808 - Effect of Gestational Age on a Touchless Measure of Neonatal Activity: A Preterm/Term Analysis

Publication Number: 808.4822

Paul S. Addison, Medtronic, Edinburgh, Scotland, United Kingdom; Dale R. Gerstmann, Retired, Orem, UT, United States; Rangasamy Ramanathan, Cedars Sinai Guerin Children's, Cedars Sinai Medical Center, Los angeles, CA, United States; Manoj Biniwale, Cedars Sinai Medical Center, Los Angeles, CA, United States; Rena Nelson, Wasatch Neonatal Research, Orem, UT, United States; Jeffrey Clemmer, Timpanogos Regional Hospital, Orem, UT, United States; Dean Montgomery, Medtronic, Edinburgh, Scotland, United Kingdom; Mridula Gunturi, Medtronic, Edinburgh, Scotland, United Kingdom

Background: Neonatal activity varies with physiological state, level of sedation or underlying pathology and may be a marker of disease onset and progression. Additionally, motion noise in physiological signals may lead to false or failed reporting of vital signs such as heart rate or respiratory rate. Given these considerations, we developed a novel neonatal activity monitoring technology using a machine learning algorithm trained on data from depth sensing camera. This non-contact, or touchless, technology allows continuous monitoring without attaching probes to the neonate.

Objective: To compare the performance of the neonatal activity monitoring algorithm between preterm and term neonates.

Design/Methods: Depth data was collected from 61 neonates (N=32 LA Site / N=29 Utah Site) using an Intel RealSenseTM D415 camera. The data was divided into term and preterm age groups corresponding to ages < 37 and >=37 weeks respectively (N=43 preterm / N=18 term). Preterm infant gestation ranged from 26.4 to 36.7 weeks. The average chronological age was 18 days for preterm and 6 days for term infants. Up to 6 video clips were collected from each neonate resulting in a total of 248 videos (comprising 22 h, 4 m, 33 s of data) in which only the neonate was present in the scene. Neonatal motion was manually labelled using synchronous RGB video data. Figure 1(a) shows an RGB image of a participant in the study with the associated depth image (Figure 1(b)) and depth changes over 1-second (Figure 1(c)). Features based on depth changes corresponding to activity were input into a random forest machine learning model. Training and testing employed a leave-one-out-cross-validation (LOOCV) paradigm which provides an indication of the variability of the resulting model. We calculated the accuracy of each test set by comparing the labelled and predicted motion on a second-by-second basis.

Results: The accuracy results for all video clips are shown in Figure 1(d). The mean [standard deviation] of the calculated accuracies for the preterm versus term splits using the LOOCV were 0.95 [0.08] v 0.93 [0.15]. In total, only 26 of 248 (=10.5%) had an accuracy below 90% and this was split preterm 16 of 171 (=9.4%) and term 10 of 77 (=13.0%).

Conclusion(s): We found close agreement between the accuracies of both preterm and term infants which indicates that the activity algorithm is unaffected by gestational age. These preliminary findings suggest the potential viability of non-contact, continuous, neonatal activity monitoring across a wide age range. Further work with a larger cohort of patients will be required to confirm this finding.

Figure 1 - (a) RGB Image of Neonate. (b) The Corresponding Depth Image. (c) The Corresponding Depth Difference Plot indicating Degree of Motion in the Scene. (d) Test Accuracies.