Safety-First AI Maps Vaccine Hesitancy to Prevent Communication Backfire

Public health campaigns have long operated on a simple assumption: if you explain the science clearly enough, people will accept vaccines. Decades of behavioral research have shown how fragile that assumption is. When health messages collide with deep-seated distrust, they can trigger psychological reactance, a defensive response in which people harden their opposition precisely because they feel pressured. A new study published in BMC Public Health takes this problem seriously and proposes something unusual for the machine learning world: an algorithm designed not to maximize its own accuracy, but to know when it should stay silent.

The research, led by Mehrdad Askarian of Shiraz University of Medical Sciences together with colleagues at Amirkabir University of Technology, the University of Saskatchewan, and Shiraz University of Medical Sciences, presents a safety-first machine learning framework for what the authors call precision public health. Instead of treating an entire population as a single audience, the framework sorts individuals into distinct psychometric profiles, each of which calls for a different communication strategy. The goal is to route people toward messages they can actually use, and away from messages that might backfire.

The study draws on a cross-sectional survey of 457 Iranian adults, of which 404 records made it into the de-identified machine learning dataset after processing. From demographic and geographic information, the framework maps each person onto one of three operational segments: Accepting, Ambivalent, and Resistant. These labels are not arbitrary. They were constructed from fixed boundaries on Likert-scale responses to theory-grounded composites built on two established models of health behavior: the Health Belief Model, which explains protective behavior through perceived susceptibility, severity, benefits, and barriers, and the Confidence, Complacency, and Convenience model, known as the 3C model, which is one of the most widely used frameworks for understanding vaccine hesitancy worldwide.

The classification engine itself is deliberately conventional. The researchers evaluated a logistic regression model using nested stratified cross-validation, a rigorous scheme in which the data is repeatedly split so that the model is tested on samples it has never seen, with an inner loop used for tuning. The results are strikingly modest, and the authors are refreshingly candid about that. The base classifier achieved an accuracy of 0.446, with a 95 percent confidence interval of 0.401 to 0.493, a balanced accuracy of 0.447, and a macro-F1 score of 0.419. For context, simply guessing the most common class every time would yield an accuracy of 0.438. In other words, the model barely outperformed a coin weighted toward the majority.

That near-baseline performance is not a failure of the study; it is the point around which the entire design pivots. The authors argue that in high-stakes health communication, a wrong prediction is not merely a statistical error but a potential harm. If a genuinely resistant person is misclassified as accepting and receives a pushy pro-vaccination message, the result could be reactance and entrenchment. If an ambivalent person is pushed too hard, they might slide toward refusal. The framework therefore treats misclassification asymmetrically, prioritizing the avoidance of what it calls critical errors over raw coverage.

The centerpiece of this safety philosophy is a mechanism the authors call the Safety Valve. Rather than forcing the classifier to commit to a prediction for every individual, the system computes a confidence threshold for each fold of the cross-validation procedure, with a mean threshold of 0.406, a standard deviation of 0.038, and a range spanning 0.36 to 0.45. When the model’s confidence in a prediction falls below this threshold, the case is not assigned to any segment at all. Instead, it is deferred to a human path the authors label Dialogue/Review, where communication is handled through conversation and professional judgment rather than automated targeting. In the reported evaluation, the Safety Valve deferred 16.8 percent of cases to this route.

The effect on harm was measurable, if modest. Observed critical errors, the mistakes most likely to cause communication damage, fell from 15.25 percent without the valve to 11.86 percent with it. The authors report that this reduction was not statistically significant, with a McNemar test yielding a p-value of 0.50, and they are careful not to overclaim. But the design logic is clear: the system trades some autonomous coverage for a lower rate of potentially harmful misrouting, and it explicitly flags the cases where the data is too ambiguous to justify automated action. This is a notable departure from the prevailing culture in applied machine learning, where models are typically optimized to predict everything and confidence calibration is an afterthought.

To make the model interpretable, the team applied SHAP analysis, or SHapley Additive exPlanations, a technique from game theory that assigns each input feature a contribution to individual predictions. The analysis identified the largest grouped predictors as healthcare professional status, education, employment, marital status, and province of residence. The authors stress that these are statistical correlates, not causal determinants. Knowing that a person works in healthcare or holds a certain educational level does not explain why they hesitate; it only signals where they tend to fall in the observed distribution. This distinction matters for anyone hoping to translate the framework into policy, because targeting based on correlates can entrench stereotypes if it is mistaken for causal understanding.

The study’s limitations are as instructive as its results. The sample was enriched with healthcare professionals and highly educated respondents, a consequence of recruitment channels that the authors acknowledge openly. This means the segment distribution and the model’s decision boundaries reflect a population that is not representative of Iranian adults at large, let alone other countries. The authors state plainly that the findings are not directly generalizable without recalibration and external validation, and that scalability, effectiveness, and feasibility all require field testing. No scalability claim is made. The framework is offered as a descriptive blueprint, a proof of concept for how risk-stratified health communication could be structured, rather than a deployable system.

Even in that modest form, the study lands at a moment when the questions it raises are urgent. Vaccine confidence has become one of the most politically and psychologically charged issues in global health, and blunt messaging campaigns have repeatedly shown their limits. The idea that an algorithm should be built to defer, to hand uncertain cases to human dialogue rather than push automated messages, inverts the usual logic of predictive systems. Whether safety valves like this one can be validated at scale, and whether segmenting populations by trust profile can be done ethically and without stigmatizing the Resistant segment, are questions the authors themselves place on the future agenda. What the study demonstrates now is that precision public health can be designed with harm prevention as its first objective, and that admitting uncertainty may be the most valuable prediction a health algorithm can make.

Subject of Research: A safety-first machine learning framework for segmenting vaccine hesitancy among Iranian adults into tailored public health communication profiles

Article Title: A safety-first machine learning framework for precision public health: segmenting vaccine hesitancy among Iranian adults

Article References: Askarian, M., Bayati, S., Rajabi, H., Askarian, A., Hatam, N., & Ayareh, N. (2026). A safety-first machine learning framework for precision public health: segmenting vaccine hesitancy among Iranian adults. BMC Public Health. https://doi.org/10.1186/s12889-026-29763-2

Image Credits: AI Generated

DOI: 10.1186/s12889-026-29763-2

Keywords: vaccine hesitancy, machine learning, precision public health, psychological reactance, explainable AI, health communication, logistic regression, SHAP, Health Belief Model, 3C model, Iran, psychometrics