Related Experiment Videos
Developing a Theory-Driven Construct of Clinical Usefulness for Evaluating Post-Hoc Explainable AI: Foundational Work
Jenny M Bauer1, Martin Michalowski1
1School of Nursing, University of Minnesota, Minneapolis, Minnesota, USA.
Rationale And Objective:
Model-centered evaluations of explainable artificial intelligence (XAI) rarely account for how clinicians interpret and use explanations during clinical decision-making. This study advances a novel method for systematic inquiry into clinical practice by defining and operationalizing clinical usefulness (CX), a theory-driven, clinician-centered construct that reflects how healthcare providers evaluate the usefulness of AI-generated explanations in context. Guided by pragmatic explanation theory and contemporary explainability scholarship, this work represents the foundational step of a multi-phase instrument development program.
Methods:
Following the MEASURE framework, thematic analysis was used to develop the construct and identify its core dimensions. Qualitative findings were systematically transformed into quantitative indicators through item-generation procedures that mapped thematic codes to measurable evaluative judgments. An international panel of experts reviewed the resulting dimensions and item pool, providing structured feedback on item clarity and relevance. Member checking supported interpretive alignment among the qualitative themes, construct definition, and item wording. Content validity was assessed using the Content Validity Index.
Results:
Analysis revealed three interdependent dimensions clinicians use when evaluating XAI explanations: comprehension, evidence support, and case relevance. These dimensions reflect clinicians' pragmatic reasoning about whether an explanation is clinically succinct, substantiated, and insightful. Qualitative themes were operationalized into an initial set of quantitative items representing the latent construct. Expert ratings were used to refine the pool to 17 items demonstrating strong content validity.
Conclusion:
This study establishes the theoretical foundation and initial operationalization of a clinician-centered construct for evaluating XAI explanations. It fills a critical gap by offering a theory-driven, contextually grounded approach to explanation evaluation, and provides the basis for subsequent psychometric testing in a multi-phase instrument development program.
