Sudden death is a natural, unexpected and rapid death that occurs within a time frame of 24 Hours after the beginning of the pathological process. The most common causes of sudden death are Cardiovascular Diseases (CVD) that happen to be among the leading causes of death worldwide. In 2012, the World Health Organization (WHO) recorded 17.5 million CVD deaths, accounting for 31% of all deaths recorded in the world. One of the cardiovascular diseases with the highest mortality is Ventricular Fibrillation (VF), which is a cardiac arrhythmia caused by a disorganized electrical activity in the heart. During VF, the ventricles contract asynchronously causing an absence of an effective beat as well as a failure of blood pumping. This could result in a sudden death of the patient, if not treated properly within a few minutes. Defibrillation is the only definitive treatment of VF. It consists of the application of a high-energy electric shock to the patient's chest, thus restarting the normal cardiac electrical activity. The success of defibrillation is Inversely proportional to the interval of time it takes from the beginning of the episode to the application of the electrical charge. There are many difficulties in diagnosing VF: on the one hand there are some Intrinsic characteristics of the VF (lack of organization, irregularity, etc.), and on the other, there is a great similarity between VF and other cardiac pathologies including Ventricular Tachycardia (VT). The differentiation between VT and VF is quite complex. A wrongful diagnosis of VT instead of VF can cause serious injuries when applying the therapy to the patient. VT treatment consist of giving a high voltage electric shock to the patient, which could cause a VF. Conversely, if the VT is incorrectly interpreted as a VF, the result might be also dangerous for the patient´s life. Therefore, an effective detection method to distinguish VF from VT is very important in clinical research. In order to diagnose and treat cardiovascular diseases (for example VF), two major groups of diagnostic methods are established in cardiology: Invasive and noninvasive diagnoses. Invasive techniques require the introduction of catheters into the body, to measure the pressures in the cardiac chambers in order to record the electrical activity. Non-invasive techniques are focused on characterizing the pathophysiological state of the heart through electrodes placed directly on the patient's skin. The electrocardiogram (ECG) is a non invasive, low cost, examination which has been used as the basic diagnostic method for cardiac conduction disorders, by studying the frequency and the morphology of different waves that constitute the cardiac cycle. The ECG analysis is a good source of information to detect different types of heart disease. Due to the fact that the ECG signal is a non-stationary random signal, the time domain analysis does not prove to be sufficiently sensitive to the distortions of the waveforms on the ECG. However, these methods not always show all the information that can be extracted from ECG signals. Thereby, the frequency information, which shows additional data from the signal, is lost. Diagnosis within the frequency domain uses methods such as the fourier transform.Therefore, an analysis in the frequency domain allows to determine the frequencies within the ECG signal. On the other hand, using a frequency domain method losses the time information from the signal, resulting in a very limited method which is not useful for the analysis of non-stationary signals. Several studies have used mathematical models that combine time and spectral information in the same representation. This is Time-Frequency Representation (TFR) technique, very important in the treatment of non-stationary signals such as the ECG signal, as it distributes signal energy in a two-dimensional time-frequency space. In addition, multiple factors affect the acquisition and recording of the ECG signal: on the one hand there is the influence of the environment (50-60 Hz interference from the electrical grid, baseline, etc.), and on the other hand, physiological disturbances such as an electromyography (EMG). In the last decades, ECG noise reduction has been one of the main fields of research, since an adequate reduction of noise allows a good pre-processing of the signal, extracting from it the maximum amount of relevant information while eliminating the irrelevant one. The use of time-frequency representation images (t-f) as the direct input to the classifier is the novelty of this doctoral thesis. It is hypothesized that this method makes it easier to improve the classification results by eliminating the extraction of typical characteristics, and their corresponding loss of information. They are also used for evaluation and comparison with other authors. Materials and Methods: MIT-BIH´s Malignant Ventricular Arrhytmia and AHA (2000 series) standard data bases have been used to obtain the records of ECG signals, creating from them one set for training and one for testing the classification algorithms used. A total of 24 monitoring records (22 MIT-BIH records plus two additional from the AHA data) have been used, with a sampling frequency of 125 Hz. As part of the method a filter algorithm has been implemented in order to reduce the base line. A report that indicates the beginning of the time Window (tW) of the ECG signal is developed. First, the Hilbert Transform (HT) was applied to each window obtained. Then the RTF was applied as input to the five classifiers: Artificial Neural Networks Classification (ANNC), Smooth-type support vector machine (SSVM), bagging (BAGG), 𝐿2Regularized Logistic Regression (L2_RLR) and K Nearest Neighbors (KNN).Their respective training and individual tests were made by choosing the KNN classifier for best quality detection. The RTF was converted into an IRTF and a reduced dimensionality IRTF by means of image size reduction techniques (Average Intensity of the Pixels (AIP), KNN, Bilineal, Bicubic) and the techniques of Selection of Characteristics (SC), specifically the Sequential Forward Selection technique (SFS). Combinations of the classification algorithms were performed when applied over the same set of data (TFRI, reduced TFRI, SC), in order to compare their behavior with the results obtained from the individual classifiers. These strategies and methodologies were compared with different methods cited in the literature to verify to what extent the results obtained support our hypothesis. Results: using performance evaluation parameters (sensitivity, specificity and accuracy) as a methodology to validate the results obtained by each one of the strategies employed, it was found that the KNN classifier achieves the best performance for TFR. For 'VF' it was obtained a sensitivity of 94.97% an overall specificity of 99.21%, and an overall accuracy of 98.47%. Whereas for 'VT, it was obtained a Sensitivity of 93.47%, an overall specificity of 99.39%, and an overall accuracy of 98.97% within a time of execution 0,1763s. The main results of the classification for the detection of VF obtained using the reduced TFRI technique (AIP) as input to the KNN classifier, showed an 88.27% of sensitivity, a 98.22% of overall specificity and a 96.35% of overall accuracy. For ´VT', it was obtained an 88.31% sensitivity, a 98.80% for global specificity, and a 98.05% of global accuracy within a runtime of 0.024s. When the classifiers were combined, it resulted that the ANNC_KNN_ANNC combination when using the Hierarchical Method (MJ) gave a sensitivity of 92.6%, a specificity of 98.68% and an overall accuracy of 97.52% for 'VF'. Whereas a Sensitivity of 89.3%, a specificity of 99.01%, and a 98.32% of overall accuracy for 'VT 'were obtained with a runtime between [0,0487s; 0,0492s]. Conclusions: the classification of VF and VT discrimination showed how the different classification algorithms (both individual and combined) behave over the same set of characteristics. The results obtained by the ANNC_KNN_ANNC combination of algorithms using the reduced TFRI data were better at detection when compared to those obtained by the algorithms used individually as well as to other multi classifiers applied over a single set of reduced TFRI data. In addition, when comparing the results of the combination with those obtained by the KNN using the non-reduced TFR and the TFRI, the latter are slightly lower. On the other hand, a shorter execution time is obtained, so the method is more useful for Real-time detection systems. After a long analysis it is possible to conclude that the proposed methodology provides useful information for the detection of VF with a low computation time. Furthermore it can separate satisfactorily the VF from the rest of cardiac pathologies at the time of diagnosis, significantly improving the patient's ability to be effectively treated when presenting an episode with one of these arrhythmias. Thus making this work a source of clinical help in the diagnosis of arrhythmias.
