Abstract

    The importance of the electrocardiogram for the diagnosis of many heart diseases, both by visual inspection and by current techniques for automatic inspection is well known. As with any other signal, the amount of information extracted and its quality will depend on properties such as signal to noise ratio, resolution analog-digital converter, sampling rate, etc. For this reason, it is of great importance that the signal provides an adequate `quality', especially when the diagnosis of severe cardiovascular diseases depends on itself. Noise reduction in the electrocardiogram (ECG) has been one of the issues raised in the literature on ECG signal processing. There are very different ways of solve the problem and unfortunately a universally method applicable to all noise sources and all cases does not exist. In this thesis the main sources of noise that appear in the ECG recording have been studied. Some of these can be minimized in the acquisition phase of the signal to provide special care to certain standards or rules. However, others, such as muscle noise, or drift of the baseline and artefacts cannot be eliminated or minimized in the acquisition phase. In this case signal processing techniques for subsequent reduction to an acceptable level are necessary. One of the first options is the filtering of the signal, by linear or nonlinear filter to maximize the signal to noise ratio, for example, with the Wiener filter. When the spectrum of the interest signal and the noise are overlapping the filtering techniques become ineffective. In that case, other common techniques are applied to the signal averaging. Its use is effective as long as the signal and noise fulfil certain conditions. The adaptive filtering is another technique that has achieved acceptable results in the reduction of noise in the ECG, as it is able to adapt to changes in the ECG signal. However, due to this, this type of filtering may cause changes in the clinical information contained in the ECG. On the other hand, techniques based on wavelet transform have achieved an improvement over traditional filtering but not completely reduce interference. The thresholding used in some sub-bands with the intention of eliminating noise can affect the clinical information of the signal.
    As for cardiac pathologies that might contain the ECG under analysis there is one in particular, atrial fibrillation (AF). AF is one of the most common arrhythmias and causes the highest number of admissions in the casualty department of hospitals. Frequently, AF is associated with a cardiac pathology but, in most cases, it appears in patients without any detectable disease. Nowadays, the treatment and analysis of AF is not completely satisfactory. And the high levels of morbidity and mortality and associated costs give rise to many scientific works and publications about this topic. An efficient non-invasive study of this type of arrhythmias requires the isolated observation of the registered atrial activity (AA) that usually has low amplitude levels and is overlapped to ventricular activity, which makes the use of lineal filtering techniques impossible. Nowadays, there are several techniques that can extract the AA with a good performance blind source separation (BSS), spatiotemporal cancellation (STC) but poor results are obtained when the number of used reference signals (leads) is less than three, or when the duration of these signals is reduced. On the other hand, classic techniques, such as template match subtraction (TMS) and recurrent adaptive filtering, have developed AA extraction from only one lead, but these systems are very sensitive to the presence of ectopic complexes.
    This thesis doctoral addresses two main objectives: firstly, to design systems based on Artificial Neural Networks with its different types in order to reduce interference and noise in ECG signals preserving the clinical information. The second objective consists in checking the potential of neural networks, in all its forms, for the extraction of atrial activity from the ECG with atrial fibrillation, as an alternative to current techniques for cancellation of QRST complex. The results obtained in this thesis are an improvement in the quality of the ECG signal, which can enable a breakthrough in the subsequent clinical analysis.
    A database of artificial signals with several types of signals, synthesized, pseudo real and real AF episodes, has been created to validate the obtained results. The fact that noise signal and atrial activity are known and simulated in the first two types, makes the quality extraction and reduction measure possible, using correlation indexes between the reduction and extracted signal with the original one. The different behaviours of these artificial recordings show the advantages and the limitations of each method in each situation. Finally, all the algorithms have been tested over real signals diagnosed by cardiologists. In all cases, the comparison with results obtained using classical techniques of noise reduction and AA extraction, with the most accepted techniques by the international scientific community, will be a reference for proposed methods.