Feature Extraction From Parametric Time–Frequency Representations for Heart Murmur Detection

Abstract

The detection of murmurs from phonocardiographic recordings is an interesting problem that has been addressed before using a wide variety of techniques. In this context, this article explores the capabilities of an enhanced time-frequency representation (TFR) based on a time-varying autoregressive model. The parametric technique is used to compute the TFR of the signal, which serves as a complete characterization of the process. Parametric TFRs contain a large quantity of data, including redundant and irrelevant information. In order to extract the most relevant features from TFRs, two specific approaches for dimensionality reduction are presented: feature extraction by linear decomposition, and tiling partition of the t-f plane. In the first approach, the feature extraction was carried out by means of eigenplane-based PCA and PLS techniques. Likewise, a regular partition and a refined Quadtree partition of the t-f plane were tested for the tiled-TFR approach. As a result, the feature extraction methodology presented, which searches for the most relevant information immersed on the TFR, has demonstrated to be very effective. The features extracted were used to feed a simple k-nn classifier. The experiments were carried out using 45 phonocardiographic recordings (26 normal and 19 records with murmurs), segmented to extract 548 representative individual beats. The results using these methods point out that better accuracy and flexibility can be accomplished to represent non-stationary PCG signals, showing evidences of improvement with respect to other approaches found in the literature. The best accuracy obtained was 99.06 +/- 0.06%, evidencing high performance and stability. Because of its effectiveness and simplicity of implementation, the proposed methodology can be used as a simple diagnostic tool for primary health-care purposes.

Time-Frequency Based Feature Extraction for Non-Stationary Signal Classification

“…Further discussion on the properties and advantages/disadvantages of each TFR estimation approach is out of the scope of this study. For more information on this subject, the reader can refer to the following papers: linear, quadratic and parametric TFR (Marchant, 2003); parametric t-f analysis (M Tarvainen & Karjalainen, 2004;Poulimenos & Fassois, 2006); comparison of different TFR on EEG classification (Tzallas et al, 2008); parametric TFR based classification (Avendano-Valencia et al, 2010). Time-frequency Dynamic Features -TFDF are a set of variables describing the temporal dependency of some spectral-related quantity.…”

“…However, for classification purposes, the obtained components are not always related to the most discriminative information, as their transformation is based only on feature variability while class labels are neglected. Supervised methods, like PLS, can be used to improve performance of linear transform methods taking into account label variability as well as feature variability (Avendano-Valencia et al, 2010). Nonetheless, computing transformation matrices of linear transform methods becomes more expensive as the dataset dimension increases.…”

“…When PCA transformations are used, these methods are known as two dimensional PCA (2D-PCA) (Yang et al, 2004). Two dimensional methods can be extended for different kinds of linear transforms, as in the case of PLS Avendano-Valencia et al, 2010), with improved classification performance. Nonetheless, these methods do not impose a priori a dynamic structure for the t-f feature ordering.…”

“…Subsequently, based on previous work in ( Avendano-Valencia et al, 2010), a new methodology is proposed, oriented to reducing the dimensionality of t-f based features. The proposed methodology consecutively performs both the selection of the most relevant features as well as a linear transform of the t-f planes.…”

Identification of Abnormal Heart Sounds

“…Among these studies, artificial neural networks [13], support vector machines [14] and HMM based [15] approaches are common. Classification based on clustering has also been shown to be effective in heart sound pathology classification [16]. Typical features used in these studies comprise time-domain, frequency-domain, time-frequency and complexity-based features [13] …”