Grant: Project Grant - PA : Population and Public Health - Equitable AI

Machine learning (ML) is a form of artificial intelligence that has increasingly been used over the last decade in a variety of sectors, as computing power and the availability of data has increased exponentially. ML applies algorithms to vast amounts of data, and through doing so, are programmed to learn over time in an autonomous fashion. ML has begun to be applied to population-level data, both in health services and in public health. However, it is now well recognized that ML models can replicate biases related to numerous factors including that the data used to train models contains biases, the people developing the models bring their biases to their work, and ideas about causation are biased. As ML is applied in population health, including the COVID-19 pandemic, guidelines to identify and mitigate biases are needed. This project consists of two parts that will occur in tandem. Part 1: We will conduct a scoping review of the literature to identify ML models used in specific areas of population health and to examine whether and how biases were identified. Part 2: We will develop guidelines for model developers to identify and reduce bias in ML models used in population health. Our approach will be informed by standards from the Grading of Recommendations Assessment, Development and Evaluation (GRADE), the UK National Institute for Health and Care Excellence (NICE), and guidelines for statistical model development. We will engage a broad set of stakeholders in the guideline development process. Our diverse team includes national and international experts in computer sciences, statistical modeling, epidemiology, ethics, sociology, public health, and the social determinants of health. This project builds on previous work on public perspectives of AI, ML applied to predicting population health outcomes, and the ethics of AI. Our findings will support efforts to ensure AI applications contribute to reducing, rather than exacerbating, health inequities.