Numerical methods for improved signal to noise ratios in spatiotemporal biomedical data
Magnetocardiography (MCG) is a technique to measure the magnetic fields produced by electrical activity in the heart. The interpretation of MCG signals is difficult because of different disturbances and noise. Several methods have been suggested for noise reduction in MCG data such as averaging, pass or stop band filters, and statistical based methods, but a unified framework that takes into account different typologies of MCG signals (rest, stress, and patients with an already ICD– Implanted Cardioverter Defibrillator- implanted) using an adequate number of recordings is still missing. Consequently, the main aim of the thesis is to develop methods for noise and artifacts treatment. Due to the non-stationarity (NS) of the noise, the conventional ensemble averaging of the data does not yield the theoretical improvement. In order to overcome this problem an average procedure that ignores the noisiest beats is applied. The results of this averaging procedure confirms that in case of NS, the Signal to Noise Ratio (SNR) does not behave as expected, but reaches a maximum after a certain number of selected beats. Furthermore, a theoretical proof of this result is given. The second part of the thesis deals with techniques based on Blind Source Separation (BSS), as preprocessing step for the averaging procedure, in case of MCG signals with low SNR. Different BSS algorithms are compared in order to find the best one in terms of noise reduction, separation, and computational time for each data typology. A drawback of BSS techniques is the order of the sources that cannot be determined a priori; for this reason 3 methods (based on different statistical principles) have been developed for the retrieval of cardiac signals. The last part of the thesis deals with the application of BSS methods to a category of signals not yet analyzed: patients with ICD implanted. It is shown that it is possible to extract the cardiac signal also in such noisy data, although not automatically. The Temporal Decorrelation source SEParation (TDSEP) algorithm outperforms the other BSS methods. This thesis shows that, applying novel automatic routines for the removal of noise and artifacts, MCG data could be used in clinical environments.