Incremental support vector machines for monitoring systems in intensive care unit

In this paper, we propose a new on-line monitoring system in intensive care unit (ICU). Two incremental versions of support vector machines (SVM) have been adapted to the current monitoring system. In fact, the monitoring system in ICU have many issues to detect real states of patients namely critical or normal states. It frequently generates false alarms that have bad effects on the working conditions. Our aim, in this paper, is to avoid false alarms by improving the current monitoring system. To this end, we apply two machine learning techniques namely the LASVM and I-SVM to the current system to improve its performance. Both of them are characterized by dealing with large amount of data streams. Besides, each technique provides its own optimal prediction model in order to correctly describe the patients' states over time. As a result, we get two on-line monitoring systems that classify patients' states to two main classes. The first class consists of patients with normal state with no need to specific care and the second one contains patients with critical state that have to be immediately hospitalized. Results of the two monitoring systems based on the LASVM and I-SVM, when using real-medical databases, show their performances. However, the I-SVM proves that it is slightly better compared to the SVM in batch mode, the LASVM and the current system.