Physician in the Loop
For youSign in
Anesthesiology ·

AI decision support shows no significant delirium reduction in Mount Sinai trial; GPT-4.5 falls short on nerve block ultrasound targets

An anesthesia face mask with a clear shell and cushioned rim, joined by corrugated breathing tubing to an amber reservoir bag.
THE SHORT READ

An AI-driven decision support tool for anesthesia teams did not significantly reduce postoperative delirium in a 7,294-patient Mount Sinai trial in Anesthesia & Analgesia, though fewer than one in five patients were assessed for delirium. Adherence to best practices also did not differ significantly. A separate study found that nerve block ultrasound targets drawn by GPT-4.5 showed limited agreement with expert consensus.

THE DETAILS
  • AI Oct 9, 2026
    AI-driven decision support shows no significant delirium reduction in 7,294-patient pragmatic trial at Mount Sinai
    At The Mount Sinai Hospital in New York City, 7,294 surgical patients with cognitive impairment identified before surgery, in part by a natural language processing algorithm, were randomized to a noninterruptive AI-driven clinical decision support intervention for anesthesia teams or routine care, according to Anesthesia & Analgesia. Only 1,255 (17.2%) had documented delirium assessments; among them, delirium incidence did not differ significantly (35.1% versus 32.7%; P = .37), nor did adherence to 12 best practices. The authors said the low assessment rate limited their ability to conclude the tool was not effective.
    Sources: Anesthesia & Analgesia, PubMed, Crossref
  • AI Oct 7, 2026
    GPT-4.5 nerve block targets show limited agreement with anesthesiologists, with 5% of regions usable without changes
    Target regions drawn by the general-purpose multimodal model GPT-4.5 on 100 pre-procedural ultrasound images for interscalene, supraclavicular, femoral and popliteal sciatic nerve blocks had a median Dice similarity of 0.47 with a consensus of three experienced anesthesiologists, versus 0.88 between experts, according to the Journal of Imaging Informatics in Medicine. Only 5% of AI-generated regions were acceptable without modification, although repeatability across sessions was high. The authors said the results support supervised evaluation or educational use rather than standalone procedural targeting.
    Sources: Journal of Imaging Informatics in Medicine, Springer

The weekly innovation roundup for anesthesiology, published every Sunday for the week before; every item links to its sources. Latest anesthesiology news · All specialties