Few-shot Cross-country Generalization of Tabular Machine Learning and Foundation Models for Childhood Anemia Prediction under Distribution Shift

Researchers evaluated transformer-based foundation models against classical machine learning methods for predicting childhood anemia across 16 countries using DHS data. TabPFN, a tabular foundation model, demonstrated superior performance in low-data environments with better calibration metrics, suggesting foundation models offer practical advantages for global health prediction in resource-constrained settings.

This research addresses a critical gap in applying advanced machine learning to global health challenges, specifically childhood anemia affecting 40% of children in developing regions. The study's cross-country evaluation reveals that foundation models like TabPFN excel where traditional methods struggle most—in low-resource settings with limited training data. This finding challenges the prevailing assumption that model architecture drives predictive performance; instead, population heterogeneity and contextual factors emerge as primary performance drivers.

The research builds on growing momentum in applying AI to healthcare equity. Foundation models, initially developed for language and vision tasks, are now proving valuable in tabular data prediction where domain-specific expertise traditionally dominated. The leave-one-country-out evaluation provides rigorous evidence of generalization capabilities across diverse populations, addressing a longstanding concern about AI model bias in global health applications.

For the AI and global health sectors, these findings validate foundation models as practical tools for developing nations lacking extensive healthcare datasets. TabPFN's improved discrimination and calibration in few-shot scenarios has direct implications for deploying predictive systems in low-resource hospitals and clinics. The stable subgroup performance across demographics suggests the approach doesn't introduce systematic bias—a critical requirement for ethical healthcare AI.

Future work should explore deployment feasibility and cost-effectiveness of foundation models versus classical approaches in actual clinical settings. The asymmetric transferability findings warrant investigation into which country characteristics facilitate or hinder cross-border model generalization, potentially enabling better prediction of generalization failures before deployment.

• →TabPFN foundation model outperforms classical ML in low-data regimes with fewer than 200 samples, achieving better discrimination and calibration
• →Population variation and country context drive prediction performance more than model architecture choice, with AUC-ROC differences between models remaining small
• →Foundation models show no systematic demographic bias across sex, age, residence, maternal education, and wealth subgroups in childhood anemia prediction
• →Child age, altitude, and height-for-age z-score emerge as dominant predictors, with wealth and maternal education as secondary factors
• →Cross-country generalization shows asymmetric transferability patterns, suggesting foundation models require careful evaluation before deployment in new regions