Designing Trust, Autonomy, and Acceptance in AI-Based Clinical Decision Support Systems: A Comparative Experimental Analysis

Abstract

Introduction: Artificial intelligence-based clinical decision support systems (AI CDSS) have the potential to improve clinical decision-making and patient outcomes, yet their adoption remains limited. Prior research highlights trust in AI and concerns about professional autonomy as key barriers to implementation. While explainability and system integration may foster trust, they may simultaneously threaten professional identity and decision authority. At the same time, perceived usefulness has been identified as the central driver of technology acceptance, although it is shaped by autonomyrelated concerns, particularly among physicians. This study integrates two experimental investigations (1,2) to examine how AI-CDSS design features jointly influence trust, autonomy-related perceptions, and acceptance across patients and physicians. Methods: The analysis combines and compares findings from two vignette-based experimental studies. The first study (n = 292) employs a 2 × 2 × 2 factorial design manipulating AI explainability, workflow integration, and accountability to assess their effects on trust and professional identity threat among physicians. The second study consists of three experiments across patients (n = 209), potential future users (n = 1062), and physicians (n = 185), manipulating AI-CDSS system type (hybrid vs. autonomous) and recommendation risk (lifestyle vs. medical) (2 x 2 factorial design). Regression-based mediation and sequential mediation models (SPSS PROCESS Macro) were applied using bootstrapping procedures. Results: Results from the first study show that explainability and workflow integration significantly increase trust in AI-CDSS. At the same time, explainability and system-induced accountability increase perceived professional identity threat. Trust is negatively associated with identity threat and partially mediates the relationship between explainability and identity threat, indicating that trust can mitigate but not fully offset these concerns. Results from the second study demonstrate that perceived usefulness is the strongest predictor of intention to use among both patients, potential future users, and physicians. However, the underlying mechanisms differ. Patients and potential future users evaluate AI-CDSS primarily based on perceived usefulness, whereas physicians’ evaluations depend on professional autonomy preservation. For physicians, hybrid systems increase intention to use only indirectly through reduced professional autonomy concerns and increased perceived usefulness. High-risk medical recommendations consistently increase perceived or anticipated professional autonomy concerns across all groups but do not directly reduce perceived usefulness or intention to use. Conclusion: AI-CDSS acceptance is shaped by the interplay of trust, perceived usefulness, and professional autonomy-related concerns. Design features exhibit ambivalent effects, simultaneously enabling and constraining adoption. Successful implementation requires aligning system design with professional expectations, particularly by preserving physician decision authority while demonstrating clear clinical value.