Publikation
Prescribing a dose of transparency: a qualitative evaluation of AI explanations with cardiovascular healthcare professionals
Sophie Haas; Malte Högemann; Oliver Thomas
In: Sanad Aburass (Hrsg.). Frontiers in Artificial Intelligence, Vol. 9, No. 1830201, Pages 1-15, Frontiers Media SA, 2026.
Zusammenfassung
Introduction: Chronic diseases, particularly heart failure, are associated with high mortality rates and frequent hospital readmissions, thereby placing a significant burden on healthcare systems and patients. Recent advancements in AI-supported telemonitoring offer a means to address these challenges through the early detection of deterioration. Yet, the black-box nature of AI can undermine physicians’ trust, underscoring the importance of explainable AI (XAI).Methods: This exploratory study examines clinicians’ perceptions of XAI in a prototype early-warning scenario for imminent heart failure decompensation based on tabular laboratory data. Three exemplary explanations are compared in interviews with 15 German cardiovascular healthcare professionals using the explanation satisfaction scale: SHapley Additive Explanations (SHAP), Counterfactual Explanations (CFEs), and Anchors.Results: SHAP was widely preferred for its intuitive graphical representation. CFEs were valued as an action-oriented method although their suitability for deriving therapeutic steps requires careful constraint and interpretation. Meanwhile, Anchors were considered too superficial.Discussion: Our findings suggest that no single explanation method fully meets clinicians’ needs. Instead, layered explanation designs that combine quick overviews with optional deeper insights may better support trustworthy AI use in early-warning scenarios based on tabular laboratory data.
Projekte
- KardioInterakt - Interaktionstechnologien zur Patientenversorgung bei betreuungsintensiven Herzerkrankungen
- CRAI - Center of research and development of trustworthy AI applications for mid-sized companies