NutriAI: AI-Powered Child Malnutrition Assessment in Low-Resource Environments

Misaal Khan, Shivang Agarwal, Mayank Vatsa, Richa Singh, Kuldeep SinghIJCAI 2023 · DOI

Introduction

Malnutrition is a primary driver of morbidity and mortality for children under five, particularly in developing nations. Traditional screening relies on manual physical measurements—such as height, weight, and Middle-Upper-Arm-Circumference (MUAC)—which are compared against WHO or CDC growth charts. However, this approach is often hindered by subjectivity, the need for specialized equipment, and limited scalability in remote areas.

NutriAI is a novel, smartphone-based screening tool designed to address these challenges. By leveraging 2D images and a dual-branch deep learning framework, the system offers a low-cost, objective, and non-invasive alternative to manual assessments.

Methodology

The core of NutriAI is a dual-branch framework designed to provide "bidirectional validation" of a child's nutritional status:

  1. Branch 1 (Few-Shot Classification): This branch analyzes facial and full-body features to perform binary classification (Healthy vs. Malnourished). It utilizes few-shot learning to handle the relatively small datasets typical in medical AI.
  2. Branch 2 (Domain-Guided Regression): This branch uses multipose body segmentation to predict specific anthropometric measurements (height, MUAC, head circumference, etc.). These predicted values are then compared against WHO/CDC growth standards to determine nutritional status, adding a layer of interpretability to the model's output.

For baseline evaluations, the researchers tested several standard architectures, including ResNet (18, 50, 101), VGG (16, 19), and DenseNet, finetuning them with additional convolutional layers and dropout to optimize performance.

Datasets and Visual Evidence

The system was trained and validated using three distinct datasets:

  • MalDB: 3,090 internet-curated images annotated by professional doctors, featuring diverse skin tones and uncontrolled settings.
  • IITJ (Adult): 479 images used for transfer learning to help the model learn basic body proportions.
  • Child Health Dataset: The primary research dataset, containing 2,729 images from 334 children, paired with physical ground-truth measurements and clinical labels.
Example of a malnourished child from the study dataset, illustrating the physical markers the AI seeks to identify.
Example of a healthy child from the study dataset used for binary classification training.

Key Results

The researchers established baseline accuracies for binary classification across their datasets:

  • MalDB: ResNet101 achieved the highest accuracy at 71.65%.
  • Child Health Dataset: ResNet50 performed best with an accuracy of 67.34%, while VGG16 followed closely at 68.5%.
  • IITJ (Adult): VGG19 outperformed other models on the adult dataset with 64.70% accuracy.

While these accuracies are modest for a clinical tool, they serve as a proof-of-concept for a fully automated, image-based screening pipeline that can function in diverse environments.

Conclusion and Future Work

NutriAI demonstrates the potential of combining "black-box" deep learning classification with "white-box" anthropometric regression to create a more trustworthy decision-making process for healthcare providers. Future work will focus on improving segmentation accuracy, expanding data collection to more diverse geographical regions, and conducting field pilots with community health workers to refine the tool for real-world deployment.

Summary updated

BibTeX
@inproceedings{khan2023nutriai,
  author = {Misaal Khan and Shivang Agarwal and Mayank Vatsa and Richa Singh and Kuldeep Singh},
  title = {{NutriAI: AI-Powered Child Malnutrition Assessment in Low-Resource Environments}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.24963/ijcai.2023/708},
}
Figure