AI Flu-Risk Messages Lifted Vaccination in 90,000-Patient Trials

Telling patients that a machine learning model had flagged them as high risk for serious flu complications made them more likely to get vaccinated, according to three randomized field trials with more than 90,000 patients published on 5 October 2026 in Nature Human Behaviour. Patients told they were high risk were vaccinated at rates 1.1 to 1.4 percentage points higher than those who received a plain reminder. Saying that an algorithm was behind the risk estimate made no measurable difference either way.

The paper, led by researchers in the Behavioral Insights Team at Geisinger, the Pennsylvania health system, tackles a question that matters as hospitals put predictive AI into patient outreach: does disclosing the AI help, hurt or simply not register with patients?

How the Geisinger flu trials were designed

The team started with a previously validated machine learning model that scores patients on their risk of influenza and related complications. According to the paper’s competing interests statement, Geisinger paid the analytics company Medial EarlySign to generate the risk scores and the personalized reasons used in the studies. Patients the model identified as high risk were then randomly assigned either to receive no message or to receive one of several messages encouraging them to get a flu shot, which was the primary outcome.

All three trials were preregistered and listed on ClinicalTrials.gov, including the second study’s registration, and were sized to detect an effect of about 2 percentage points, in line with earlier research on vaccination nudges. The message arms differed in what they told patients. Some received a standard reminder to get vaccinated. Others were told they were at high risk, with variants that said the risk came from a review of their medical records, said it came from an algorithm, or listed specific reasons for their risk, a simple form of explainable AI.

The scale was substantial. In the second study, which began sending messages on 9 September 2021, each of the five arms held about 8,500 patients. The third study ran from mid-September to late October 2022 with roughly 7,800 patients per arm, and everyone in its high-risk arms was told they were in the top 20% of risk. In the first study, patients in the top 3% were told exactly that, while those in the top 4% to 10% were randomly told either that they were in the top 10% or simply that they were at high risk.

Risk messages beat plain reminders in all three trials

The headline finding held across the studies. Among patients informed of their high risk, vaccination was 1.1 to 1.4 percentage points higher than among patients who got only a reminder, a relative increase of 3.3% to 5.4%. Compared with patients who received no message at all, the high-risk messages raised vaccination by 1.7 to 3.5 percentage points, or 3.3% to 14.7% in relative terms.

The baseline rates show why even small gains matter. Among patients who received no message, 50.8% were vaccinated in the first study, 31.5% in the second and 23.8% in the third. In a large health system, a lift of one or two points can mean thousands of additional shots for the people most likely to end up in hospital.

The effect sizes are modest, and the authors do not claim otherwise. Behavioral nudges rarely move behavior by large margins, and the researchers designed the trials around the small effects typical of that literature. What stands out is that the gain came from personalized risk information rather than from simply sending another reminder.

Patients shrugged at the word algorithm

The more novel result concerns trust in AI. Laboratory studies have found conflicting evidence on whether people avoid or prefer algorithmic judgments, a debate often framed as algorithm aversion versus algorithm appreciation. In these real-world trials, vaccination was similar whether or not the message mentioned an algorithm and whether or not it explained the reasons behind the risk estimate.

The authors interpret that as a sign that patients were neither put off by nor drawn to algorithm use or explainability in this setting. For health systems weighing whether to disclose AI involvement in outreach, that is useful evidence: transparency did not cost them uptake, but adding explanations did not buy extra uptake either.

The finding fits a broader shift in which AI is moving from research tools into routine care. PromptCrates has reported on ChatGPT for Healthcare adding Epic EHR access and on the FDA TEMPO pilot letting AI care tools reach patients early. Those efforts raise the same question this study answers for one narrow case: how patients respond when software, not a clinician, is the source of the advice.

What health systems can take from the results

The practical lesson is that predictive models can be useful for targeting as well as for diagnosis. Instead of sending identical reminders to everyone, a health system can use a risk model to find the patients who would benefit most and tell them, specifically, why the shot matters for them. Similar targeting logic underpins diagnostic tools such as the AI ECG that flags heart failure in under two seconds.

There are limits. The trials ran within one US health system, and the paper reports outcomes for flu vaccination only, so the results may not transfer to other conditions, populations or countries with different attitudes toward AI. The effect sizes, while statistically meaningful, are small, and the algorithm’s risk scores came from a commercial vendor. The paper notes that part of the work was funded by the National Institute on Aging of the National Institutes of Health.

The researchers have posted de-identified data and analysis code on the Open Science Framework, which should let other teams check and extend the analysis. As flu season begins in the Northern Hemisphere, the study offers health systems a tested template: use the model to find high-risk patients, tell them plainly that they are high risk, and do not worry too much about whether to mention the algorithm.

• Nature Human Behaviour: Algorithm-assisted personalized risk communication to encourage flu vaccination in the USA
• ClinicalTrials.gov: Using Explainable AI Risk Predictions to Nudge Influenza Vaccine Uptake
• PromptCrates: ChatGPT for Healthcare adds Epic EHR access
• PromptCrates: FDA TEMPO pilot lets AI care tools reach patients early
• PromptCrates: AI ECG flags heart failure in under two seconds