A scoping review of robustness concepts for machine learning in healthcare
Alan Balendran, Céline Béji, Florie Bouvier, Ottavio Khalifa, Theodoros Evgeniou, Philippe Ravaud, Raphaël Porcher
Abstract
While machine learning (ML)-based solutions-often referred to as artificial intelligence (AI) solutions-have demonstrated comparable or superior performance to human experts across various healthcare applications, their vulnerability to perturbations and stability to variations due to new environments-essentially, their robustness-remains ambiguous and often overlooked. In this review, we aimed to identify the types of robustness addressed in the literature for ML models in healthcare. A total of 274 eligible records were retrieved from PubMed, Web of Science, IEEE Xplore, and additional sources. Eight general concepts of robustness emerged. Furthermore, an analysis of those concepts across types of data and types of predictive models revealed that the concepts were differently addressed. Our findings offer valuable insights for stakeholders seeking to understand and navigate the robustness of machine learning models during their development, validation, and deployment in healthcare settings, where interpretation of robustness may vary.