Weaning Failure Prediction from Heterogeneous Time Series using Normalized Compression Distance and Multidimensional Scaling

Abstract

Abstract¿Scientific evidence has shown that a failed weaning, defined as the process of gradual reduction in the level of mechanical ventilation support, increases the risk of death in prolonged mechanical ventilation patients. Different methods for weaning outcome prediction have been proposed, using variables and time series from the monitoring systems, however, monitored data are often non-regularly sampled, hence limiting its use in conventional automatic detection systems. In this work, we propose the joint use of two statistical data techniques, the Normalized Compresion Distance (NCD) and the Multidimensional Scaling (MDS), to deal with the data heterogeneity in the monitoring systems for failure weaning prediction. NCD technique determines a similarity measure between two sequences from compression length and entropy principles, whereas MDS provides each sequence with a point in an N-dimensional space suitable for training a classifier for weaning prediction. A total of 104 weaning events were collected in 253 patients under mechanical ventilation from Intensive Care Unit of Hospital Universitario Fundaci¿on Alcorc¿on; for each weaning event, 20 time series (TS), 15 clinical laboratory parameters (CLP) and 12 general descriptors (GD) were collected during 48 hours previous to the weaning event. Only 18 TS could be considered as candidates to the classifier input space, and one of them, diastolic blood preassure, reached significant results by itself, yielding 90.4% of accuracy prediction. These results show that weaning prediction systems can be designed with NCD and MDS, providing a compact input space to the classifiers.