436 - Optimizing Near-Infrared Spectroscopy Source–Detector Separation for Noninvasive Neonatal Blood Pressure Monitoring

Publication Number: 3421.436

Animesh Giri, Brigham and Women's Hospital, Boston, MA, United States; Ceilidh Smith, Brigham and Women's Hospital, Boston, MA, United States; Julia Davis, MassGeneral Hospital for Children, Cambridge, MA, United States; Tina Steele, Brigham and Won=men's Hospital, North Billerica, MA, United States; Myrene Johnson, Purdue Univeristy, Boston, MA, United States; Deborah Cuddyer, Brigham, Boston, MA, United States; Marco Renna, Massachusetts General Hospital, Boston, MA, United States; Zachary Starkweather, Massachusetts General Hospital, Andover, MA, United States; Hoda El-Shibiny, Brigham's Women Hospital, Boston, MA, United States; Maria Angela Franceschini, Massachusetts General Hospital, Harvard Medical School, boston, MA, United States; Emily M. Herzberg, Mass General Brigham for Children / Harvard Medical School, Lynnfield, MA, United States; Mohamed El-Dib, Brigham and Women's Hospital / Harvard Medical School, Boston, MA, United States; John Sunwoo, Brigham and Women's Hospital / Harvard Medical School, Boston, MA, United States

Background: We have developed and are evaluating FlexNIRS, a wearable, wireless LED-based near-infrared spectroscopy (NIRS) device that integrates deep learning models for continuous, noninvasive blood pressure (BP) monitoring in neonates. FlexNIRS acquires NIRS-photoplethysmography (NIRS-PPG) signals through three source-detector separations (SDS: 19, 25, and 31 mm), each probing different tissue depths-from superficial layers to deeper cerebral regions (Figure 1). However, variations in SDS can alter both signal quality and BP estimation performance; determining the optimal distance may guide future probe design and miniaturization.

Objective: To determine whether deeper or superficial NIRS-PPG signals differ in signal quality and contribute differently to mean arterial pressure (MAP) prediction performance.

Design/Methods: : FlexNIRS-PPG and clinically deployed arterial line signals were collected from infants admitted to neonatal ICUs. In each SDS, we computed and compared the following four metrics: 1) Pulsatile-to-Noise Ratio (PNR) of the band-passed NIRS-PPG signal (1.5-25 Hz); 2) percentage of rejected data points after applying motion artifact rejection algorithm; and 3) Spearman correlation (r) and 4) mean absolute error (MAE) between true (arterial line) and predicted MAP values from our deep learning model running on bidirectional long short-term memory (LSTM) and convolutional layers. 20 random seeds were used for MAP prediction per channel per subject (Figure 2).

Results: Among fourteen enrolled infants, five were analyzed, each with approximately four hours of data. The PNR was −2.43±3.27 dB, −2.44±3.63 dB, and −1.78±3.31 dB; motion artifact rejection was 17.5±18.1%, 20.3±19.2%, and 17.9±22.8%; MAE was 2.38±1.55, 2.42±1.63, and 2.72±1.54 mmHg; and r was 0.56±0.17, 0.58±0.09, and 0.49±0.17, for 19, 25, and 31 mm, respectively. Across all evaluated metrics, none of the SDS channels demonstrated superior performance (Figure 3), and Tukey's post hoc tests confirmed no statistically significant differences between all metrics.

Conclusion(s): Preliminary findings show comparable MAP prediction performance across all SDS channels (19, 25, and 31 mm). Ongoing enrollment will increase statistical power and validate these results. If performance remains consistent across depths, future designs could utilize only the 19 mm channel, enabling a smaller, lower-power consumption probe without compromising accuracy.

FlexNIRS device probe attached on infant's forehead with approximate path of IR collected by each channel
Detectors on the device are placed at 19mm, 25mm and 31mm distances from the source to collect PPG from different depth across cerebral layers.

Signal quality and model performance metrics are collected across processing pipelines to compare 19mm, 25mm, and 31mm SDS
Heart rate band in frequency domain is calculated and is used to calculate PNR. Motion artifacts are rejected from the band-passed signals and percentage of artifact rejection is calculated, and MAE and correlation(r) are collected while testing the prediction model.

Metric distributions across the three SDS channels for each subject
Statistical testing using Tukey-HSD multiple comparison on Percent of motion artifact rejection, PNR, MAE and Correlation r validate there is no significant difference between 19,25, and 31mm channels.