Foundation Models Improve Childhood Anemia Prediction in Data-Scarce African Contexts

Childhood anemia remains a significant global health challenge, affecting nearly 40% of young children worldwide due to diverse underlying factors that hinder the generalizability of predictive models. This research investigates the effectiveness of transformer-based tabular foundation models, specifically TabPFN, against traditional supervised machine learning methods in predicting childhood anemia, particularly in settings with limited data and across varied international contexts.

The study leveraged Demographic and Health Surveys (DHS) data from 16 countries spanning Africa, Asia, Latin America, the Caucasus, and the Middle East, encompassing over 68,000 children. Researchers compared the performance of Logistic Regression, XGBoost, LightGBM, and TabPFN v2.6. The evaluation focused on metrics such as AUC-ROC, Brier score, and Expected Calibration Error (ECE), assessing generalization through leave-one-country-out, reverse-LOCO, and few-shot learning scenarios. Subgroup analyses were also conducted to identify potential demographic biases.

Key findings revealed that TabPFN significantly outperformed classical models in low-data environments, specifically with fewer than 200 samples, demonstrating superior discrimination and calibration capabilities. Across the diverse countries included in the study, TabPFN achieved the lowest Brier score and ECE, indicating better predictive accuracy and reliability. While full-data settings showed smaller differences between models, the overall stability of predictions across countries was largely influenced by specific country contexts, rather than model choice.

This research highlights that the accuracy of childhood anemia prediction is often more driven by population-specific variations than by the particular machine learning model employed. Crucially, the study underscores the advantage of foundation models like TabPFN in resource-constrained settings, where data scarcity is common. Their enhanced discrimination and calibration make them promising tools for advancing global health prediction efforts, especially in regions like Africa where detailed, large-scale health data can be limited.

The implications for Africa are substantial, suggesting that advanced AI models can offer robust solutions for public health challenges even when comprehensive datasets are unavailable. By providing more accurate and reliable predictions of childhood anemia, these models can support targeted interventions and resource allocation, ultimately contributing to improved child health outcomes across the continent.