Trust Through Transparency: Developing a Stakeholder-Informed Explainability Framework for Clinician-Facing Medical Imaging AI Outputs
Artificial intelligence is increasingly being used to support healthcare decision-making, but its predictions can be difficult for healthcare professionals to interpret and trust. There is a need for clear and clinically meaningful explanations that help users understand how an AI system reached its conclusions without creating information overload. This project aimed to identify how AI-generated predictions and explanations can be presented to strengthen transparency, informed trust, and usefulness in clinical practice.
Background
This project focused on improving how AI-generated predictions are communicated to healthcare professionals. Two web-based prototype interfaces were developed and compared, presenting predictions from a previously trained glioma classification model through different formats and levels of explainability. The interfaces were designed to explore whether presenting model outputs more clearly and in a clinically relevant manner could improve understanding and support informed trust.
The prototypes were evaluated through user-testing interviews in which participants reviewed representative glioma cases, rated each interface for clarity and trust, and provided feedback on clinical usefulness, information overload, and workflow integration. Quantitative ratings and qualitative feedback were analyzed to identify preferred features, refine the prototype, and develop recommendations for designing transparent and user-centred clinical AI interfaces.
