AI Agent for Delirium Screening among Patients in Oncology and Cardiac Intensive Care Units: A Proof-of-Concept Study

To develop and evaluate an autonomous artificial intelligence (AI) agent to support nurse-led delirium screening and guideline-concordant prevention and management.

We constructed a delirium-specific knowledge graph from publicly available clinical guidelines and implemented an autonomous AI agent that integrates retrieval-augmented generation with validated delirium assessment tools to simulate decision-support in nursing workflows. In this proof-of-concept evaluation, the agent was benchmarked on 20 clinical patient cases to assess (i) tool selection accuracy, (ii) fidelity of clinical conclusions, and (iii) adherence to guideline-based delirium care recommendations. Human experts review rated coherence, relevance, and clinical interpretability of the outputs. The 20 cases were drawn from postoperative care (n=6), Cardiac Intensive Care Unit care (n=5), and cancer wards (n=9), including 6 males (30%) and 14 females (70%), with a mean age of 56.5 years (SD 17.9; range 21-82). Two experienced registered nurses independently validated the agent's outputs, benchmarked them against guideline-based recommendations. The agent achieved 100% accuracy in selecting appropriate tools, and 90% overall accuracy in generating conclusions and care recommendations, compared with 45% for other large language model baselines. Recommendations were presented in a structured, actionable format and aligned with guideline-based delirium care.

This proof-of-concept study suggests that an autonomous AI agent can deliver clinically interpretable, guideline-aligned delirium decision support and may help reduce missed or delayed recognition while standardizing nursing actions. Given the high delirium burden in cardiovascular pathways (e.g., cardiac ICUs and postoperative care), prospective validation in cardiovascular settings is warranted to evaluate clinical impact, safety, and workflow integration.