Artificial intelligence may eventually help perfusionists organize information and evaluate changing conditions during cardiopulmonary bypass. A human-centered design study by Seelam and colleagues asks a more immediate question: what should that support look like to the clinicians using it? Participants favored understandable explanations, consolidated patient and team information, customizable displays, and continued control over decisions.
Designing decision support around perfusion practice
Perfusionists monitor multiple physiological and circuit parameters while coordinating with surgery, anesthesia, and nursing. Information may be spread across monitors, charts, and conversations. Long procedures, competing alarms, interruptions, and documentation demands can make this work more difficult.
The investigators explored how a future clinical decision support system (CDSS) could fit those demands. The study focused on interface design and user requirements. It did not develop, deploy, or validate a real-time AI model.
How the study was conducted
Eight practicing perfusionists, aged 27–59 years, participated across three overlapping phases. Recruitment used convenience sampling through professional networks and social media. Six participants completed exploratory interviews, four contributed to prototype-feedback sessions, and five took part in usability testing. These phase counts overlap and should not be added together as separate groups.
Interviews examined monitoring strategies, decision-making, communication, and information needs. Researchers then used structured feedback to refine a prototype and compare three explanation formats: decision trees, probability charts, and counterfactual examples.
The final interactive prototype was assessed through 15 tasks within two simulated surgical scenarios. Tasks included finding patient information, interacting with recommendations, and requesting explanations. Recommendations and their explanations were manually constructed by clinical experts using historical surgical data and clinical expertise; they were not predictions generated by an operational AI system.
What perfusionists wanted from the interface
Explanations that can be inspected quickly
Participants preferred decision-tree explanations among the three formats evaluated. They requested visual highlighting of the relevant decision path, support for more complex branching, and access to explanations during a procedure and afterward for review. This preference identifies a promising design direction; it does not establish that decision trees improve clinical decision accuracy.
Patient and team information in one place
Participants wanted consolidated access to comorbidities, surgical history, medications, allergies, and both historical and day-of-procedure laboratory values. They also valued a timestamped event feed describing activities across perfusion, surgery, anesthesia, and nursing, with filtering and color cues to help locate relevant events.
Such features could support situational awareness and debriefing, but their effects on those outcomes were not directly measured. Information consolidation, reminders, and event tracking also do not necessarily require AI.
Personalization and clinician control
Institutional protocols, patient context, and user experience influence how information is interpreted. Participants therefore requested configurable thresholds, alerts, and information priorities. Recommendations should remain advisory, with clear options to accept, reject, dismiss, question, or review them.
Participants were generally receptive to decision support, while emphasizing that their own reasoning and professional judgment must remain central. Mandatory interaction with every suggestion could create additional workload during time-sensitive tasks.
Usability findings
The five participants in the final phase independently completed 84% of the predefined prototype tasks. The most frequent difficulties involved filtering the live event feed and switching between surgery and review modes.
The 84% figure measures interface task completion. It is not an AI accuracy rate, a measure of treatment success, or evidence that patient outcomes improved. Exploratory survey responses suggested participants anticipated feeling more comfortable and relaxed than distracted, anxious, or annoyed when using the prototype. These anticipated responses should not be interpreted as validated reductions in cognitive workload.
Limitations and unanswered safety questions
The small convenience sample limits generalizability. Uncompensated participation may have favored clinicians with a particular interest in research or AI. The study evaluated simulated recommendations and anticipated use, rather than performance during live cardiopulmonary bypass.
It did not establish how clinicians respond to incorrect recommendations, how trust changes with system reliability, or whether users can recognize and recover from automation errors. It also did not demonstrate improvements in patient safety, clinical outcomes, decision accuracy, or workload. Usability errors were not evaluated through a formal failure-analysis framework, and subjective measures were exploratory.
Further work requires operational models, broader evaluation, validated measures, and testing of human–system behavior under realistic conditions, including unreliable outputs. Positive perceptions of a prototype are an early development finding rather than evidence of clinical readiness.
Relevance for perfusion teams
The study offers a practical set of questions for evaluating future tools: Can the perfusionist understand the rationale? Does the interface fit ongoing monitoring? Can information and alerts be adapted to local practice? Does the clinician retain control? Can recommendations and actions be reviewed after the case?
Its central contribution is a set of user-informed design requirements for future decision support. Clinical benefit and safe real-world deployment remain to be established.
Source: Seelam L, Yang S, Chen L, et al. Human-Centered design requirements for clinical decision support in cardiac surgery perfusion. Human Factors in Healthcare. 2026;10:100152. doi:10.1016/j.hfh.2026.100152. Open-access source.
Funding and disclosures: Supported by NIH grant R01HL157457. The authors disclose research funding and consulting or honoraria relationships as detailed in the source manuscript. The original article is available under the Creative Commons Attribution (CC BY 4.0) license. This IPA educational analysis is a paraphrased summary; the downloadable PDF is the supplied author manuscript.
