Search

The Electroencephalography (EEG)-based motor imagery (MI) paradigm is one of the most studied technologies for Brain-Computer Interface (BCI) development. Still, the low Signal-to-Noise Ratio (SNR) poses a challenge when constructing EEG-based BCI systems. Moreover, the non-stationary and nonlinear signal issues, the low-spatial data resolution, and the inter- and intra-subject variability hamper the extraction of discriminant features. Indeed, subjects with poor motor skills have difficulties in practicing MI tasks against low SNR scenarios. Here, we propose a subject-dependent preprocessing approach that includes the well-known Surface Laplacian Filtering and Independent Component Analysis algorithms to remove signal artifacts based on the MI performance. In addition, power- and phase-based functional connectivity measures are studied to extract relevant and interpretable patterns and identify subjects of inefficency. As a result, our proposal, Subject-dependent Artifact Removal (SD-AR), improves the MI classification performance in subjects with poor motor skills. Consequently, electrooculography and volume-conduction EEG artifacts are mitigated within a functional connectivity feature-extraction strategy, which favors the classification performance of a straightforward linear classifier.

Este artículo presenta un método de mapeo acústico-a-articulatorio que permite hacer seguimiento del movimiento de articuladores críticos en voces del tipo fricativo. El método propuesto hace uso de aquellas características de tiempo-frecuencia mejor relacionadas desde un punto de vista estadístico con el movimiento de los articuladores, en donde, para el proceso de estimación de características, se aplica normalización por longitud del tracto vocal. El método se prueba sobre la base de datos MOCHA-TIMIT, la cual contiene señales provenientes de un articulógrafo electromagnético. El conjunto de características propuesto se prueba sobre un sistema de mapeo acústico-a-articulatorio basado en modelos de mezclas Gaussianas, en donde la correlación de Pearson se utiliza para medir la bondad de las estimaciones. Se hacen pruebas de una manera independiente del hablante, es decir, los conjuntos de entrenamiento y prueba pertenecen a diferentes hablantes. Como resultado se obtiene que el valor de correlación entre las estimaciones, hechas de manera independiente del hablante, es comparable a la medida de desempeño obtenida desde un punto de vista dependiente del hablante.

Este artículo presenta un nuevo método para la estimación de actividad neuronal a partir de señales electroencefalográficas usando análisis de series de tiempo ponderadas. El método considera un modelo lineal basado en restricciones fisiológicas que tiene en cuenta tanto la dinámica espacial como la temporal, y una etapa de ponderación que modifica las suposiciones del modelo a partir de las observaciones. La matriz de pesos calculada es incluida en la función de costo usada para solucionar el problema inverso dinámico, y por lo tanto en la formulación del filtro de Kalman. De esta forma, se propone un filtro de Kalman ponderado que incluye la matriz de pesos. El desempeño del filtro (en términos del error de localización) se analiza para varios SNRs. El desempeño óptimo se alcanza usando el modelo lineal con matriz de ponderación calculado por el método de producto interno.

En el presente trabajo se propone una metodología para la reducción de dimensión en representaciones tiempo frecuencia (TFRs) enfocada a la clasificación de bioseñales no estacionarias, que trata directamente su cantidad de datos irrelevantes y redundantes, combinando una etapa de selección de características con una etapa de reducción de dimensión por medio de métodos de descomposición lineal extendidos a datos bidimensionales. La metodología se prueba sobre un conjunto de TFRs paramétricas calculadas sobre una base de datos de señales fonocardiográficas (FCG) para la detección de soplos cardiacos. Los resultados muestran una mejora comparados con otras metodologías que no tienen en cuenta la presencia de datos irrelevantes y redundantes en las representaciones, además, el uso de las metodologías de descomposición lineal bidimensionales reducen adecuadamente la redundancia de las TFRs, obteniendo un nuevo conjunto de características 2D de menor dimensión que el conjunto inicial.

In this paper, a comparison of methods for nonlinear dimensionality reduction is proposed in order to determine which technique preserves better the local properties, without losing the overall structure of the original data. We seek to establish which of these methods is the most appropriate for visualization tasks. The embeddings obtained with each technique are evaluated by two criteria Preservation Neighborhood Error and Preserved Neighbors Average. The methodologies were tested on artificial and real-world data sets which allow us to visually confirm the quality of the embedding. The results obtained show that Maximum variance unfolding computes high quality embeddings, because the optimization problem pretends to preserve exactly the local pair-wise distance between neighbors and conserve the global manifold structure.

In this paper is presented a method for neural activity estimation over the brain that take into account, for the solution of the inverse problem, a dynamic model for the neural activity in a realistic head model calculated with bounded elements method, according to a physiologically based model that describes the real interaction between neurons. The solution of the inverse problem is calculated using high performance computing. This analysis is performed for simulated EEG signals for SNR of 25 dB, 15 dB and 5 dB. The obtained results show the robustness of the estimation method that includes the dynamic model in comparison with the static model for several levels of noise.

A very desirable attribute in automatic pathology detection systems is noise robustness, such that the presence of noise should not significantly affect the ability to detect pathologies. This paper explores the discriminatory potential that nonlinear dynamic features, in particular, the Hurst exponent, the maximum Lyapunov exponent and correlation dimension, can provide in the detection of heart murmurs using phonocardiographic signals contaminated with different levels of noise. At the same time comparing with features obtained from time frequency representations. The results show the strength of features based on nonlinear dynamics for functional state classification tasks, even on signals with high noise levels.

People with a defective velopharyngeal mechanism speak with abnormal nasal resonance (hypernasal speech). Voice analysis methods for hypernasality detection commonly use vowels and nasalized vowels. However to obtain a more general assessment of this abnormality it is necessary to analyze stops and fricatives. This study describes a method with high generalization capability for hypernasality detection analyzing unvoiced Spanish stop consonants. The importance of phoneme-by-phoneme analysis is shown, in contrast with whole word parametrization which includes irrelevant segments from the classification point of view. Parameters that correlate the imprints of Velopharyngeal Incompetence (VPI) over voiceless stop consonants were used in the feature estimation stage. Classification was carried out using a Support Vector Machine (SVM), including the Rademacher complexity model with the aim of increasing the generalization capability. Performances of 95.2% and 92.7% were obtained in the processing and verification stages for a repeated cross-validation classifier evaluation.

In this work, a comparative study of feature selection methods for supervised and unsupervised inference obtained from classical PCA is presented. We deduce an expression for the cost function of PCA based on the mean square error of data and its orthonormal projection, and then this concept is extended to obtain an expression for general WPCA. Additionally, we study the supervised and unsupervised Q – α algorithm and its relation with PCA. At the end, we present results employing two data sets: A low-dimensional data set to analyze the effects of orthonormal rotation, and a highdimensional data set to assess the classification performance. The feature selection methods were assessed taking into account the number of relevant features, computational cost and classification performance. The classification was carried out using a partitional clustering algorithm.

Usually, in pattern recognition problems we represent the observations by mean of measures on appropriate variables of data set, these measures can be categorized as Static and Dynamic Features. Static features are not always an accurate representation of data. In these sense, many phenomena are better modeled by dynamic changes on their measures. The advantage of using an extended form (dynamic features) is the inclusion of new information that allows us to get a better representation of the object. Nevertheless, sometimes it is difficult in a classification stage to deal with dynamic features, because the associated computational cost often can be higher than we deal with static features. For analyzing such representations, we use Principal Component Analysis (PCA), arranging dynamic data in such a way we can consider variations related to the intrinsic dynamic of observations. Therefore, the method made possible to evaluate the dynamic information about of the observations on a lower dimensionality feature space without decreasing the accuracy performance. Algorithms were tested on real data to classify pathological speech from normal voices, and using PCA for dynamic feature selection, as well.

El análisis multiescala es un área de gran actividad investigativa con fuerte impacto en computación científica y matemática aplicada, ocupando un lugar de privilegio en la forma como se entiende la relación entre la matemática y las demás ciencias. Aunque el estudio matemático de problemas multiescala está bastante documentado, estas técnicas heredan en el entorno discreto la naturaleza mal condicionada del problema. La molificación discreta ha sido empleada con éxito en la solución numérica de diversos problemas mal condicionados. Este árticulo propone una nueva técnica de análisis multiescala, basada en molificación discreta. El procedimiento aquí propuesto usa resultados de algebra lineal numérica para implementar molificación discreta en el algoritmo de Mallat. La nueva técnica tiene una teoría simple, una implementación eficiente y proporciona resultados de calidad comparable a técnicas multiescala clásicas tipo onditas (wavelet). El proceso es aplicado en señales electrocardiográficas contaminadas con ruido no blanco, para fines de comparación.

In order to characterize physiological signals, which may have highly nonlinear structures, it’s common to use methodologies derived from fractal techniques that make part of complexity analysis. This work proposes is proposed an evaluation function based on measuring the capacity of prediction of a neural network trained with Kalman filter to predict points in a reconstructed state space attractor, so measuring the quality of the attractor from a onedimensional signal. We propose use of statistic measures such as Kullback –Leibler, Kolmogorov-Smirnov and Hellinger to determine difference between the embedded statistic structure in the predicted points and the original signal points. Results were obtained on attractor reconstruction from ECG signals of MIT-BIH database and EEG signals obtained from Clinic for Epileptologie Epileptologie Bonn University database. In this way, it was possible to evaluate the prediction capacity corresponding to reconstruct attractors from records, from which we concluded that an attractor with high capacity of time series prediction implies good embedding properties in state space.