Addressing bias in big data and AI for health care: A call for open science

Summary

Artificial intelligence (AI) has an astonishing potential in assisting clinical decision making and revolutionizing the field of health care. A major open challenge that AI will need to address before its integration in the clinical routine is that of algorithmic bias. Most AI algorithms need big datasets to learn from, but several groups of the human population have a long history of being absent or misrepresented in existing biomedical datasets. If the training data is misrepresentative of the population variability, AI is prone to reinforcing bias, which can lead to fatal outcomes, misdiagnoses, and lack of generalization. Here, we describe the challenges in rendering AI algorithms fairer, and we propose concrete steps for addressing bias using tools from the field of open science.

The bigger picture

Bias in the medical field can be dissected along three directions: data-driven, algorithmic, and human. Bias in AI algorithms for health care can have catastrophic consequences by propagating deeply rooted societal biases. This can result in misdiagnosing certain patient groups, like gender and ethnic minorities, that have a history of being underrepresented in existing datasets, further amplifying inequalities.

Open science practices can assist in moving toward fairness in AI for health care. These include (1) participant-centered development of AI algorithms and participatory science; (2) responsible data sharing and inclusive data standards to support interoperability; and (3) code sharing, including sharing of AI algorithms that can synthesize underrepresented data to address bias. Future research needs to focus on developing standards for AI in health care that enable transparency and data sharing, while at the same time preserving patients’ privacy.

Abstract

Artificial intelligence (AI) has an astonishing potential in revolutionizing health care. A major challenge is that of algorithmic bias. Most AI algorithms need big datasets to learn from, but several groups of the human population are absent or misrepresented in existing datasets. AI is thus prone to reinforcing bias, which can lead to fatal outcomes and misdiagnoses. Here, we describe challenges in rendering AI algorithms fairer, and we propose concrete steps for addressing bias using open science tools.