An AI prompt for patients with cognitive impairment did not change delirium care

Design

Pragmatic randomised controlled trial embedded in routine care

Population

7,294 surgical patients with preoperative cognitive impairment, single US hospital

Primary outcome

Postoperative delirium (4AT) within 7 days

Effect

35.1% vs 32.7% in those assessed (P = .37); no difference in best-practice adherence

A pragmatic randomised trial in Anesthesia & Analgesia, published on 9 October, ran inside routine care at a large US hospital. A natural language processing algorithm, MMSE scores or diagnostic codes flagged 7,294 surgical patients with cognitive impairment immediately before surgery. Half had a non-interruptive decision support prompt in the electronic record encouraging 12 perioperative delirium best practices; half had routine care.

Adherence to the best practices did not differ. Only 17% of patients had any documented delirium assessment, and among them delirium occurred in 35.1% with the prompt and 32.7% without (P = .37).

The primary result is uninformative because most patients were never assessed, but the secondary result is clear: a passive prompt did not change what anaesthesia teams did. Identifying the at-risk patient is the easy part; the effort needs to go into reliable screening and changes that do not depend on someone noticing a message.

• Screen older surgical patients for cognitive impairment before surgery, so the risk is known.
• Build delirium prevention into default order sets rather than relying on passive alerts.
• Make postoperative delirium assessment routine — most patients here were never assessed.
• Avoid benzodiazepines and anticholinergic drugs where possible in patients with cognitive impairment.

Why it matters

Hospitals investing in AI alerts should expect identification without behaviour change unless the workflow changes too.

Don't overread it

With fewer than 1 in 5 patients assessed for delirium, the trial cannot say whether the tool affects delirium itself.

The statistics, in plain English

When only 17% of the randomised patients have the outcome measured, and those who were measured may differ from those who were not, randomisation no longer protects the comparison. The delirium result should be treated as unknown, not as no effect.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.