Global generalisability of AI-driven COVID-19 vaccination policies: a cross-sectional observational study (Preprint)

Abstract

Objective: To discover global determinants of vaccine uptake behavior and to develop a generalizable machine learning model to predict the vulnerability of vaccine uptake behavior at the individual level. Methodology: 23135 Respondents across the 23 countries were interviewed for the survey questionnaire, after preprocessing and cleaning data, we performed Bayesian networks and generalized linear models to identify the key determinants of vaccine uptake. Markov Blankets obtained from the Bayesian networks were used to estimate the important predictors of the vaccine uptake. These variables were then used to build the models. To build generalizable models, we used country-wise data splitting. Model evaluation is assessed for the prediction performance on the new countries. We also developed income specific models cross validated within the income group. Results: We found 16 important predictors of vaccine uptake using the Bayesian network and Markov Blanket approach. We found that the trust of the central government (Log-Odds 0.55[0.25, 0.84] (p= 0.0002)), Vaccination restriction for national and international travel (Log-Odds 0.4[0.14, 0.65] (p= 0.0034)) as the key determinants of Vaccine uptake. Our Generalized mixed effects model approach achieved an AUC of 89%, Precision 90% and Recall of 82% on the prediction task on new countries, thus, generalizing to new countries.