化研究

撤稿:没有正常大脑:人口学排除如何损害神经影像AI的效度

RETRACTED: No Normal Brain: How Demographic Exclusion Undermines Neuroimaging AI Validity.
PubMed 文献 · J Clin Epidemiol · 2026-02-20

该分析指出,五大神经影像数据集(HCP、ADNI、UK Biobank、ABCD、OpenNeuro)中白人占比52-94%,北美/欧洲来源占85-100%,导致AI性能差异,如青春期女性癫痫检测敏感性低27%,黑人痴呆筛查假阳性率翻倍。结论认为人口学排除威胁研究效度与临床安全。

要点

  • 神经影像AI训练数据存在严重人口学集中,白人占比52-94%。
  • 数据排除导致性能差异:青春期女性癫痫检测敏感性低27%。
  • 黑人痴呆筛查假阳性率翻倍,需警惕临床转化风险。

核心结论

  • 神经影像AI中的人口学排除对研究效度和临床安全构成根本威胁。

研究设计与样本

  • 分析五个主要神经影像数据集(HCP、ADNI、UK Biobank、ABCD、OpenNeuro),代表超过55,000名参与者。
  • 收集参与者种族/民族、性别、年龄、地理来源、社会经济地位和教育程度特征,并分析已发表文献中的AI性能差异。

主要发现

  • 主要数据集白人占比52-94%,而全球人口占比仅16%。
  • 85-100%参与者来自北美/欧洲,而全球占比仅15%。
  • 青春期女性癫痫检测敏感性低27%。
  • 黑人参与者痴呆筛查假阳性率翻倍。

关键数值与推荐

  • 白人占比52-94%,对应全球人口占比16%。
  • 北美/欧洲来源占比85-100%,对应全球占比15%。
  • 青春期女性癫痫检测敏感性低27%。
  • 黑人参与者痴呆筛查假阳性率翻倍。

临床怎么用

  • 临床使用神经影像AI时应评估训练数据的人口学代表性。
  • 对青春期女性癫痫检测和黑人痴呆筛查需警惕性能偏差。
  • 推动多样性、公平与包容举措以保障AI科学严谨性。

局限性

  • 原文未报告

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 AND OBJECTIVE: Neuroimaging AI systems increasingly influence clinical decisions, yet demographic exclusions in training datasets may compromise their scientific validity and clinical safety. This analysis examines how demographic concentration in major neuroimaging repositories undermines the fundamental neurobiological principle that there is no single "normal" brain, potentially creating AI systems that pathologize neurological diversity rather than embrace it as biological reality. METHODS: I analyzed five major neuroimaging datasets (HCP, ADNI, UK Biobank, ABCD, OpenNeuro) representing >55,000 participants. For each dataset, I collected participant characteristics across race/ethnicity, sex/gender, age distribution, geographic origin, socioeconomic status, and educational attainment. I analyzed published literature documenting performance disparities in neuroimaging AI systems across demographic groups. RESULTS: Major neuroimaging datasets demonstrate severe demographic concentration: 52-94% White representation despite 16% global population share, with 85-100% from North America/Europe representing only 15% globally. These exclusions produce predictable AI performance disparities, including 27% lower seizure detection sensitivity in adolescent females and doubled false positive rates in dementia screening for Black participants. Contemporary COVID-19 brain aging research demonstrates bias amplification in demographically restricted training data. CONCLUSION: Demographic exclusion in neuroimaging AI represents a fundamental threat to research validity and clinical safety. Diversity, equity, and inclusion initiatives provide essential mechanisms for ensuring scientific rigor by acknowledging neurological diversity as biological reality rather than statistical inconvenience.

来源:J Clin Epidemiol,查看原文。版权归原作者所有。

引用

  1. Rudroff T. RETRACTED: No Normal Brain: How Demographic Exclusion Undermines Neuroimaging AI Validity. J Clin Epidemiol. 2026. PMID: 41724217. doi:10.1016/j.jclinepi.2026.112199

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。