Development of a Low-Cost Near Infrared Device for Measure of Subcutaneous Fat for Newborn Malnutrition

Abstract

The World Health Organization (WHO) reported that half or more of all under five deaths were caused by undernutrition in developing countries with most of these deaths occurring in the first week of life. Body fat percentage has been recognized to be closely associated with undernutrition in neonates. There is a need for a medical device to measure body fat percentage.

While such devices do exist, they are either very expensive or require highly trained professionals. This means that in developing countries, where the device is most needed it is not available. Our solution to this problem has been to use Near-Infrared Spectroscopy (NIRS) to measure fat levels in newborn babies. A study was conducted in Soweto, South Africa where measurements were taken on 650 newborn babies ranging from 0 to 2 years old using our NIRS device. With 164 subjects receiving acceptable deuterium dilution and DXA results, we were able to use a 4-compartment model to predict fat mass. We developed a model to predict fat mass from NIRS. The feature selection was undertaken on a randomized division of the dataset and then evaluated on the remaining two fifths. The wavelength range we used was 850-1100 nm. Two wavelength ratios along with weight over length were used in the final model. The model was developed using the 4-compartment model predicted fat mass as the reference and evaluation model. An r value of 0.885 was reached with Bland Altman limits of agreement of -78.9 (-954.3, 796.6) measured in grams.

My role throughout the project primarily was a data scientist. While all the data was available, we were unsure how we wanted to create the model. Extensive research and testing had to be performed to determine the best approach, which ultimately landed with using the 4-compartment model to train the dataset. This thesis will give an overview on the background of measuring fat in newborns as well as explaining the approach to developing the NIR model to predict fat mass.