Five children die of malnutrition every minute. Such deaths are preventable, but one of the hurdles to stopping them is accurately identifying those in need. Normally, making the necessary measurements requires bulky equipment and trained specialists. Soon that could all be replaced by a mobile phone.
The idea comes from Kenya-based non-profit Kimetrica. They’ve been working on a system that uses artificial intelligence to detect a child’s level of malnutrition from a single photo. The system is called MERON – Method for Extremely Rapid Observation of Nutritional status – which they presented on 15 and 16 May at the AI for Good Global Summit in Geneva.
Advertisement
To prove the concept, Kimetrica first developed a prototype for adults. Using a dataset from the University of North Carolina Wilmington consisting of 60,000 photos of faces along with the person’s height and weight, they trained an AI to assess someone’s body mass index and weight category – underweight, normal, overweight, or obese – from their picture alone. Overall, the prototype had an accuracy of 78 per cent, which was enough to convince UNICEF that the project had legs.
Picture of health
UNICEF then helped fund a project at Kimetrica to focus on detecting malnutrition in children under the age of 5 in Kenya. But no dataset of Kenyan children along with their weight and height exists. So, at the beginning of this year the team piggybacked on other ongoing health surveys in the country to gather 4,000 new images to train their system.
Unfortunately, of the 4,000 images, nearly a quarter of them have turned out not to be useable. This is because the children were either crying or not looking straight into the camera, or the picture had poor contrast or lighting. Some also had an obstruction, such as a hand, in the way. The team is currently adjusting these pictures to try to make them useable, and they expect to find out how well the new system works in the next few weeks.
To diagnose malnutrition, a specialist is usually required to measure a child’s weight, height, and the circumference of a particular part of their upper arm. Getting this last one right, can be especially tricky, as it’s important to get the correct part of the arm and to have the tape measure at the correct tightness.
Children can often find this process stressful, moving around whilst it takes place. “Overall, a single examination can take between 20 minutes to 30 minutes,” says Anita Shah at Kimetrica.
Aid in conflict zones
But with MERON, the plan is to take just a single picture. That can then be uploaded to a server for processing, which will return the results a few seconds later. The results will say either that the child is well-nourished, or has moderate, severe, or acute malnutrition. Each of these categories have well-established treatments.
With this approach there would be no need to use trained professionals to carry out nutrition surveys, as local caregivers could easily take the photos. This has an additional advantage in conflict zones, where it can often be very dangerous to send people, but there is still a need to assess the situation.
MERON still needs to be fully tested and proved, but Shah says the hope is that it will be faster, more accurate, and more cost effective way to address malnutrition, which accounts for nearly half of all global deaths in children under five. In the future, Kimetrica hopes to extend it to cover different age groups and ethnicities as well.
Read more: AI helps answer thousands of health queries in Zambia via SMS
