Pre-trained Encoders for Global Child Development: Transfer Learning Enables Deployment in Data-Scarce Settings

TL;DR

This work tackles the data bottleneck in global child development monitoring by introducing a globally pre-trained Tabular Masked Autoencoder trained on $357{,}709$ children across $44$ countries to learn a transferable developmental prior. With just $N=50$ local samples, it achieves an average AUC of $0.65$ and reaches $0.73$ at $N=500$, while zero-shot predictions for unseen countries reach as high as $0.84$, backed by a domain-adaptation bound explaining the role of cross-country diversity. The approach yields substantial data-efficiency gains, enabling deployment in resource-constrained settings and proposing a practical two-stage framework: global pre-training followed by local fine-tuning ($N$ in the low hundreds). Across regional tests and small-island cases, the encoder shows robust transfer and calibration, though full-data performance remains edges ahead for some baselines. These results advance SDG 4.2.1 monitoring by providing a scalable, transferable, and equity-conscious ML paradigm for global child development surveillance.

Abstract

A large number of children experience preventable developmental delays each year, yet the deployment of machine learning in new countries has been stymied by a data bottleneck: reliable models require thousands of samples, while new programs begin with fewer than 100. We introduce the first pre-trained encoder for global child development, trained on 357,709 children across 44 countries using UNICEF survey data. With only 50 training samples, the pre-trained encoder achieves an average AUC of 0.65 (95% CI: 0.56-0.72), outperforming cold-start gradient boosting at 0.61 by 8-12% across regions. At N=500, the encoder achieves an AUC of 0.73. Zero-shot deployment to unseen countries achieves AUCs up to 0.84. We apply a transfer learning bound to explain why pre-training diversity enables few-shot generalization. These results establish that pre-trained encoders can transform the feasibility of ML for SDG 4.2.1 monitoring in resource-constrained settings.

Figures (4)

• Figure 1: Few-Shot Transfer Performance. The pre-trained encoder (solid lines) consistently outperforms cold-start gradient boosting (dashed lines) in data-scarce regimes ($N < 1000$). Shaded areas represent 95% confidence intervals across 100 bootstrap resample-stratified splits. The pre-trained encoder enables deployment-grade performance ($>0.70$ AUC) with significantly fewer local samples than state-of-the-art baselines.
• Figure 2: Regional Generalization Maps. Zero-shot AUC across Latin America, Africa, and Asia. The Pre-trained Encoder generalizes effectively to diverse national contexts without local training data.
• Figure 3: Model and Regional Comparison. Small multiples show consistent performance gains across Africa, Asia, and Latin America. The Pre-trained Encoder (blue) maintains superiority over cold-start baselines (red) across sample sizes.
• Figure 4: Feature Importance (SHAP). Real-world SHAP values from N=10,000 samples confirm that Child Age, Mother's Education, and Wealth are the primary drivers of prediction, aligning with developmental science.

Theorems & Definitions (9)

• Theorem 3.2: Transfer Learning Generalization Bound
• proof : Proof Sketch
• Proposition 3.3: Sample Complexity Reduction
• proof : Proof Sketch
• Definition 5.1: $\mathcal{H}$-divergence
• Lemma 5.2: Ben-David et al., 2010
• proof : Proof of Theorem \ref{['thm:transfer']}
• Lemma 5.3: VC Dimension Bound
• proof : Proof of Proposition \ref{['prop:sample']}