Project Lead(s): Armando Torre
Issue
Chronic malnutrition can severely affect children, causing lifetime consequences if not prevented in time.
In 2012, it was estimated that 20% of children in Peru under five years of age suffered from malnutrition.
To prevent malnutrition, nutritional status screening and monitoring are paramount. However, they are not routinely done and are subject to measurement errors.
Solution
The goal of this project was to develop an automated nutritional assessment system, based on the analysis of digital images captured with a mobile phone or tablet.
This system consists of an Android-based mobile application and a web application.
The mobile app is used to collect data and capture a set of body part images, which are then processed using image-processing techniques to obtain anthropometric data. These data, along with child gender and age, are submitted to a neural network to determine the nutritional status of children, based on the Z‑scores of the three most commonly used indices: weight-for-age (W/A), height-for-age (H/A) and weight-for-height (W/H).
The results are sent to the web application in order to store, analyze and visualize the data.
A total of 313 children under five were included in the study from 11 public and private schools in the districts of San Juan de Lurigancho, Los Olivos, Callao and San Borja.
Images were used in two different processes. The first 49 children’s images helped to determine the morphological traits, number of images, image acquisition methodology and image enhancement techniques to be used. The remaining 264 children’s images were used to evaluate image-based anthropometric measurements and for training of the neural networks.
Outcome
The team managed to develop a prototype of the nutritional system for the screening and monitoring of children less than five years of age, and were also able to implement the system on an affordable Android device.
The H/A indicator showed a low root mean square (RMS) value of 0.52, but other metrics were not very high, possibly because of the unbalanced distribution of classes in the sample (sensitivity 66.7%, specificity 88.9% and overall accuracy 96.6%).
The W/A indicator displayed better results compared to the other indices, mainly because the samples in the dataset had only two classes, normal and overweight (RMS 0.49, sensitivity 100%, specificity 100% and overall accuracy 100%).
The last indicator, W/H, showed slightly better results than the H/A indicator, because of the more balanced distribution of classes in the dataset and the lack of cases with moderate and severe wasting (RMS 0.58, sensitivity 84.1%, specificity 92.3% and overall accuracy 86.2%).
Results suggest this nutritional assessment system could aid public health programs to detect weight changes in children in a timely manner by getting accurate W/A results. Although sensitivity and specificity of H/A and W/H indicators are around 80%, overall accuracy was above 85%.
A larger sample of children of less than three years of age is felt to be necessary in order to generate specific neural network systems to estimate Z-score values for each age group.
The research team has not yet decided whether to apply for Transition To Scale (TTS) funding.