(P-IT-13) Use of Pulse Oximeter and Pulse Rate Time Series Processing to Understand Physiologic Changes Associated with Transfusion in Very Preterm Infants

Background/Case Studies: Current strategies to treat anemia of prematurity are based on hemoglobin thresholds for transfusion and do not account for measures of physiology. There is limited evaluation of newer signal processing approaches to determine how cardiorespiratory physiology is altered by anemia and both gestational and postnatal maturation.

Study

Design/Methods: We analyzed prospectively collected data from very-preterm infants < 1250 grams who underwent longitudinal monitoring will pulse oximetry. Signal processing approaches were applied to time series measurements of pulse rate and SpO2 acquired every second. Neonates who received a transfusion and had at least 80% of the 24 hours immediately following the transfusion available during a monitoring period were selected for target data. The last monitoring session, per patient, that satisfied this requirement was included in the target data. Control neonates were selected from patients who had at least 80% of data during a 24-hour interval within a monitoring period, as long as a transfusion was not received during the period or in the 24-hour period before the session or the 24-hour period after the end of the session. The first monitoring period with sufficient data, per patient, was utilized, and the last 24 hours, in the session, were selected for processing. tsfresh, a python programing language package, extracted 1566 features, total, for the pulse and SpO2 data for the entire 24-hour period. For each of the physiological data streams (Pulse and SpO2), features were extracted using statistical and time-frequency domain methods. A detailed outline of the features being extracted can be found. After the feature extraction, we performed a significance test based on the p-values with a significance level of 0.0001, resulting in 35 final features.

Results/Findings: We evaluated 111 monitoring sessions from 73 infants. 46 of the evaluations were post transfusion. Support Vector Machines, Linear Regression, and XGBoost machine learning models were evaluated on identifying the post transfusion time period. XGBoost had the best performance, using 5-fold cross validation, with AUROC of 0.821. Most of the most influential features in the model were derived from the SpO2 measurement, although pulse rate did contribute. The top feature was from the frequency domain, but the other top 8 of 9 were in the time domain.

Conclusions: Signal processing measures of pulse rate and pulse oximetry measurements were incorporated in machine learning models to identify post transfusion differences in hemodynamics among neonates. These tools may be relevant in identifying infants with altered physiology due to anemia to guide more precise treatment with transfusion.

Importance of research: Use of signal processing on pulse oximeter-based heart rate and oxygen saturation show potential to inform on physiology related to efficacy of transfusion. Findings suggest the potential for signal-processed pulse oximetry data to identify infants with altered physiology due to anemia which can be used to evaluate treatment.