Human-Centered design requirements for clinical decision support in cardiac surgery perfusion

Perfusionists manage critical physiological parameters of cardiac surgery patients during cardiopulmonary bypass (CPB), requiring rapid, high-stakes decision-making under uncertainty. While artificial intelligence (AI) has the potential to enhance perfusionists' decision-making and improve patient safety, there is limited understanding of how future AI-enabled clinical decision support systems (CDSS) should be designed to align with perfusionists' workflows and cognitive demands.

This study aims to identify human-centered design requirements for CDSS in cardiac perfusion and to conduct a formative evaluation of a prototype interface presenting simulated decision-support recommendations.

A three-phase, multiple-method qualitative study was conducted with eight practicing perfusionists. Phase 1 involved semi-structured interviews to characterize workflow challenges and design requirements. Phase 2 engaged participants in structured prototype-feedback sessions to refine an interface concept for future AI-enabled decision support. Phase 3 assessed interface usability using scenario-based tasks with simulated AI recommendations designed by clinical experts and measured task success and exploratory self-reported perceptions.

Overall, perfusionists expressed positive perceptions of future AI-enabled CDSS concepts but emphasized the need to retain human control and receive contextually relevant explanations. Among three explanation formats evaluated, decision trees emerged as the most preferred explanation format across the study participants. In usability testing, participants achieved an 84% task success rate, reported positive perceptions of the prototype and perceived compatibility with perfusion workflows.

This work contributes empirically grounded design requirements for future AI-enabled CDSS in safety-critical intraoperative settings. Key design considerations include: (1) prioritizing interpretable explanations (such as decision trees), (2) ensuring system recommendations align with perfusionists' reported monitoring strategies, and (3) preserving clinician oversight through human-in-the-loop design frameworks.