Researchers have demonstrated that OpenAI's GPT-4o-mini can predict child malnutrition with comparable accuracy to traditional machine learning models, according to a new study using Bangladesh health survey data spanning 15 years.

The research, accepted to the AIME 2026 Workshop, tested zero-shot inference capabilities on child stunting prediction using Bangladesh Demographic and Health Survey data collected between 2007 and 2022.

GPT-4o-mini achieved balanced accuracy matching a random forest baseline while showing substantially higher sensitivity for identifying stunting cases. The model maintained consistent performance across different survey waves and child sex groups.

Researchers transformed maternal, child, healthcare, and household characteristics into prompt-based representations that the language model could interpret without prior training on the specific task.

Fairness concerns emerge

The study revealed important disparities in model performance across residence and household wealth categories. These fairness issues highlight potential risks before deploying foundation models in public health settings.

Child malnutrition remains a critical challenge in low- and middle-income countries, particularly in South Asia, where early identification enables timely intervention and resource allocation.

The zero-shot approach offers potential advantages over traditional supervised learning methods, which require extensive labeled datasets and domain-specific model training.

Temporal robustness testing showed the model maintained stable predictive behavior across multiple BDHS survey waves, suggesting reliability over time periods.

The research was conducted by Muhammad Ashad Kabir and Md Ahshanul Haque, focusing on the feasibility of using pretrained language models for population health prediction tasks.

The findings suggest foundation models could support public health decision-making while emphasizing the need for careful evaluation of bias and fairness before real-world deployment in healthcare applications.