Search

Machine learning model deployment is on the rise in clinical settings, but clinician-specific design objectives are not generally taken into account in the development of visualizations of model outputs. We are motivated by Tonekaboni et al’s findings that to increase the actionability of such integrations, we must increase trust in the model. We explore how to visually communicate to clinicians the uncertainty associated with a model’s outputs, using Tonekaboni et. al’s cardiac arrest risk model, implemented using SickKids Hospital Critical Care Unit’s T3 system, as a case study in design approach for visualizing predicted risk, and prediction confidence in an actionable, trustworthy way. We find that whether increased actionability is achieved can be determined by whether our end-users see the relevance of predictions in their clinical context, and that it is critical to employ a co-design approach when developing both these systems and their visualizations. We query three visualization settings: errorCloud, errorBlob, errorFade, and two uncertainty settings: Low, and High.