Title : Analysis of intersectional bias in artificial intelligence in health care
Abstract:
With the rising use of artificial intelligence in health care, it is becoming increasingly important to ensure that machine learning algorithms developed to recommend disease management be free of bias. Existing research has explored bias in a variety of machine learning applications in healthcare, including classification of melanoma in skin tones and in interpretation of diagnostic images (Larrazabel, 2020, Seyyed-Kalantari, 2021). Intersectional bias occurs when populations do not fit into discrete data sets, such as race, gender, class, or disability. Further, intersectional bias can be extended to target multi-comorbid patients who lie outside a discrete data set and therefore, not be captured in an algorithm. As a result, intersectional bias has resulted in under diagnosis of certain diseases (Seyyed-Kalantaru, 2021). Few attempts have been made to determine the consistency of bias occurring in the classification of patients who have overlapping characteristics that do not fit into a discrete data set. Health care providers need to be equipped to understand how an AI algorithm may not capture an individual who lies outside the parameters of a specific data set. The purpose of this paper is to provide an overview of artificial intelligence in healthcare. The derivation of a data set, awareness of intersectional bias and strategies to mitigate bias will be proposed. We conclude that machine learning algorithms need to use more inclusive databases in training so as to not reinforce implicit bias.
