Morphology Analysis of Physiological Signals Using Hidden Markov Models D. Nov´ ak, L.Lhotsk´ a Department of Cybernetics Czech Technical University in Prague xnovakd1@hpk.felk.cvut.cz T.Al-ani,Y. Hamam A2SI Laboratory Group ESIEE-Paris, France alanit@esiee.fr D.Cuesta-Frau, P. Mic´ o Department of Computer Science Polytechnic University of Valencia, Spain dcuesta@disca.upv.es M.Aboy Biomedical Signal Processing Laboratory Portland State University, USA mateoaboy@ieee.org Abstract We describe a clustering algorithm based on continu- ous Hidden Markov Models (HMM) to automatically clas- sify both electrocardiogram (ECG) and intracranial pres- sure (ICP) beats based on their morphology. The algorithm detects, classifies and labels each beat based on morphol- ogy. In order to avoid the numerical problems with classi- cal Expectation-Maximization (EM) algorithm we apply a novel method of simulated annealing (SIM) for HMM op- timization. We show that better results are achieved using simulated annealing approach. 1. Introduction Computer-aided medical applications is a field of enor- mous development in recent years. One of these appli- cations consists of extracting significant information from raw data like in case of Holter ECG signals and intracranial pressure signals. Holter signals are ambulatory long-term ECG registers used to detect heart diseases which are difficult to find in normal electrocardiograms. These signals normally include a quantity of beats greater than 100000, and doctors must visually examine all of them in order to find possible ab- normal beats. Examining every beat present in the Holter register is time consuming, and is quite likely some beats could be omitted in the visual inspection because of subjec- tive reasons. Nevertheless, these signals include many sim- ilar beats, and only a few are different, namely, most of the time doctors are examining the same kind of beat. There- fore, it would be very useful to have a method to simplify the Holter register prior to its visual inspection. Traumatic brain injury (TBI) remains a significant cause of mortality and morbidity in both children and adults. TBI often leads to increased intracranial pressure (ICP) that may result in worsening brain injury and outcome. Several re- searchers, however, have performed preliminary studies on the ICP beat morphology and suggest that the ICP beat may contain indirect information about the intracranial compli- ance [5]. We introduce an automatic clustering algorithm based on continuous HMM which can be used to perform mor- phological analysis. In the past, cluster analysis techniques have focused on data described by static features. In many real applications, the dynamic characteristics, i.e., how a system interacts with the environment and evolves over time, are of interest. Such behavior or characteristic of these systems is best described by temporal features whose values change significantly during the observation period, like in case of electrocardiogram or intracranial pressure signals. An HMM is a very suitable tool for coping with temporal information [8]. This stochastic based modelling approach has found its use in many application areas including speech processing, molecular biology and robotics. Due to the numerical and initialisation problems we cast the problem of HMM opti- mization under the framework of simulated annealing [4]. 2. Method 2.1 Hidden Markov Models An HMM is a stochastic finite state automata defined by the following parameters λ =(A, p, B), where A is a state transition probability, p is the initial state probability and B is the emission probability density function of each state is defined by a finite multivariate Gaussian mixture. Each model can be used to compute the probability of observing