Time series are an important class of data objects that arise from various sources and their analysis typically involves huge amounts of information requiring usage of data mining techniques. Measuring similarity in long time series plays an important role in searching for similar patterns, classification, clustering, prediction and knowledge discovery. In clinical context any estimation of future values based on its past values can be useful in disease prognosis. In this thesis different methods of measuring similarity between time series of arterial blood pressure (ABP) signals are described and experimental results are provided. To classify an ABP record within a particular diseases’ class (a cluster), the typical procedure is the prior determination of the similarity of the ABP record with a reference signal characterizing a cardiovascular disease (CVD) and then identifying the strength of that similarity to enable a true positive classification of the illness (or not). Several methods of measuring similarity among time-series are referred in literature, the most commonly employed one were object of this research. Since the goal was the application of the similarity results to perform clustering of the ABP signals, similarity methods were investigated particularly in what concerns their performance when proceeding for the clustering following step. So, this thesis reports the usage of seven different similarity methods, five working in the time domain and two in the transform-based domain, and explores their usage when clustering by Partitioning Around Medoids is implemented. As data records are noisy and signals suffer from variations due to other sources than heart, six types of variations were imposed on the reference signal and 20 degrees of possible variations were tested. The time series considered on this study were 10 seconds length, referring to healthy, electrocardiogram (ECG) long term ST’s, atrial fibrillation and a collection of diagnostic ECGs. Three clusters were considered, each involving healthy and pathological records, in different proportions. Results demonstrate that the Discrete Wavelet Transform using a Haar wavelet decomposition with the Karhunen-Loève transforms, besides reducing the computational processing load enables clustering with an accuracy between 76% and 84% among the three diagnostic classes considered. The organization of this thesis is as follows. A short representation of Time-series is in chapter.1. A brief description of various similarity methods and clustering methods are given in chapters 2 and 3. Experiments performed and results obtained are described in chapter 4. Finally, the conclusion of this work is presented in chapter 5 where the list of publications resultant from this thesis is included.