来自十个国家的研究团队发现，大多数健康应用和人工智能诊断工具在编码前就已埋下系统性偏见的隐患1。这些工具的训练数据代表性狭隘，反映西方中心的健康观念，边缘化社区缺乏实质参与1。
皮肤诊断算法的偏见问题尤为突出1。主要基于浅色皮肤训练的算法在近十年间已导致深色皮肤患者的明显误诊1。在美国医院，某风险评分算法因基于医疗支出数据而对黑人患者产生了严重低估1。仅18%被该算法分配额外护理的患者是黑人，但按实际需求应该超过46%1。这一差异的根本原因在于，黑人患者由于社会经济地位较低和对医疗系统的不信任，医疗支出往往更少，导致他们必须在病情更严重的情况下才会被标记为需要额外护理1。
研究团队指出，关键问题需要在编码前就被提出：什么算是健康？谁的知识被重视？哪些福祉形式被优先考虑？谁参与塑造这些技术？1 为了开发更具包容性和公平性的数字健康工具，研究建议在设计阶段就审视西方中心假设，纳入社区知识和数据主权1。数据主权概念要求社区能够管理其数据的收集、分析、链接、解释、共享和管理方式1。加拿大、澳大利亚和美国的研究表明，原住民更倾向采用能增强家庭和社区参与护理、促进相互依赖而非独立的技术1。
A research collaboration involving partners from ten countries has identified widespread bias embedded in health applications and artificial intelligence diagnostic tools before any code is even written.1 The bias stems from training data that narrowly represents populations primarily from Western perspectives, while excluding meaningful participation from marginalized communities.1
Skin diagnostic algorithms trained predominantly on light-skinned individuals have been documented over the past decade to produce misdiagnoses in patients with darker skin tones.1 Similarly, risk assessment algorithms used in American hospitals have demonstrated significant disparities—only 18 percent of patients assigned additional care resources are Black, despite Black patients comprising more than 46 percent of those who should receive such care according to need-based metrics.1 This discrepancy occurs because Black patients typically spend less on healthcare due to lower socioeconomic status and mistrust of medical systems, meaning they must reach more severe health conditions before being flagged as requiring additional support.1
The research highlights fundamental questions that must be addressed before development begins: What constitutes health? Whose knowledge is valued? Which forms of wellbeing are prioritized? Who participates in shaping these technologies?1 The concept of data sovereignty—enabling communities to control how their data is collected, analyzed, linked, interpreted, shared, and managed—is presented as essential to more equitable solutions.1 A study across Canada, Australia, and the United States found that Indigenous populations more readily adopt technologies that strengthen family and community involvement in care while promoting interdependence rather than independence.1
评论
还没有评论，欢迎留下第一条。