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

Traditional one-size-fits-all public health communication often fails to account for individual psychological barriers, potentially triggering psychological reactance and backfire effects. To address this, a Precision Public Health approach is required to categorize populations into tailored communication paths based on their unique trust profiles. Utilizing data from a cross-sectional survey of 457 Iranian adults, we analyzed 404 records in the de-identified machine-learning dataset. We developed a descriptive computational framework to map demographic and geographic features onto three operational psychometric segments: Accepting, Ambivalent, and Resistant. Target classes were constructed from fixed Likert-scale boundaries on theory-grounded composites based on the HBM and 3 C models. Logistic regression was evaluated using nested stratified cross-validation. A fold-specific confidence threshold (mean τ = 0.406; SD = 0.038; range 0.36–0.45) was used as a Safety Valve to route low-confidence predictions to Dialogue/Review. The base classifier achieved accuracy 0.446 (95% CI 0.401–0.493), balanced accuracy 0.447 (95% CI 0.395–0.500), and macro-F1 0.419 (95% CI 0.375–0.465). The majority-class baseline accuracy was 0.438. The Safety Valve deferred 16.8% of cases to Dialogue/Review and reduced observed Critical Errors from 15.25% to 11.86%, although this reduction was not statistically significant (McNemar p = 0.50). SHAP analysis identified Healthcare professional, Education, Employment, Marital status, and Province as the largest grouped predictors; these are statistical correlates, not causal determinants. This study provides a descriptive blueprint for risk-stratified health communication that prioritizes harm prevention over autonomous coverage. However, given the sample’s enrichment with healthcare professionals and highly educated respondents, these findings are not directly generalizable without recalibration and external validation. Scalability, effectiveness, and feasibility require external validation and field testing; no scalability claim is made here.