Reconocimiento de patrones en señales biomédicas para la detección temprana de eventos no fisiológicos

Permanent URI for this collection

https://repositorio.uai.edu.ar/handle/123456789/385

Browse

Browsing Reconocimiento de patrones en señales biomédicas para la detección temprana de eventos no fisiológicos by Title

  • Item
    Bayesian inversion approach for ECG denoising
    (Asociación Argentina de Matemática Aplicada, Computacional e Industrial (ASAMACI), 2019) Bergamini, Maria Lorena ; Liberczuk, Sergio Javier
    Stochastic or Bayesian filtering is an inverse problem in the sense that from given noisy observations we want to estimate hidden state variables knowing models for states evolution and measurement noises. In the present work we propose a Particle Filter method for denoising ECG signals based on Monte Carlo filter techniques estimating the state (filtered signal value) from noisy observations simulated with different SNRs. We use Mc Sharry dynamical model whose solution trajectories reproduce realistic ECG waves. The improvement in the denoised signal is higher when the SNR in the input signal is lower. Particle Filter method allows any noise distribution to be considered. This property is very interesting for physiological signal processing, where the noise is often complex and non Gaussian.
  • Item
    Heart beat parametric modeling based on Monte Carlo fitting techniques
    (Sociedad Argentina de Bioingeniería, 2019-3-28) Liberczuk, Sergio Javier ; Bergamini, María Lorena ; Arini, Pedro
    Synthesis of electrocardiogram (ECG) signals is closely linked to the modeling process since precise knowledge of the parameters of the heartbeat to be modeled is required. The knowledge of these parameters is achieved through methods of adjusting curves between simulated beats and real beats. These traditional optimization methods, such as nonlinear least squares or similar, suffer from the drawback of falling at local minima especially when the initial conditions are not given in an accurate fashion. In the present work, we have designed a novel method robust to deviations in the initial conditions based on Monte Carlo techniques derived from the ideas of the Particle Filtering. Our method allows to adjust the heart beat and to determine the parameters of a model already known in the literature that consists of the sum of five Gaussian curves. The method fits with errors very similar to the traditional method when the initial conditions are good, but better results are obtained in terms of squared error when the initial conditions are sufficiently degraded. Validation was carried out with real physiological and pathological ECG records from international databases.
  • Item
    Model based feature extraction method for myocardial infarction detection
    (Asociación Argentina de Mecánica Computacional, 2018-11) Liberczuk, Sergio Javier ; Bergamini, María Lorena
    The electrical activity of the heart represented by the electrocardiogram (ECG) has been widely used for the detection of heart diseases. Long-term records require the automatic detection of cardiac events. In this work, the detection of myocardial infarction (MI) is performed by means of novel ECG features based on a synthesis ECG model previously described in the literature. The model consists of a sum of five Gaussians centered on each wave of the ECG (P, Q, R, S and T). Each Gaussian is fully specified by three parameters; location in time, amplitude and broadness. By fitting this set of Gaussians, and performing numerical and nonlinear optimization procedures in the resulting 15-dimensional space, we get this set of 15 parameters as features for classification. Although the model was widely used previously with different purposes, its parameters had never been used as features for heartbeat classification even though they reflect the morphology of the ECG in an accurate manner. Physikalisch-Technische-Bundesanstalt (PTB) database was used to validate training and testing algorithms. Data was obtained from 48 healthy subjects (HS) and 95 patients with MI and was split into two datasets. The first dataset contains 190 beats from 26 HS, and 140 beats from 60 patients with MI and was used to train a support vector machine (SVM) classifier with linear kernel. The second dataset contains 88 beats from 22 HS, and 70 beats from 35 subjects with MI and was used to provide a detection performance assessment of the previously trained SVM. This assessment yielded an overall accuracy above 93%. The results show the feasibility of performing the separation between infarcted beats and physiological beats based on the new model-based features proposed. The simplicity of the linear kernel used in the SVM classifier shows the power of the proposed features for classification tasks.