459 - Association of RsO2-SpO2 coherence with intraventricular hemorrhage in very preterm infants: A prospective observational study using machine learning

Poster Number: 459
Publication Number: 459.342

Souvik Mitra, Dalhousie University & IWK Health, Halifax, NS, Canada; Timothy Bardouille, Dalhousie University, Halifax, NS, Canada; Tara R. Hatfield, IWK Health, Halifax, NS, Canada; Muhaimen Siddiqui, EVERSANA, Markham, ON, Canada; Walid El-Naggar, Dalhousie University, Halifax, NS, Canada; MARSHA CAMPBELL-YEO, Dalhousie University School of Nursing, Faculty of Health, Halifax, NS, Canada; Georg Schmolzer, University of Alberta Faculty of Medicine and Dentistry, Edmonton, AB, Canada; Douglas D. McMillan, Dalhousie University, Calgary, AB, Canada

Background: Cerebral hypoxia and disturbance of cerebral autoregulation are implicated in the pathogenesis of intraventricular hemorrhage (IVH) which is associated with cerebral palsy and cognitive impairment in infants born very preterm. However, there are currently no reliable tools to predict IVH occurrence.

Objective: To determine the coherence of regional cerebral oxygen saturation (RsO2) with other trended vital parameters [peripheral oxygen saturation (SpO2), heart rate and hemoglobin index (HbI)] in the first 72 h of age and their ability to predict IVH in very preterm infants.

Design/Methods: A single center prospective observational study was conducted between 2018-2021 at the IWK Health, Halifax, Canada. Infants born at < 32 weeks of gestational age (GA) were included in the study. Simultaneous cerebral oximetry and pre-ductal pulse oximetry (time-stamped data recorded every 4 seconds) using the NONIN SenSmart(TM) device were initiated immediately after birth and continued for the first 72h. RsO2, SpO2, heart rate and HbI data were downloaded after the study period, and potential artifacts were manually eliminated by a study team member blinded to clinical outcomes prior to analysis. Cross-correlation of RsO2-SpO2, RsO2-heart rate and RsO2-HbI were computed using wavelet coherence analysis that measures how consistent the phase difference between two signals remains over time. We further used machine learning (ML) using a support vector machine (SVM) to predict occurrence of IVH based on the average coherence values. A linear SVM was used with standard scaling of features, in a ten-fold cross-validation framework.

Results: We included 63 infants (GA 28±2.6 weeks) of whom 29 (46%) had IVH of any grade (Table 1). High coherence was observed between RSO2 and SpO2 signals (Figure 1). No other pairs, besides RsO2-SpO2 showed clear peaks in fraction of study period with significant coherence.
ML using linear SVM showed that average coherence values for RsO2-SpO2 could be used to predict if an IVH occurred with a classification accuracy (CA) of 64% (+/-5%). CA values for each pair of signals is shown in Table 2. Classifying severe IVH, or death, was not possible given the low incidence for these events.Conclusion(s): There is high in-phase synchrony between RsO2-SpO2 signals in very preterm infants in the first 72h. Using ML, average coherence of RsO2-SpO2 had the best combination of classification accuracy and certainty for prediction of IVH. Future research should focus on exploring other ML models to improve the predictive ability of RsO2-SpO2 for IVH to explore its potential use in clinical practice.
Table 1. Demographic profile of participants with and without intraventricular hemorrhage (any grade)IVH, Intraventricular hemorrhage
Figure 1. Analysis of coherence between the RsO2-SpO2 signalsThe analysis measures the synchronization of two signals, as a function of the signal period [represented in the X-axis]. Synchronization is measured using a mathematical construct called coherence, which measures how similar two signals are, over a range of frequencies (or periods).
The magnitude of coherence (0 to 1, as a fraction of the period of observation denoted on the Y-axis) denotes the strength of synchronization of the two signals. The phase of coherence denotes how the signals are related for a given period and time interval. The two signals can be:
(1) In-phase – they oscillate up and down together;
(2) Out-of-phase – one is going up when the other is going down;
(3) Leading – signal 1 is ahead of signal 2;
(4) Following - signal 2 is ahead of signal 1;
Coherence between the RsO2-SpO2 signals was significant for 25% of the study period, and the signals were in phase with each other during these periods with 4 to 80-minute period oscillations that synchronized between the two signals.