IN BRIEF
More than 90,000 patients took part in three field trials.
Personalized high-risk messages increased vaccination by 1.1–1.4 points versus a simple reminder.
Saying the estimate came from an algorithm, or explaining its reasons, did not increase the effect further.
WHY IT MATTERS
Many health systems can already identify vulnerable patients; the harder problem is turning that information into action.
Here a risk prediction became a real-world tested intervention with a measurable effect on vaccination.
WHAT IS NOW POSSIBLE
Algorithmic risk prediction moved from identification alone to a tested behavioral intervention that changed a real health action in routine care.
WHAT MADE IT POSSIBLE
An already validated algorithm identified high-risk patients, and the trial tested different ways of communicating that information.
THE STORY
Knowing that someone is high risk matters only if it changes something. In three randomized field trials involving more than 90,000 US patients, directly telling people they were at high risk for influenza and complications increased the chance that they got vaccinated.
Compared with a simple reminder, vaccination increased by 1.1–1.4 percentage points. Compared with people who received no message, the increase was 1.7–3.5 points.
That is a modest individual effect, but it can become meaningful when applied to tens of thousands of people. Crucially, the result comes from recorded behaviour in a healthcare system rather than stated intentions in a survey.
One counterintuitive result concerns AI: explicitly telling patients that an algorithm determined their risk, or showing personalized reasons behind the prediction, did not increase vaccination further.
The signal is therefore not that algorithm explanations persuade people more. It is simpler: making risk personal and concrete can shift a preventive decision. The study does not show that the messages directly reduce influenza cases or hospitalizations.
WHAT WE KNOW
The outcome was recorded vaccination behaviour. [1]
The effect was observed across three preregistered trials.
Personalized risk messaging performed better than a simple reminder.
Algorithm disclosure added no measurable effect.
WHAT WE DON'T KNOW
How well the result generalizes outside this US health system.
Whether the same approach works for other vaccines or preventive behaviours.
How much the vaccination increase reduces infections, complications or hospitalizations.
WHAT HAPPENS NEXT
Other health systems can test the strategy across populations, vaccines and risks while measuring downstream clinical outcomes.
SOURCE TRACE
The programme included three preregistered randomized trials with more than 90,000 unique healthcare patients.
High-risk messages increased vaccination by 1.1–1.4 percentage points versus simple reminders.
Vaccination was 1.7–3.5 percentage points higher than among patients receiving no message.
Disclosing algorithm involvement or its personalized explanations did not measurably improve vaccination beyond the risk message.
SOURCES
- Algorithm-assisted personalized risk communication to encourage flu vaccination in the USA: three randomized field trialsOpen source
CURRENT EVIDENCE
VALIDATED EVIDENCE
Real healthcare messaging and observed vaccination behavior.
WHAT WOULD CHANGE THIS?
- Replications in other health systems finding no vaccination increase.
- Evidence of harmful effects or substantial inequities caused by the targeting strategy.