Predicting Measles Vaccination Defaulting in Uganda and Kenya: Comparative Evaluation of Custom Soft Voting and TabNet Stacked Ensemble Architectures.

BACKGROUND: Measles is a major cause of death among infants despite free routine immunization (RI) in low- and middle-income countries (LMICs). Vaccination defaulting contributes to measles outbreak occurrences. Conventional immunization defaulter identification methods are expensive to implement, prone to data falsification, reactive, and overestimate uptake. Several machine learning studies leveraged the use of child and maternal characteristics as predictors of immunization defaulting, overlooking community, demographic, and socioeconomic factors. Similarly, there is limited research regarding the use of interpretable deep learning and advanced ensemble methods to improve model performance in predicting RI defaulters and establishing significant predictors at the national level in the LMIC context, like Uganda. OBJECTIVE: This study predicts infants at a risk of defaulting on measles vaccination in Uganda and Kenya using advanced ensemble learning architectures. METHODS: The study benchmarked classical and deep learning models using Demographic and Health Survey (DHS) datasets (2016-2022) with child, maternal, community, social-economic, and demographic characteristics as predictors. The robust modeling pipeline began with feature selection according to the literature and the expert domain. It was followed by vaccination eligibility, missing value handling, and feature elimination using the phi-k correlation framework. To mitigate class imbalance, synthetic minority over-sampling technique for nominal continuous features was used in a training pipeline after the RobustScaler. Nine baseline classifiers, including Tabular Network (TabNet), multilayer perceptron (MLP), random forest (RF), gradient boosting (GB), adaptive-boosting (AB), among others, underwent stratified k-fold cross-validation and hyperparameter optimization via grid search. Custom soft voting was evaluated alongside TabNet stack architecture for superior ensemble prediction. Models were evaluated for accuracy, precision, recall, and F1-score. Model explainability was decoded using Shapley additive explanations (SHAP), while Streamlit and SQLite formed the tool architecture. RESULTS: The proposed TabNet+GB+RF custom soft voting achieved significant, nonrandom performance over tuned base models with an F1-score of 0.7111, area under the curve (AUC) of 0.8847 for Kenya. Conversely, for Uganda, the TabNet hybrid ensemble (GB+RF+ AB→TabNet) was the best with AUC of 0.8862 and F1-score of 0.6756, which was statistically insignificant. SHAP identified child's age and vitamin A1 as universal key predictors in both countries; others varied. In Kenya, time to water source, respondents' education, district lived previously, and current district were the most influential global features, respectively. In Uganda, Polio inactive, Pentavalent3, current district, and respondents' education were the most influential factors to defaulting, respectively. CONCLUSIONS: The proposed advanced ensemble learning models improved the classification of measles vaccine defaulters compared to individual models. The study recommends distinct country-tailored strategies. In Kenya, outreaches are recommended to areas with low literacy, distant water sources, and mobile populations. Conversely, in Uganda, the study recommends reminders to avoid missing preceding vaccines and outreaches to high-default districts.