Detecting delirium with agents: implications of AI enabled screening for clinical care

This invited commentary refers to ‘AI Agent for Delirium Screening among Patients in Oncology and Cardiac Intensive Care Units: A Proof-of-Concept Study’ by Y. Zeng et al. European Journal of Cardiovascular Nursing. 2026; https://doi.org/10.1093/eurjcn/zvag029. Delirium remains one of the most common, distressing, and costly complications across the spectrum of clinical care.1 It is associated with increased morbidity and mortality, prolonged hospitalization, long-term cognitive decline, and substantial health system burden.2 Despite the availability of validated screening tools and clear guideline recommendations, delirium, particularly the hypoactive form, continues to be under-recognized in everyday practice. Nurses play a central role in detection, prevention, escalation, and treatment. The limitation lies not in clinical knowledge but in using tools, guidelines, and embedding them into everyday practice. In this context, the proof of concept study by Zeng et al. is both timely and welcome.3 The authors describe an artificial intelligence-enabled agent designed to support nurse-led delirium screening through appropriate tool selection, synthesis of clinical information, and generation of structured, guideline-concordant recommendations. Evaluated across a small but clinically diverse set of cases, the system demonstrated high concordance with expert judgement and produced outputs that are interpretable and actionable within nursing workflows. The principal concept of this work lies not in replacing clinical judgement, but in closing the gap between best practice and everyday practice. By integrating guidelines and validated tools into a single workflow, the agent targets inconsistent delirium screening and follow-through in routine care. In principle, such systems could help standardize nursing actions, prompt earlier recognition, and reduce missed/delayed responses, particularly in cardiovascular settings where delirium risk is amplified by acute illness, invasive monitoring, and physiological instability. Recent work in cardiac surgical populations has reinforced the clinical value of earlier and more structured delirium detection in these pathways.4 Importantly, the authors appropriately frame their findings as preliminary. A proof of concept study is a starting point, not an endpoint. Delirium is a fluctuating and context-dependent syndrome; performance in curated case vignettes must be followed by prospective evaluation in real-world clinical environments. Future studies will need to demonstrate not only accuracy, but also safety, usability, and impact on outcomes that matter to patients and healthcare professionals. A further point of caution relates to the application of guideline-based recommendations to individual patients. Evidence-based medicine depends not only on the availability of high-quality evidence, but on its conscientious and judicious integration with clinical context, patient values, and professional judgement. Haynes et al. has emphasized that clinical expertise lies precisely in this integrative step, with different weights applying to evidence, context, and values in each case.5 While guideline-based systems offer clear advantages in standardization and reliability, they risk obscuring uncertainty and individual trade-offs if applied mechanistically. This is particularly relevant in delirium care, where multimorbidity, frailty, fluctuating physiology, and competing priorities are common. Similar challenges have been described in decision making for older adults with complex illness, where the limits of guideline-driven care are most evident.6 Artificial intelligence-enabled decision support should therefore be designed to make uncertainty explicit, support shared decision making, and preserve space for clinical discretion. Equally critical is how such systems are implemented. Artificial intelligence tools should function as decision support rather than decision makers. Experience from co-designed digital care pathways suggests that alignment with clinical workflow and professional roles is central to safe and sustainable adoption.7 When implemented with appropriate guardrails, agent-based systems have the potential to strengthen rather than undermine nursing expertise. Delirium has long been recognized as an important clinical problem within cardiovascular care, yet translation of evidence into consistent practice remains challenging. This study represents an encouraging step towards bridging that gap by aligning digital innovation with nursing workflows and guideline-based care. In delirium management, the value of artificial intelligence lies not in innovation alone, but in supporting reliable, consistent, and context-sensitive bedside practice. Louis John Koizia (Conceptualization [lead]; Supervision [equal]; Writing—original draft [lead]), Vincent E. S Allott [Writing—review & editing (equal)], Benjamin Howell Lole Harris (Supervision [lead]; Writing—review & editing [lead]) B.H.L.H. would like to thank the Orthogeriatric Research Fund for their support. This commentary does not contain new data.