Mitigating Bias in Clinical Machine Learning Models

Purpose of review: Identifying the risk for and addressing bias in clinical machine learning models is essential to reap its full benefits and ensure health equity. We provide a review of the machine learning landscape in clinical medicine, highlight ethical concerns with a particular focus on algorithmic bias, and offer a framework for mitigating bias. Recent findings: Machine learning, the computational framework that supports artificial intelligence, now plays a significant role in everyday life and its potential role in clinical medicine continues to increase exponentially. Multiple machine learning models have demonstrated outstanding performance, surpassing human abilities with specific tasks, and are poised to revolutionize clinical research and practice over the next few years. While machine learning can augment clinician’s diagnostic capabilities, support clinical decision-making, and improve health care efficiency, they are not infallible. One key concern with the use of machine learning models is algorithmic bias, which if present poses a non-trivial risk to patient care particularly if algorithms are used in a population different from that used to create the algorithm. Recommendations and methods to identify and mitigate algorithmic bias to ensure responsible development of machine learning models are summarized. Summary: With the anticipated widespread adoption of machine learning in medicine, significant ethical concerns remain, particularly the risk for bias. Researchers, model developers, and end users need to be aware of the potential for bias, its associated risk, and methods to guard against it prior to deploying it for clinical use.