JMIR Publications

ABSTRACT

Background:

Measles is a major cause of death among infants despite free routine immunization in low-and middle-income countries (LMIC). Vaccination defaulting contributes towards measles outbreak occurrences. Conventional immunization defaulter identification methods are expensive to implement, propagate data falsification, overestimate uptake and reactive. Several machine learning scholarly 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 ensemble methods to improve model performance in predicting basic RI defaulters and establishing significant predictors at the national level in the LMIC context, like Uganda.

Objective:

This study presents a proactive and low-resource-demanding tool designed based on the proposed ensemble model for predicting infants at a risk of defaulting measles vaccination.

Methods:

The study benchmarked classical and deep learning models using Demographic and Health Survey datasets (2016 - 2022) with child, maternal/household, community, social-economic and demographic characteristics as predictors. The robust modeling pipeline begun with feature selection according to literature and expert domain. It was followed by vaccination eligibility, missing value, min-max scaling for normalisation and balancing with synthetic minority over sampling technique. Subsequently, k-fold cross validation ensured reliable performance metrics while exhaustive hyperparameter tuning via grid and randomized search enabled model optimisation feeding into stacking experiments that leveraged on hard, soft and meta classifiers for superior ensemble prediction. TabNet, Multilayer perceptron (MLP), random forests, gradient-boosting, Adaboost, logistic regression, among others were bench marked. Models were evaluated for accuracy, precision, recall and F1-Score. SHapley Additive Explanations (SHAP) values were used for local and global feature influence interpretability while streamlit, SQLite formed the tool architecture.

Results:

Experimental results revealed that TabNet-MLP ensemble significantly outperformed other models by improving immunisation defaulter prediction by 39%(U) and 6%(K) compared to the best base models with an F1-Score of 0.8553(K) and 0.7286(U). For clinical utility, a functional tool based on the proposed ensemble was scaled to a production-ready analytical dashboard. Similarly, given the growing need for interpretability, Shapley additive explanations (SHAP) analysis identified respondents’ education and the age of a child as key global predictors for model output in both countries, whereas the marginal contributions of the remaining features exhibited significant variance across observations.

Conclusions:

The proposed TabNet-MLP based functional prototype significantly improved classification of measles vaccine defaulters presenting proactively tool. The study also recommends prioritising education, outreaches targeting families with many infants and highly mobile populations in Kenya. In Uganda, strengthening immunisation services in underserved districts and creating a robust system for managing health-cards.