Human Categorization with “Dirty” Confounders in AI and ML Medical Models: The Role of Religion

Abstract

Aim: This study was conducted to evaluate the acceptance among healthcare practitioners and scientific researchers of the current official regulatory recommendations regarding the incorporation of human categorization through confounders, such as “Religion”, into AI and ML-based clinical research and healthcare settings. Materials and Methods: An anonymous online survey was conducted using the Telegram platform, where participants were asked a single question: "Do you consider the inclusion of Religious status in Artificial Intelligence and Machine Learning models justified from the perspective of medical ethics and science?" Respondents were provided with only two response options: "Yes" or "No." This survey was specifically targeted at international groups, focusing primarily on English and Russian-speaking clinicians and scientific researchers. Results: 134 unique individuals participated in the survey. The results revealed that two-third of the respondents (87 individuals) agreed that including Religion status as predictor in the ML and AI models is inappropriate. Conclusion: Two-thirds of healthcare practitioners and scientific researchers agree that categorizing individuals within healthcare settings based on their religion is inappropriate. Educational programs are needed to inform healthcare and scientific professionals that AI and ML applications should be built on unbiased and ethically appropriate predictors. ML is incapable of distinguishing individual human characteristics. Therefore, constructing healthcare AI and ML models based on confounders like religion is unlikely to aid in identifying the cause of or treating any pathology or disease. Moreover, the high conflict potential of this predictor may further deepen societal disparities.