J.F. Martínez-Trinidad et al. (Eds.): CIARP 2006, LNCS 4225, pp. 954 – 963, 2006. © Springer-Verlag Berlin Heidelberg 2006 Support Vector Machine with External Recurrences for Modeling Dynamic Cerebral Autoregulation * Max Chacón 1 , Darwin Diaz 1 , Luis Ríos 1 , David Evans 2 , and Ronney Panerai 2 1 Universidad de Santiago de Chile; Departamento de Ingeniería Informática, Av. Ecuador No 3659 - Casilla 10233; Santiago-Chile mchacon@diinf.usach.cl, darwindiazh@gmail.com, lrios@diinf.usach.cl 2 Medical Physics Group, Department of Cardiovascular Sciences, University of Leicester, Leicester Royal Infirmary, Leicester LE1 5WW, UK dhe@le.ac.uk, rp9@le.ac.uk Abstract. Support Vector Machines (SVM) have been applied extensively to classification and regression problems, but there are few solutions proposed for problems involving time-series. To evaluate their potential, a problem of difficult solution in the field of biological signal modeling has been chosen, namely the characterization of the cerebral blood flow autoregulation system, by means of dynamic models of the pressure-flow relationship. The results show a superiority of the SVMs, with 5% better correlation than the neural network models and 18% better than linear systems. In addition, SVMs produce an index for measuring the quality of the autoregulation system which is more stable than indices obtained with other methods. This has a clear clinical advantage. Keywords: Support Vector Machine, biological signals, cerebral autoregu- lation. 1 Introduction Support Vector Machines (SVMs) have shown their usefulness by improving over the performance of different supervised learning methods, either as classification models or as regression models. But the small number of papers involving the prediction of temporal series or signal modeling [1-2] shows a lack of assessment in this respect. When these applications are restricted to the field of biomedical signals, SVMs are used as classical classifiers following a process of extraction of signal characteristics [3-4]. At this time we are not aware of any applications that use SVMs as recurrent structures for modeling biomedical signals. To evaluate SVMs in the field of biomedical signal modeling and prediction we have chosen to model the Autoregulation Blood Flow System (ABFS). Describing this system adequately is a complex problem for which currently there is no model that represents the phenomenon properly. This means that we can not have reliable methods that allow detecting, diagnosing and monitoring different cerebrovascular * This works was supported by FONDECYT, Chile, under project 1050082.