Time-Frequency Analysis of Hemodynamic Waveforms to Predict the Occurrence and Severity of Periventricular Leukomalacia

This paper is concerned with the prediction of the occurrence and severity of Periventricular Leukomalacia (PVL), a form of white-matter brain injury that occurs often in neonates after heart surgery. The data which is collected over a period of twelve hours after the cardiac surgery contains vital measurements. The fact that the exact cause of the PVL have still not been clearly understood renders a mathematical modeling approach for fault diagnosis impractical, if not impossible. Hence, the decision tree classification technique has been selected for its capacity for discovering rules and novel associations in the data. It classifies groups based on reducing uncertainty in the classified data. From a physiological point of view we know that there are several regulatory mechanisms responsible for fluctuation of the hemodynamic variables at different time scales. To discover the most important active physiological components which might lead to the occurrence of PVL and possibly affect its severity, we focus on the variation in the data in one minute, twenty minute and two hour periods. We calculate the energy of continuous wavelet transform coefficients of vital data at these time scales as a measure of variation in the different time frames. The results obtained from developing decision tree classifiers show that among all variations in all the variables, 2 hour and 20 minute variations in the heart rate, 1 minute and 20 minute variations in Oxygen saturation, and 2 hour variations in the mean arterial pressure are the most important parameters to be able to predict PVL occurrence.