Pooling SAX-BoP Approaches with Boosting to Classify Multivariate Synchronous Physiological Time Series Data

As the current standard practice of manually recorded vital signs through a few hours is giving way to continuous, automated measurement of high resolution vital signs, it brings a tremendous opportunity to predict patient outcomes and help to improve the early care. However, making predictions in an effective way is fairly challenging, because high resolution vital signs data are multivariate, massive and noisy. Inspired by the max-pooling approaches in Convolutional Neural Networks (CNN), we propose extensions of vanilla SAXBoP approach, called Pooling SAX-BoP to successfully predict patient outcomes from multivariate synchronous vital signs data. Our experiments on two standard datasets demonstrate the Pooling SAX-BoP approaches are competitive with the current state-of-thearts on multivariate time series classification problems. We also integrate Boosting algorithm as one of the most powerful ensemble learning approaches on the BoP representations to further improve the performance. Our experimental results on the clinical data demonstrate that our methods are accurate and stable for classifying multivariate synchronous vital signs time series data.