Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-2050
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-20T22:03:34.046581Z
status: published
lens: g_space
sector: health
headline: Predicting Measles Vaccination Defaulting in Uganda and Kenya: Comparative Evaluation of Custom Soft Voting and TabNet Stacked Ensemble Architectures
dek: 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, d…
gain_title: Ensemble models combining TabNet with tree-based classifiers identified likely measles vaccine defaulters better than individual models, and SHAP pinpointed country-specific drivers that could guide targeted outreach.
problem_title: (none)
trace_subject: (none)
gain_reading: Ensemble models combining TabNet with tree-based classifiers identified likely measles vaccine defaulters better than individual models, and SHAP pinpointed country-specific drivers that could guide targeted outreach.
gain_evidence: The proposed advanced ensemble learning models improved the classification of measles vaccine defaulters compared to individual models.
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers benchmarked nine classical and deep learning classifiers on Demographic and Health Survey data from 2016 to 2022 to predict which infants in Kenya and Uganda would default on measles vaccination. They paired a custom soft voting ensemble and a TabNet stacked ensemble with SHAP explanations, and built a tool using Streamlit and SQLite.

In Kenya the voting ensemble reached an AUC of 0.8847, and in Uganda the stacked hybrid reached 0.8862. The results matter because conventional defaulter identification is costly, reactive and prone to falsification, so a predictive approach could help target outreach. How well the models work in routine practice remains untested in the source.
limitation: In Uganda, the best hybrid ensemble's performance gain was statistically insignificant, so the improvement over simpler models there is not firmly established.
tag: Evidence-backed gain
key_points: Kenya's best model was a TabNet+GB+RF custom soft voting ensemble with an F1-score of 0.7111 and an AUC of 0.8847. | Uganda's best model was a TabNet hybrid ensemble (GB+RF+AB→TabNet) with an AUC of 0.8862 and an F1-score of 0.6756. | SHAP identified child's age and vitamin A1 as key predictors in both countries, while other predictors differed. | The study recommends outreach to areas with low literacy, distant water sources and mobile populations in Kenya, and reminders plus district outreach in Uganda.
rundown: The pipeline used feature selection from literature and expert input, vaccination eligibility filtering, missing value handling and phi-k correlation-based feature elimination. Class imbalance was addressed with SMOTE for nominal continuous features after RobustScaler, and models were tuned using stratified k-fold cross-validation and grid search.

In Kenya the most influential features were time to water source, respondents' education, previous district and current district. In Uganda they were polio inactive, Pentavalent3, current district and respondents' education.

The authors recommend country-tailored strategies: outreach in Kenya to low-literacy, distant-water and mobile areas, and in Uganda reminders to avoid missing preceding vaccines and outreach to high-default districts.
sources:
- peer_reviewed | JMIR AI | https://doi.org/10.2196/95727 | 2026-10-13
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

d1e25da5fe622f207ebe3b44c8083d72400be29b413460cd38e9258d9a8476be

previous

0000000000000000000000000000000000000000000000000000000000000000