68/100

Moderate evidence

Cite
More
EmailLinkedInX

AI note-scanning nearly doubled smoking referrals at a London hospital trust, a 12-week study finds

Study title: Expanding tobacco dependency service reach through AI-assisted identification of hospital inpatients who smoke: A prospective study

Evidence type
Controlled or validated comparison
Sources
1 sourcePeer-reviewed
Published
2 min read

Study title: Expanding tobacco dependency service reach through AI-assisted identification of hospital inpatients who smoke: A prospective study

A 12-week prospective study at a large London acute NHS trust tested ClinicalSmokeBERT, a language model that reads free-text clinical notes each day to find patients who currently smoke. Its referrals ran alongside the standard pathway based on structured records, and the AI pathway raised referrals to the tobacco dependency service by 99.6%.

Read the study abstract

Objectives: Tobacco smoking remains a major public health challenge, and the NHS Long Term Plan commits hospitals to identifying and treating tobacco dependency in all admitted patients who smoke. Systematic identification of patients who smoke within hospitals is limited by incomplete structured electronic health record (EHR) data, leaving many patients unreferred to cessation support. Natural language processing (NLP) offers a means of identifying smoking status from unstructured clinical text. We evaluated th…

58Gain side

How the two sides scoreClose call

54Problem side

The gain

In a prospective 12-week study at a large London acute NHS trust, an AI model that scanned clinical notes daily raised referrals to the tobacco dependency service by 99.6% compared with the standard non-AI pathway.

The problem

In the same 12-week London NHS trust deployment, referrals from the AI note-scanning pathway were less likely to be confirmed as current smokers: 66.0% positive predictive value versus 77.1% for the standard pathway, and non-inferiority was not shown.

A fixed formula sets both scores from the evidence and its sources. Reader votes don’t move them.

What the scores mean

G is the gain side, out of 100. It goes up when the gain is specific, when it measurably helped people and when the sources behind it are strong.

P is the problem side, out of 100. It goes up when the problem is specific, when it measurably hurt people and when the sources behind it are strong.

In brief

The NHS Long Term Plan commits hospitals to identifying and treating tobacco dependency in every admitted patient who smokes, but incomplete structured records leave many patients unreferred. Reading free text is one way to close that gap.

Quality was mixed. Among patients confirmed as smoking, 89.2% of AI referrals accepted support versus 85.0% for standard referrals. But the share of patients seen by an advisor who were confirmed as smoking was 66.0% for AI versus 77.1%, and non-inferiority was not shown.

Main points

  1. The study ran for 12 weeks in routine clinical deployment at a large London acute NHS trust.
  2. ClinicalSmokeBERT scanned clinical notes daily and generated referrals alongside the standard structured-data pathway.
  3. The primary outcome was the number of referrals from each pathway; the AI pathway produced 99.6% more than standard care.
  4. Among patients confirmed as currently smoking, 89.2% accepted support on the AI pathway versus 85.0% on the standard pathway.
  5. Positive predictive value was 66.0% for AI referrals versus 77.1% for standard referrals, and the 15-point non-inferiority margin test did not demonstrate non-inferiority.
  6. No significant sociodemographic differences were seen between pathways, which the authors read as no evidence of model bias.

The full story

The trial setting was a large acute NHS trust in London. The AI model supplemented the existing referral route rather than replacing it, so the comparison was between referrals generated by notes-scanning and those from structured data.

Secondary outcomes used a non-inferiority framework with a 15-percentage-point margin. Positive predictive value was the proportion of patients seen by a tobacco dependency advisor who were confirmed as currently smoking, and engagement was the proportion of confirmed smokers who accepted support.

The authors conclude that AI-assisted identification extends service reach without compromising referral quality or introducing demographic bias, and support using it as an additive part of routine clinical pathways.

What this doesn’t fix

Only patients seen by a tobacco dependency advisor were included in the secondary analyses, because the smoking status of unseen patients could not be verified, and PPV non-inferiority was not demonstrated.

Sources

1 on file
  1. Public Healthdoi:10.1016/j.puhe.2026.106560
    Peer-reviewed2026-11-12

Talk it over

Three questions to start on, two more if you have time
The AI pathway raised referrals by 99.6%, but 66.0% of its referrals were confirmed as current smokers versus 77.1% for the standard pathway. How do you weigh the extra reach against the lower confirmation rate?

Start with a question

AI note-scanning nearly doubled smoking referrals at a London hospital trust, a 12-week study findshttps://truvace.com/item/expanding-tobacco-dependency-service-reach-through-ai-assisted-identification-of

  1. The AI pathway raised referrals by 99.6%, but 66.0% of its referrals were confirmed as current smokers versus 77.1% for the standard pathway. How do you weigh the extra reach against the lower confirmation rate?

    Answer in the thread
  2. The quality figures only cover patients seen by a tobacco dependency advisor, because the smoking status of unseen patients could not be verified. How might that limit change what we conclude about the AI referrals?

    Answer in the thread
  3. If you were an advisor at the London trust receiving referrals from both pathways, how would you handle the AI referrals that turned out not to be current smokers?

    Answer in the thread
If you have more time
  • The authors read the lack of significant sociodemographic differences between pathways as no evidence of model bias. What other checks would you want before accepting that reading?

    Answer in the thread
  • What would you look up about a hospital you know to judge whether free-text note scanning could add referrals there, given the NHS Long Term Plan commitment to treat every admitted patient who smokes?

    Answer in the thread

What readers say