Use Explainable AI to Improve the Trust of and Detect the Bias of AI Models

项目概要/摘要 AD/ADRD 是一个日益严重的全国性公共卫生危机，因为 65 岁以上的美国人数量 预计到 2050 年将翻一番。我们的家长补助金旨在衡量心肺功能 在最广泛的流行病学研究中，健身（CRF）作为体力活动的生物标志物 近 100 万退伍军人使用 VA 世界一流的电子健康记录和先进的人工 情报技术。家长补助金的目的是（1）确定之间的关系 CRF 和事件 AD/ADRD，考虑非线性关系和潜力 CRF 与其他风险因素的相互作用，以及 (2) 定义相关的增量 CRF 水平 逐步降低总体上以及按年龄、性别和种族划分的亚组中 AD/ADRD 的风险。我们的目标 三是开发并验证基于深度学习的风险预测模型，以确定最佳风险预测模型 CRF 水平使个人实现 AD/ADRD 风险最低。深度学习是关键的人工智能 智能（AI）技术。人工智能在过去十年中取得了巨大进步。然而，人工智能 模型通常被视为“黑匣子”，因为它们难以解释。了解什么是人工智能 模型确实是人工智能使用道德的先决条件，因为利益相关者不能信任模型或 检测与开发相关的潜在有意和无意的偏见或 在不理解模型的情况下使用模型。我们相信可解释的人工智能是一个强大的工具 解决信任和偏见的生物伦理问题。可解释人工智能的目的是使其 人类用户有可能理解并信任由以下人员提供的决定或建议 AI 模型，并对其进行调试和完善。具体来说，本次补充将测试AI的效果 通过招募一组在模拟环境中进行信任和偏见检测的模型解释 利益相关者并使用基于场景的方法。拟议的潜在广泛影响 其工作在于它将促进人工智能/机器学习在生物医学和行为领域的道德发展和使用 使用可解释的人工智能方法进行科学。

Project Summary/Abstract AD/ADRD is a growing national public health crisis as the number of Americans ≥65 years is projected to double by 2050. Our parent grant was designed to measure cardiorespiratory fitness (CRF) as a biomarker of physical activity in the most extensive epidemiological study in nearly 1 million Veterans using VA’s world-class electronic health record and advanced artificial intelligence technologies. The parent grant aims are (1) to determine the relationship between CRF and incident AD/ADRD, taking into consideration a non-linear relationship and potential interactions of CRF with other risk factors and (2) to define incremental CRF levels that are linked to progressively lower risk of AD/ADRD, overall, and in subgroups by age, sex, and race. Our Aim 3 is to develop and validate a deep learning-based risk prediction model to determine the optimal CRF level for individuals to achieve the lowest risk of AD/ADRD. Deep learning is a key Artificial intelligence (AI) technique. AI has demonstrated great strides in the past decade. However, AI models are often viewed as “black box” as they are difficult to explain. Understanding what an AI model does is a prerequisite to the ethical use of AI, because stakeholders can’t trust a model or detect the potentially intended and unintended biases associated with the development or utilization of the model without understanding it. We believe that explainable AI is a powerful tool to address the bioethics issues of trust and bias. The purpose of explainable AI is to make it possible for human users to understand and trust the decisions or recommendations offered by the AI model, and to debug and refine it. Specifically, this supplement will test the effect of AI model explanation on trust and bias detection in a simulated environment by recruiting a set of stakeholders and using a scenario-based approach. The potential broad impact of the proposed work is that it will advance the ethical development and use of AI/ML in biomedical and behavioral sciences using explainable AI methods.