Identifying people who smoke at scale

Smoking kills around 80 000 people a year in the UK. Two in three long-term smokers will die from a smoking-related disease. The people most affected are also, disproportionately, those already dealing with social and economic disadvantage. This  means the gap between those who get support and those who don't tends to follow existing fault lines in health inequality.

Hospital admission is a recognised opportunity to intervene. Patients are already in contact with clinical teams, and a period of illness can be a prompt to act. But that conversation only happens if the right patients are identified first.

The problem with free text

Health Improvement Team (HIT) advisors at Imperial College Healthcare NHS Trust were already working to identify patients who smoke. Their work has a measurable impact with around 1 in 3 patients they support remain smoke-free at a three month follow-up.

But with finite resources and a large patient population across five hospitals, the team could only identify and approach a limited number of patients each day. "Timing is everything with tobacco dependency," one HIT advisor explains. "The sooner you speak to a patient, the better."

Researchers at iCARE wanted to see if machine learning tools could help. Working closely with HIT advisors, they set out to find a solution.

From clinical notes to a daily list

From the start, HIT advisors were central to how the tool was designed and deployed. Their knowledge of the clinical workflow shaped every decision. "The solution wasn't imposed on us," a member of the HIT team notes, "we helped shape it."

"Now our advisors can plan their day before 10am and reach patients much sooner." 

The model achieved approximately 90% accuracy on routine Trust data and now runs across five ICHT hospitals. Each morning it produces a prioritised list of patients likely to smoke, allowing advisors to focus their time on patients rather than records.

"Working alongside the Health Improvement Team from the start has been invaluable," says Maite Arribas Ardura, iCARE researcher. "It meant we built a model that delivers real clinical value, rather than simply moving the bottleneck somewhere else in the clinical pathway."

What could this mean for patients and the NHS?

Within three months, the number of patients identified had nearly doubled.

Researchers modelled the potential downstream impact if those additional patients go on to receive effective cessation support:

• Around 1,500 additional smokers identified in 2026
• Around 40 additional lives saved
• Around 80 fewer hospital readmissions
• The equivalent of one hospital bed freed up every day across the Trust

Translating AI research into clinical practice

The work illustrates how machine learning can be used to address a practical challenge in healthcare: making clinically relevant information within unstructured records accessible to the teams who need it.

Rather than introducing a new data collection process, the system uses information already generated through routine clinical care. The deployment across five hospitals at ICHT represents a transition from the development and evaluation of an AI method to its use within a live NHS service.

Smoking status is just one example. So much valuable information sits in free text where it's invisible to the systems that plan and deliver care. We hope other services can use similar approaches to unlock it. Maite Arribas Ardura Senior Data Scientist, iCARE

Could this apply to your work?

This project used machine learning to extract structured insight from unstructured clinical text - a challenge that appears across many NHS services. If your team is working with a similar problem, get in touch with iCARE →

If you or anyone else you know is looking to quit smoking or nicotine products, free and effective professional support is available through local NHS stop smoking services.