Time-series analysis techniques combined with Gaussian Process Classifiers for prediction of clinical stability after coronary bypass surgery

In this study we present a combination of time-series analysis tools and a machine learning algorithm (Gaussian Process classifier) for the task of predicting the time frame in which the minimal clinical conditions of stability to start weaning of mechanical ventilation are reached. We perform a retrospective analysis of clinical data obtained from a Patient Data Management System of 103 elective coronary bypass surgery patients. Four hours of ICU data of 14 physiological variables, was used as input for five different time-series analysis models. A Gaussian Process Classifier, with the parameters of the calculated models as inputs, assigned to each patient a probability of belonging to the defined classes for clinical stability: within the first 8 hours, between 8 and 16 hours, between 16 and 24 hours, and after 24 hours. Including parameters of different types of time-series models as a representation of the time-varying signals, we incorporate knowledge of the dynamical behaviour of the patients. As a result we obtained aROCs above the medical requirements of 0.8 for some of the classes and above 0.7 for all classes. The use of the dynamics captured by the model representations led to increased performance in further ahead predictions.