Fairness of Medical Artificial Intelligence: A Literature Review : Emerging Trends in Internet Technologies, Summer Term 2021

Background: With the increasing digitization and growing integration of Artificial Intelligence (AI) in healthcare (e.g. for diagnosis), the risk of incorporating unfair and biased behaviours in medical applications caused by biased AI is rising. Thus, the integration of AI in healthcare requires the need to develop trustworthy and unbiased applications that make fair decisions regardless of a patient's demographic profile.
Objective: To contribute to the development of fair and unbiased AI applications in healthcare, the aim of this paper is to provide an overview of the current state of research in the emerging research field on fairness of medical AI. In particular, it aims to elaborate issues and causes related to the occurrence of biased AI and provide possible solutions.
Methods: To provide the overview of the extant literature, a systematic literature review was conducted. By searching in eight scientific databases (ACM Digital Library, AIS Electronic Library, arXiv, IEEE Explore, Ebsco Buisness Source, medRxiv, Pubmed,
Scopus) and applying inclusion and exclusion criteria, 35 relevant articles have been identified and further analyzed.
... mehr
Results: The emergence of biased AI applications cannot be considered solely as a data or a system problem. Accordingly, the occurrence of unfair AI behaviors has different reasons. For example, under- or over-representation of a particular group in an AI's
training data can lead to bias, as it is not a baseline that is representative of the entire population. Unclear instructions to an algorithm lead to incorrect process steps. In addition, the black-box nature of many AIs makes it difficult to identify problem areas
due to a lack of transparency. By learning and forming false correlations, models also cause biases. In addition, models must make a trade-off between fairness, model performance, and model explainability, which make the development of a fair, powerful,
and transparent AI an NP-hard optimization problem. The interfaces of developers and contributors as well as users with the AI are equally a cause of the emergence of unfair medical AI.
Conclusion: Developing AI for healthcare applications is a complex problem that must holistically consider and optimize the interplay of bias as a data and system problem.