Stakeholder Guidance to Anticipate and Address Ethical Challenges in Applications of Machine Learning and Artificial Intelligence in Algorithmic Medicine: a Novel Empirical Approach

项目摘要 人工智能应用的潜力，特别是机器学习，以预防，预测和帮助 控制疾病不仅为受影响的个人，而且为整体健康带来了巨大的希望。 人口数量。这些新颖的计算策略的特别令人兴奋的例子越来越多地出现在 开发用于医疗用途的深度学习算法。算法已经嵌入我们的日常生活， 开始影响人类的决策，从招聘和雇用员工到刑事判决。 在医学之外，对算法可能反映、复制和延续偏见的方式的认识导致了 关于这个问题的理论和实证研究的激增。人们越来越意识到 潜在的算法弱点，包括一些引起人们对公平性基本问题的担忧， 公正和偏见预测和解决算法医学中新出现的伦理问题的需要是时间- 敏感随着医疗保健系统越来越多地利用算法来进行患者识别、诊断和诊断， 治疗方向，算法偏差的后果产生真实的和显著的成本。许多 利益相关者负责医学算法的开发、应用和解释， 然而，受这些学习系统和工具影响最大的利益攸关方很少参与。 这个经验和假设驱动的项目的总体目标是阐明道德的景观 在机器的开发、改进和应用方面出现的问题和问题 学习算法医学。首先，我们确定了不同的伦理问题和遇到的问题， 机器学习的发展，完善和应用，通过查询不同的观点， 一系列利益相关者参与-机器学习研究人员，临床医生，伦理学家和患者。使用 从上半年产生的新见解，我们将进行一个以证据为基础的信息共享小插曲 调查，以了解算法的背景对道德上突出的观点， 医生-那些准备在他们自己的决策中实施这种创新的人，以照顾病人。 最大限度地发挥我们在经验道德调查方面的专业知识，这一系列项目 利用斯坦福大学进行的卓越机器学习研究，包括 由NIH资助的研究人员开展工作，并提供广泛的、系统收集的有关伦理问题的数据 在整个算法的开发和实现过程中遇到和预期的。最后 项目制定并完善了一项循证信息共享调查，用于更好地了解 医生如何对智能系统做出反应。

PROJECT ABSTRACT The potential for artificial intelligence applications, specifically machine learning, to prevent, predict, and help manage disease sparks immense hope not only for the individuals affected, but also for the overall health of populations. Particularly exciting examples of these novel computing strategies are increasingly found in the development of deep learning algorithms for medical use. Already embedded in our daily lives, algorithms have begun to impact human-decision making, from recruitment and hiring of employees to criminal sentencing. Outside of medicine, recognition of the ways algorithms may reflect, reproduce, and perpetuate bias has led to an explosion of theoretical and empirical research on the subject. There is an increasing awareness of potential algorithmic weaknesses, including some that raise concerns about fundamental issues of fairness, justice, and bias. The need to anticipate and address emerging ethical issues in algorithmic medicine is time- sensitive. As health care systems increasingly utilize algorithms for patient identification, diagnosis, and treatment direction, the consequences of algorithmic bias yield real and significant costs. Numerous stakeholders are responsible for the development, application and interpretation of algorithms in medicine, and yet there has been very little engagement of stakeholders most affected by these learning systems and tools. The overarching goal of this empirical and hypothesis driven project is to articulate the landscape of ethical concerns and the issues emerging in the context of the development, refinement, and application of machine learning in algorithmic medicine. First, we determine the distinct ethical issues and problems encountered in the development, refinement, and application of machine learning, by querying the perspectives of a diverse array of stakeholders involved—machine learning researchers, clinicians, ethicists, and patients. Using the new insights generated from the first half, we will conduct an evidence-based, information-sharing vignette survey to understand the impact of the contexts of algorithms on the ethically salient perspectives of physicians—those poised to implement such innovation in their own decision-making for the care of patients. Maximizing our established record of expertise in empirical ethics investigations, this sequence of projects leverages access to the exceptional machine learning research conducted at Stanford University, including work by NIH-funded investigators, and provides extensive, systematically collected data on ethical issues encountered and anticipated throughout the development and implementation of algorithms. Finally, the project develops and refines an evidence-informed information-sharing survey for use in better understanding how physicians react to intelligent systems.