Scholarlink

<p>Current governance of clinical artificial intelligence (AI) often emphasizes downstream auditing, which can allow ethically consequential design choices to become entrenched before they are openly deliberated. We propose the Phase 1 Ethical Review (P1-ER) framework as an inception-stage approach for structured upstream ethical scrutiny of clinical machine learning projects. P1-ER combines a tiered trigger screen (routing higher-risk projects to comprehensive review) with three modular assessments: Clinical Need Alignment Score (CNAS), Stakeholder Diversity Index (SDI), and Data-Value Mapping Schema (DVMS). A dual assessment rule pairs quantitative scores with evidence grades (A/B/C) to distinguish weak ethical performance from documentation gaps and to support iterative revision rather than binary gatekeeping. We present two illustrative retrospective diagnostic applications, structured as counterfactual “thought-experiment” reviews based on publicly available documentation (a stroke imaging tool and a pharmacogenetic dosing algorithm). These examples illustrate the diagnostic logic by which P1-ER would be expected to draw attention to upstream failure modes such as proxy-outcome gaps, undocumented value trade-offs, and equity risks, while also showing how documentation gaps propagate through the evidence-grade rule. P1-ER is a conceptual, implementation-facing prototype; empirical evaluation (e.g., pilot implementation, inter-rater reliability assessment, and workload measurement) is needed to assess feasibility, reliability, and institutional fit.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Moving toward upstream governance: a Phase 1 Ethical Review framework for clinical machine learning

  • 期刊论文

  • 发表日期：

    2026年4月14日

查看全文
  • Qiqi Jiang,
  • Jian Xu,
  • Lejun Li,
  • Liefu Ye,
  • Liang Lin,
  • Jinfeng Wu,
  • Yupeng Wu

摘要

Current governance of clinical artificial intelligence (AI) often emphasizes downstream auditing, which can allow ethically consequential design choices to become entrenched before they are openly deliberated. We propose the Phase 1 Ethical Review (P1-ER) framework as an inception-stage approach for structured upstream ethical scrutiny of clinical machine learning projects. P1-ER combines a tiered trigger screen (routing higher-risk projects to comprehensive review) with three modular assessments: Clinical Need Alignment Score (CNAS), Stakeholder Diversity Index (SDI), and Data-Value Mapping Schema (DVMS). A dual assessment rule pairs quantitative scores with evidence grades (A/B/C) to distinguish weak ethical performance from documentation gaps and to support iterative revision rather than binary gatekeeping. We present two illustrative retrospective diagnostic applications, structured as counterfactual “thought-experiment” reviews based on publicly available documentation (a stroke imaging tool and a pharmacogenetic dosing algorithm). These examples illustrate the diagnostic logic by which P1-ER would be expected to draw attention to upstream failure modes such as proxy-outcome gaps, undocumented value trade-offs, and equity risks, while also showing how documentation gaps propagate through the evidence-grade rule. P1-ER is a conceptual, implementation-facing prototype; empirical evaluation (e.g., pilot implementation, inter-rater reliability assessment, and workload measurement) is needed to assess feasibility, reliability, and institutional fit.