Using simulated clinical decision-making to evaluate the benefits of AI tools

Abstract

AI-enabled clinical decision-support (CDS) tools are information-based interventions that must demonstrate capacity to improve care delivery and patient outcomes. The information value chain provides a useful framework that connects use of CDS to clinical outcomes. To be beneficial CDS needs to yield new and pertinent information, that information then needs to improve clinical decisions made and in turn improve the healthcare delivered, with the goal of improving outcomes.   Studies evaluating the effect of AI-enabled CDS generally assess performance in terms of what the AI does, for example, detecting pathology in imaging, rather than impact on clinical decisions, such as diagnosis and patient management.   This paper discusses the use of simulated clinical decision-making as a method to directly test the impact of CDS on clinical decision-making, a critical precursor for CDS to impact patient outcomes. Simulated clinical decision-making studies provide greater experimental control than what is possible in real-world clinical settings in a patient- and risk-free environment, and can also be conducted online. Such studies allow testing of critical patient safety factors that are not feasible or ethical to test in clinical settings, such as the risk of false-positive and false-negative AI results that could adversely impact critical decisions and harm patients. We showcase three studies testing the impact of CDS on clinical decisions for prescribing by medical students; for diagnosis by emergency doctors and medical students interpreting imaging; and risk assessment for cardiovascular disease by general practitioners.