What type of study is this?

September 10, 2025Open Access

Machine Learning-Driven Early Detection of Childhood Malnutrition in Low-Resource Settings: Integrating Anthropometric Imaging and Socioeconomic Data

Key Points

The ML-driven model achieved a remarkable accuracy of 92.5%, significantly improving malnutrition detection outcomes.
Using convolutional neural networks for image analysis, the study reported precision and recall rates of 0.89 and 0.91 respectively.
A comparative study found that the hybrid machine learning model outperformed manual screening by 14.2% in accuracy and reduced false negatives by 21.8%.
The framework shows potential for affordable deployment on low-cost devices like Raspberry Pi, increasing access to reliable malnutrition screening.

Abstract

Malnutrition continues to be a significant global health issue, impacting approximately 148.1 million children under five years old in 2024 (WHO). Traditional screening techniques, like mid-upper arm circumference (MUAC) and body mass index (BMI), provide moderate coverage but face limitations in rural low-resource regions because they require trained personnel and equipment. To fill this gap, we suggest an ML-based framework that combines anthropometric image analysis with socio-economic and dietary intake information for the early detection of malnutrition. The research was carried out on a dataset involving 2,000 children from 3 rural centers in Nigeria, including 6,000 anthropometric images, 2,000 dietary assessments, and 2,000 socio-economic profiles. Convolutional neural networks (CNNs; ResNet50, MobileNetV3) that were trained on images reached an accuracy of 84.7%, with a precision of 0.81 and a recall of 0.83. Ensemble methods (Random Forest, XGBoost, LightGBM) applied to tabular data reached 87.2% accuracy and a 0.85 F1-score. A blended fusion layer integrating both modalities enhanced the outcomes to 92.5% accuracy, 0.89 precision, 0.91 recall, 0.90 F1-score, and 0.95 ROC-AUC. The efficiency of deployment was evaluated on affordable edge devices. On a GPU workstation (16-core CPU, 32 GB RAM, 8 GB VRAM), the inference duration per sample was 0.18 seconds. On a Raspberry Pi 4 (4 GB RAM), MobileNet took 1.24 seconds per sample, whereas ResNet took 2.37 seconds per sample. A comparative analysis involving human nutrition workers (n = 10) showed that the hybrid ML model exceeded manual screening by +14.2% in accuracy and decreased false negatives by 21.8%, indicating possibilities for scalable deployment in community settings. These results emphasize the practicality of ML-driven malnutrition assessment as an affordable, precise, and resource-efficient approach. By directly supporting SDG 2 (Zero Hunger) and SDG 3 (Good Health and Well-being), the suggested system provides a means for enhancing nutritional monitoring and actions in marginalized communities.

Read Full Paperexternally

Ask AI

Helpful

Bookmark

View Full Paper