Author:

Keywords:

Science & Technology, Life Sciences & Biomedicine, Critical Care Medicine, Public, Environmental & Occupational Health, General & Internal Medicine, CARE-UNIT PATIENTS, APACHE-II, PREDICTION, MORTALITY, SYSTEM

Abstract:

© 2010 by Nova Science Publishers, Inc. Objective In this study, we attempted to find differences in the dynamics of physiological variables of survivors and non-survivors using multivariate autoregressive (MAR) time series analysis techniques. Materials and Methods Patient data were taken from a previously published randomized controlled trial and included subjects with length of intensive care unit stay of at least 20 days [1]. Time series data on 11 physiological variables, measured once daily were used to create the MAR models. These included variables such as minimum and maximum body temperature, glycaemia, urine output, white blood cell count, urea concentration, plasma creatinine, total bilirubin, C-reactive protein and total protein level. The methodology was developed on data from 20 patients (training set) and validated on data from 16 patients with an intensive care unit length of stay of at least 20 days. Based on the MAR coefficients, i.e. the estimated parameters of the MAR models, impulse response curves were simulated to describe the contributions of a single variable to fluctuations in the other variables. These simulations were performed by varying each of the 11 variables by applying an impulse and calculating each time the impulse responses in the other 10 physiological variables. For the amplitude of the impulse we used two times the standard deviation of the variable. The distance between the eigenvalues of the MAR models and the center of the unit circle is related to the dynamics of the impulse response functions. Results The impulse responses of non-survivors had a tendency to be either unstable or to return to the initial level after a longer time than the responses of survivors. This could be explained by the fact that the eigenvalues of the non-survivors were closer to the unit circle than those of the survivors. The eigenvalues allowed us to distinguish survivors from non-survivors in the training set. An eigenvalue of 0.953 was the optimal threshold for which the best possible classification accuracy of 85% was obtained. This resulted in a sensitivity of 7/10 and a specificity of 10/10 for the classification of survivors and nonsurvivors. The results were confirmed in the validation set where a sensitivity of 5/8 and a specificity of 8/8 were reached. Conclusion The sensitivity and specificity of this approach in which a classification between survivors and non-survivors was made on the basis of simulated impulse responses, may permit its use as an individualized monitoring tool at the bedside to help medical doctors and nurses to make their decisions.