Germán Castellanos

Feature Selection in Pathology Detection using Hybrid Multidimensional Analysis

Heuristical algorithms can reduce the computational complexity. Such methods require of some stopping criteria (cost function). Some of these cost functions are based on statistics like univariate and multivariate methods of analysis. Dimensional reduction techniques such as principal component analysis (PCA) allow to find a lower dimension transformed space based on data variance, but this procedure does not take into account information about classes separability, the direction of maximum variance does not necessarily correspond to the direction of maximum separability. In this work, we propose a feature selection algorithm with heuristic search that uses multivariate analysis of variance (MANOVA) as the cost function. This technique is put to test by classifying hypernasal from normal voices of CLP (cleft lip and/or palate) patients. The classification performance, computational time and reduction ratio are also considered by the comparison with an alternate feature selection method founded on unfolding the multivariate analysis into univariate and bivariate analysis

Parameter tuning associated with nonlinear dynamics techniques for the detection of cardiac murmurs by using genetic algorithms

In this study, nonlinear dynamics techniques toward detecting cardiac murmurs from phonocardiograms (PCG) are used. With this purpose, a methodology for tuning parameters (reconstruction delay -tau and embedding dimension -m) involved in the reconstruction of a meaningful state space from scalar time series is presented, using genetic algorithms (GA), as well as constructing a meta-algorithm combined with support vector regression to adjust the GA parameters in order to decrease the computational cost. The forecasting capacity is used as cost function of the GA. The PCG records belong to the National University of Colombia, 360 beats were chosen by specialist, 180 normal and 180 with cardiac murmur evidence. The obtained results show that by using the tuned GA an efficient procedure for the consistent determination of tau and m is achieved. Murmur detection by using nonlinear features was obtained with classification accuracy of 96% using a k nearest neighbor classifier in cross-validation with 10 folds.

Acoustic Speech Analysis for Hypernasality Detection in Children

Here, an analysis of different acoustic features and their influence in automatic identification of hypernasality is shown. Effective feature selection method includes preprocessing of the initial feature space based on statistical independence analysis. Simultaneously, the synthesis of a specialized diagnostic feature is proposed based on analyzing the acoustic emission of the hyper nasal speech. As a result, It is obtained the acoustic features can differentiate with enough precision the pathology. However, the proposed feature does not require training samples and less computational power, as well

Recognition of cardiac arrhythmias by means of beat clustering on ECG-holter records

The follow-up of some cardiac diseases may be achieved by ECG-holter record analysis. A heartbeat clustering method can be used to reduce the usually high computational cost of such Holter analysis. This study describes a method aimed at cardiac arrhythmia recognition based on this approach, by means of unsupervised inspection of morphologically similar heartbeat groups. Singular Value Decomposition (SVD) is used as the feature selection method since the complexity increases exponentially with the number of features. A modification of the k-means algorithm was developed for centroid computation, taking into account heartbeat length changes. Experimental set consisted of ECG records from the MIT database. The method yielded a 99.9% clustering accuracy considering pathological versus normal heartbeats. Both clustering error and critical error percentage was 0.01%.

Feature extraction of weighted data for implicit variable selection

Approaches based on obtaining relevant information from overwhelmingly large sets of measures have been recently adopted as an alternative to specialized features. In this work, we address the problem of finding a relevant subset of features and a suitable rotation (combined feature selection and feature extraction) as a weighted rotation. We focus our attention on two types of rotations: Weighted Principal Component Analysis and Weighted Regularized Discriminant Analysis. The objective function is the maximization of the J4 ratio. Tests were carried out on artificially generated classes, with several non-relevant features. Real data tests were also performed on segmentation of naildfold capillaroscopic images, and NIST-38 database (prototype selection).