As if sand were stone. New concepts and metrics to probe the ground on which to build trustable AI

<jats:title>Abstract</jats:title><jats:sec><jats:title>Background</jats:title><jats:p>We focus on the importance of interpreting the quality of the labeling used as the input of predictive models to understand the reliability of their output in support of human decision-making, especially in critical domains, such as medicine.</jats:p></jats:sec><jats:sec><jats:title>Methods</jats:title><jats:p>Accordingly, we propose a framework distinguishing the reference labeling (or Gold Standard) from the set of annotations from which it is usually derived (the Diamond Standard). We define a set of quality dimensions and related metrics: representativeness (are the available data representative of its reference population?); reliability (do the raters agree with each other in their ratings?); and accuracy (are the raters’ annotations a true representation?). The metrics for these dimensions are, respectively, the<jats:italic>degree of correspondence</jats:italic>,<jats:italic>Ψ</jats:italic>, the<jats:italic>degree of weighted concordance</jats:italic><jats:italic>ϱ</jats:italic>, and the<jats:italic>degree of fineness</jats:italic>,<jats:italic>Φ</jats:italic>. We apply and evaluate these metrics in a diagnostic user study involving 13 radiologists.</jats:p></jats:sec><jats:sec><jats:title>Results</jats:title><jats:p>We evaluate<jats:italic>Ψ</jats:italic>against hypothesis-testing techniques, highlighting that our metrics can better evaluate distribution similarity in high-dimensional spaces. We discuss how<jats:italic>Ψ</jats:italic>could be used to assess the reliability of new predictions or for train-test selection. We report the value of<jats:italic>ϱ</jats:italic>for our case study and compare it with traditional reliability metrics, highlighting both their theoretical properties and the reasons that they differ. Then, we report the<jats:italic>degree of fineness</jats:italic>as an estimate of the accuracy of the collected annotations and discuss the relationship between this latter degree and the<jats:italic>degree of weighted concordance</jats:italic>, which we find to be moderately but significantly correlated. Finally, we discuss the implications of the proposed dimensions and metrics with respect to the context of Explainable Artificial Intelligence (XAI).</jats:p></jats:sec><jats:sec><jats:title>Conclusion</jats:title><jats:p>We propose different dimensions and related metrics to assess the quality of the datasets used to build predictive models and Medical Artificial Intelligence (MAI). We argue that the proposed metrics are feasible for application in real-world settings for the continuous development of trustable and interpretable MAI systems.</jats:p></jats:sec>