Weaning outcome prediction from heterogeneous time series using Normalized Compression Distance and Multidimensional Scaling

In the Intensive Care Unit of a hospital (ICU), weaning can be defined as the process of gradual reduction in the level of mechanical ventilation support. A failed weaning increases the risk of death in prolonged mechanical ventilation patients. Different methods for weaning outcome prediction have been proposed using variables and time series extracted from the monitoring systems, however, monitored data are often non-regularly sampled, hence limiting its use in conventional automatic prediction systems. In this work, we propose the joint use of two statistical techniques, Normalized Compression Distance (NCD) and Multidimensional Scaling (MDS), to deal with data heterogeneity in monitoring systems for weaning outcome prediction. A total of 104 weanings were selected from 93 patients under mechanical ventilation from the ICU of Hospital Universitario Fundación Alcorcón; for each weaning, time series (TS), clinical laboratory and general descriptors variables were collected during 48 h previous to the moment of withdrawal mechanical support (extubation). The TS diastolic blood pressure variable provided the best weaning prediction, with an improvement of 37% in the error rate regarding the physician decision. This result shows that the joint use of the NCD and MDS efficiently discriminates heterogeneous time series.

► We propose a general, automated and auto-tuned system for weaning outcome prediction.
► We compare heterogeneous time series using Normalized Compression Distance.
► Then, we project the instances with Multidimensional Scaling to a vector space.
► Finally, a Partial-Least-Squares based method is used for classification.
► We improve the error rate a 37% when compared to the physician decision.