CAREER: Improving Real-world Performance of AI Biosignal Algorithms

用于处理个体生物数据流（生物信号）的基于AI的算法是智能手表和医疗监护仪等可穿戴设备的基础技术，这些设备对日常生活中的健康监测至关重要。目前的算法，尽管它们在可穿戴设备中用于健康监测，但受到性能差异的阻碍，特别是在不同的人口统计群体中。该项目解决了围绕偏见数据和不断变化的技术或漂移的关键挑战，导致准确性降低，特别是对边缘化群体。该研究的重点是评估算法在不同人群、各种类型的测量中的表现，以及随着时间的推移和技术的更新。该项目的成果有望提高用于监测生物信号的人工智能技术的公平性和可靠性，为诊所外公平可靠的健康监测提供突破性方法。 研究计划主要有两个方面。第一个推力发展强大的技术，用于评估和报告算法性能跨交叉人群，特别关注涉及连续变量的回归任务。这包括描述现有生物信号训练数据集的人口统计特征，设计报告标准，以及在考虑交叉性的同时实施基于理论的算法公平性定量评估方法。实证分析将评估关键生物信号算法和数据集的交叉公平性。第二个重点是检测和监测生物信号数据和算法中随时间推移的概念漂移，考虑交叉人口统计学变化。这涉及开发概念漂移监测的方法和指标，提供对训练数据组成变化如何影响基于AI的生物信号算法性能的微妙理解。这项工作将获得有关偏差和漂移的基本知识，推进其检测和监测技术，并最终加强人工智能在生物信号算法中的公平和可靠应用，以改善健康结果。该项目的范围还扩展到推广和教育计划，促进更多地获得生物信号监测设备，并促进STEM领域的多样性。这种多方面的方法确保了研究的影响超越了理论上的进步，通过改进用于健康监测的可穿戴设备以及开发实现更普遍的人工智能监督的方法来直接造福社会。该项目反映了NSF的法定使命，将通过开发和教育值得信赖和公平的人工智能技术及其在生物信号算法中的应用，以广泛改善健康和健康，从而提供社会效益。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

AI-based algorithms for processing individual biological data streams (biosignals) are the enabling technology underlying wearables like smartwatches and medical monitors that are pivotal for health monitoring in everyday life. Current algorithms, despite their utility in wearables for health monitoring, are hindered by performance disparities, particularly among diverse demographic groups. This project addresses critical challenges surrounding biased data and constantly changing technologies, or drift, leading to reduced accuracy, especially for marginalized groups. The research focuses on evaluating how well algorithms perform across diverse people, various types of measurements, over time, and with technology updates. Outputs of this project promise to enhance the fairness and reliability of AI technologies for monitoring biosignals, offering breakthrough methods for equitable and reliable health monitoring outside of the clinic. The research plan unfolds in two primary thrusts. The first thrust develops robust techniques for assessing and reporting algorithm performance across intersectional populations, with a particular focus on regression tasks involving continuous variables. This includes characterizing existing biosignal training dataset demographics, designing reporting standards, and implementing a theory-based method for quantitative evaluation of algorithmic fairness while considering intersectionality. An empirical analysis will assess intersectional fairness on key biosignal algorithms and datasets. The second thrust focuses on detecting and monitoring concept drift over time in biosignal data and algorithms, accounting for intersectional demographic shifts. This involves developing methods and metrics for concept drift monitoring, providing a nuanced understanding of how changes in training data composition impact the performance of AI-based biosignal algorithms. This work will result in gaining fundamental knowledge about bias and drift, advancing techniques for their detection and monitoring, and, ultimately, enhancing the equitable and reliable application of AI in biosignal algorithms for improved health outcomes. The project's scope also extends to an outreach and education plan, promoting increased access to biosignal monitoring devices and fostering diversity in STEM fields. This multifaceted approach ensures that the impact of the research transcends theoretical advancements, directly benefiting society through the improvement of wearables for health monitoring as well as the development of methods to enable more general AI oversight. Reflecting NSF’s statutory mission, this project will provide societal benefits through development of and education on trustworthy and equitable AI technologies and their applications to biosignal algorithms to improve health and wellness broadly.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.