The increasing use of artificial intelligence (AI) in medical decision support systems has intensified debate over a crucial design trade-off inherent to these systems: balancing model accuracy (performance) and explainability. Many guidelines and regulations worldwide emphasize the importance of transparency, human oversight, and explainability in health care decision support systems classified as high-risk applications. While prior research has documented laypeople's general preferences for explainable AI, many also want to sacrifice explainability for accuracy when faced with a trade-off between the two, especially in high-stakes and health care contexts. Less is known about how these preferences vary depending on the perspective individuals adopt when evaluating such systems. In an online survey experiment, we investigate whether people’s evaluations of the trade-off between explainability and accuracy differ when considering an AI-based risk assessment in clinical decision-making from the perspectives of a neutral observer, an affected patient, or an attending physician. In a quantitative study, participants read case studies on AI support systems based on health care scenarios where AI is nowadays advanced in its applications. We find that participants who adopt the perspective of an affected patient exhibit stronger preferences for accuracy over explainability than those who evaluate the situation from other perspectives. This study provides empirical evidence for ongoing discussions about trustworthy, human-centered AI in healthcare by clarifying how preferences for accuracy versus explainability depend on the social role from which individuals evaluate AI-mediated risk assessment. Findings have implications for the design and communication of AI technologies used in clinical practice.