Artificial Intelligence-Driven Clinical Decision Support for Postoperative Delirium Prevention: A Pragmatic Randomized Controlled Trial

BACKGROUND: Postoperative delirium affects up to 50% of older adults with cognitive impairment, yet preoperative identification of at-risk patients remains limited. Artificial intelligence methods, such as natural language processing, can extract cognitive impairment signals from electronic health records and enable targeted clinical decision support interventions to promote adherence to perioperative best practices and reduce postoperative delirium. We aimed to evaluate the effectiveness of an artificial intelligence-driven non-interruptive clinical decision support intervention in reducing postoperative delirium by promoting adherence to best practices among anesthesia teams caring for surgical patients with cognitive impairment. METHODS: We conducted a pragmatic randomized controlled trial embedded within routine clinical care at The Mount Sinai Hospital in New York City from June 20, 2023, to August 31, 2024, including patients with preoperative cognitive impairment identified immediately before surgery using a validated natural language processing algorithm, Mini-Mental State Examination scores, or diagnostic codes. Patients were randomized to an artificial intelligence-driven non-interruptive clinical decision support intervention within the electronic health record to promote adherence to best practices among anesthesia teams or routine perioperative care. The primary outcome was postoperative delirium, assessed with the 4 A's Test within seven days postoperatively. Secondary outcomes were adherence to 12 perioperative best practices. We used chi-square tests for primary and secondary outcome analyses. RESULTS: Overall, 7294 patients were identified as having cognitive impairment immediately before surgery and were randomized to the intervention (n = 3624) or control (n = 3670) group. In total, 1255/7294 (17.2%) patients had documented postoperative delirium assessments, of whom 426/1255 (34%) developed this condition. Between intervention and control groups, there were no statistically significant differences in POD incidence (225/641, 35.1% vs 201/614, 32.7%, P = .37) or best practice adherence (all P > .05). CONCLUSIONS: In this pragmatic randomized controlled trial, less than 20% of the sample had documented delirium assessments due to variability in routine POD screening practices, limiting our ability to conclude that artificial intelligence-driven clinical decision support was not effective in preventing postoperative delirium. However, based on our secondary outcome analyses, the intervention did not significantly change best practice adherence. Further work is needed to determine the best way to use these tools effectively.