We explore numerical and categorical feature engineering strategies in predicting COVID-19 vaccine hesitancy among Malaysian healthcare workers using machine learning models. Using a survey dataset containing 554 instances and 92 features, feature sets acquired through various feature selection and aggregation strategies are evaluated through model performance across several machine learning algorithms including Support Vector Machine (SVM), Multilayer Perceptron (MLP, K-Nearest Neighbors (KNNs), and Naive Bayes (NB). We show how we can achieve the best model and feature outcomes through different combinations of filter-based and wrapper-based feature selection methods with feature aggregation methods. For primary COVID-19 vaccine hesitancy prediction, MLP with ANOVA + CHI filters applied on aggregated features shows the best results. On the other hand, Bagging Linear SVM with ANOVA + CHI filters applied on aggregated features shows the best performing model for booster COVID-19 vaccine hesitancy prediction. Our study shows features selected to be important for primary vaccine hesitancy prediction differ from booster vaccine hesitancy prediction with only one common important predictor, views on the vaccine. The findings from our study underscore the importance of targeted feature engineering in enhancing model performance and provide insights into the critical factors driving vaccine hesitancy among healthcare workers.