TabPFN Outperforms Classical Models in Cross-Country Anemia Prediction
A new study looks into how well transformer-based tabular foundation models perform in predicting childhood anemia compared to standard supervised methods, especially with limited data across different countries. The research analyzed DHS data from 16 countries in Africa, Asia, Latin America, the Caucasus, and the Middle East, involving 68,856 samples. They evaluated methods like Logistic Regression, XGBoost, LightGBM, and TabPFN v2.6, using metrics such as AUC-ROC and Brier score, and tested generalization through various approaches. Key factors examined included sex, age, residence, maternal education, and wealth, with feature importance assessed via SHAP. Notably, TabPFN outperformed others in low-data scenarios, achieving an impressive Brier score of 0.042, highlighting its promise for global health challenges.
Key facts
- Childhood anemia affects around 40% of children aged 6-59 months globally.
- Data from 16 countries across Africa, Asia, Latin America, the Caucasus, and the Middle East (n=68,856).
- Models compared: Logistic Regression, XGBoost, LightGBM, and TabPFN v2.6.
- Performance metrics: AUC-ROC, Brier score, ECE.
- Generalization evaluated using LOCO, reverse-LOCO, and few-shot settings.
- Subgroup analyses included sex, age, residence, maternal education, wealth.
- Feature importance estimated using SHAP.
- TabPFN outperformed classical models in low-data regimes (<200 samples).
- TabPFN achieved lowest Brier score (0.042) across countries.
Entities
Locations
- Africa
- Asia
- Latin America
- Caucasus
- Middle East