Jaundice is a serious disease which affects mostly newborn infants. High levels of bilirubin in the bloodstream are the main cause of this. These bilirubin levels, total serum bilirubin (TSB), need to be determined frequently, in order to detect early onset of jaundice and start treatment. The traditional method to determine the TSB is by analysing blood samples in a laboratory. This causes a high strain on the newborn. A non-invasive technique to determine the bilirubin levels is through the skin, using a transcutaneous bilirubinometry (TcB) device. This technique is preferred because it is non-invasive, cheaper and faster than the traditional method. However, TcB measurements are less accurate, especially for preterm infants. This group is also more susceptible to jaundice. Therefore it is of interest to increase the accuracy of a TcB measurement.
This study uses the data of n = 101 newborn infants that are preterm (median gestational age: 30.5 weeks, range: 28.0 to 35.7). This data includes TcB measurements on five body locations and patient characteristics, along with the actual TSB measurements. Machine learning is applied to map the TcB measurements more accurately to a TSB value. Two models have been realized: a linear regression model and a decision tree. The root mean square error (RMSE) of the linear regression model is 21.9μmolL−1, and that of the decision tree is 30.4μmolL−1. The specified error of the TcB measurement device for preterm infants without phototherapy is 27.4μmolL−1, and 39.0μmolL−1 after phototherapy. The data used to train the models contains a mix of measurements with and without phototherapy. The error of the measurement data ranges from 30.2 to 85.2 μmol L−1, depending on the body measurement location.
The linear model reduces the number of unnecessary blood samples from 201 to 69, and has an RMSE that is lower than the specified error of the TcB device, and can therefore be accepted as a valid model to predict TSB values.
To join the presentation via Microsoft Teams click here