Title: A strategy for identifying informative variables: Prediction of developmental outcomes of preterm neonates
Authors: Sergiy Pereverzyev - Johann Radon Institute for Computational and Applied Mathematics (RICAM) (Austria) [presenting]
Sergiy Pereverzyev-Jr - Medical University of Innsbruck (Austria)
Vasyl Semenov - Delta SPE LLC (Ukraine)
Abstract: The problem of detecting the most informative coordinates of inputs allowing an accurate reconstruction of the corresponding outputs is discussed. We are motivated by predicting neurodevelopmental outcomes of preterm neonates from the ratios of the amplitudes of the peaks of metabolite spectra. One of the difficulties of the above prediction problem is that the available clinical data contain only a few input-output pairs associated with neurodevelopmental impairments. The construction of a predictor from training data can be seen as the recovery of a multivariable function from its evaluations at given points. An attempt to approximate the ideal prediction function by using an insufficient number of variables generates an error above the best guaranteed one. This remark hints that informative variables can be detected by finding the minimum value of the prediction errors observed for the predictors from the considered input combinations. The predictors employed in our study are obtained by kernel ridge regression (KRR) with various input variables as regressors. In KRR we use universal Gaussian kernels and the kernels constructed from the data according to a recently proposed approach. In this way, we try to cover a variety of predictors exhibiting universality or specificity.