760 - Noncontact Growth Monitoring Using Radar-Based Machine Learning in Neonates

Periodic measurement of height and weight in newborns is used not only as a criterion for evaluating growth status but also as the most basic indicator for determining whether nutrition is well supplied and whether there is no edema. However, especially in the case of newborns, there is a risk of having to move the child to accurately measure height and weight, which often requires a lot of effort and time.

Objective: In this study, we propose a method of measuring the height and weight of a newborn baby more conveniently, safely, and accurately by applying a machine learning technique using data obtained by a non-contact radar sensor.

Design/Methods:

A total of 9 newborns were included in the study, and height and weight were measured using frequency modulated continuous wave (FMCW) radar. The experiments were conducted in the neonatal intensive care unit (NICU) and neonatal unit in a supine position. Signature images for machine learning were created from the radar signals, and a convolutional neural network (CNN) was designed for height and weight measurements. In the case of height measurement, to reduce the error, the radar signals from the entire body and bent-legs part of the newborn were separately extracted and then integrated with CNN. In the case of weight measurement, radar images and beforehand measured height values were mixed with CNN.

Results:

In the case of radar-based newborn height measurement, the mean absolute error (MAE) and root mean square error (RMSE) were obtained 1.61 cm and 1.41 cm compared to the actual measurements, respectively. Intra-class correlation coefficient (ICC) was 0.82, and the p-value was 0.02, which confirms a significant correlation between the actual and the radar measurements. The weight measurement showed a slightly more significant difference than height. The MAE, RMSE, and ICC were 212 grams, 212 grams, and 0.86, respectively. In the case of body weight, the p-value was 0.52, which shows the insignificant correlation between the actual and the radar measurements.

Conclusion(s):

Despite the small number of newborns and the small number of experiments, we were able to measure the height and weight of newborns relatively accurately using the radar sensor. It is expected that better results can be achieved if a large dataset is collected from a large number of newborns and if a large dataset is accumulated continuously for an extended period for the same baby. It is also expected that continuous growth monitoring will be possible in NICU and even in home environment use cases.