Yasemin P. Kahya

Comparison of AR-based algorithms for respiratory sounds classification

Respiratory sounds of pathological and healthy subjects were analyzed via autoregressive (AR) models with a view to construct a diagnostic aid based on auscultation. Using the AR vectors, two reference libraries, pathological and healthy, were built. Two classifiers, k-nearest neighbour (k-NN) classifier and a quadratic classifier, were designed and compared. Performances of the classifiers were tested for different model orders. The best classification results were obtained for model order 6.

A Multi-Channel Device for Respiratory Sound Data Acquisition and Transient Detection

In this study, a multi-channel analog data acquisition and processing device with the additional feature of detecting adventitious sounds has been designed and implemented. The overall system consists of fourteen microphones attached on the backside, an airflow measuring unit, a fifteen-channel amplifier and filter unit connected to a personal computer (PC) via a data acquisition (DAQ) card, and an interface and adventitious sound detection program prepared using LabVIEW (6.0, National Instruments) and MATLAB (7.0.1, MathWorks). The system records the fourteen-channel respiratory sound data at the posterior chest wall and in addition measures the air flow to synchronize the pulmonary signal on the respiration cycle. Respiratory data are amplified and band-pass filtered, whereas flow signal is only low-pass filtered since it is a low-frequency signal with sufficiently high amplitude. All data are sent to a PC to be digitized by DAQ card, then to be processed and stored. An algorithm based on wavelet decomposition is developed which detects the adventitious pulmonary sounds, mainly the crackles and wheezes. This system is intended to be used for mapping the pulmonary sounds and detecting and locating the adventitious pulmonary sounds

Multiresolution biological transient extraction applied to respiratory crackles

A method is proposed for the detection of transients in biological signals. The method is based on enhancing the transient-to-background ratio by a series of operations such as background whitening, wavelet-based multiresolution decomposition and application of Teagers energy operator. The transients are extracted by judiciously thresholding this processed signal. The proposed detector is applied to the discrimination of crackles in pathological respiratory sounds. It is shown that both the crackle detection performance and ability to extract the transient waveforms correctly are superior to existing detectors in the literature.

Pulmonary crackle detection using time-frequency and time-scale analysis

Pulmonary crackles are used as indicators for the diagnosis of different pulmonary disorders in auscultation. Crackles are very common adventitious transient sounds. From the characteristics of crackles such as timing and number of occurrences, the type and the severity of the pulmonary diseases may be assessed. In this study, a method is proposed for crackle detection. In this method, various feature sets are extracted using time-frequency and time-scale analysis from pulmonary signals. In order to understand the effect of using different window and wavelet types in time-frequency and time-scale analysis in detecting crackles, different windows and wavelets are tested such as Gaussian, Blackman, Hanning, Hamming, Bartlett, Triangular and Rectangular windows for time-frequency analysis and Morlet, Mexican Hat and Paul wavelets for time-scale analysis. The extracted feature sets, both individually and as an ensemble of networks, are fed into three different machine learning algorithms: Support Vector Machines, k-Nearest Neighbor and Multilayer Perceptron. Moreover, in order to improve the success of the model, prior to the time-frequency/scale analysis, frequency bands containing no-crackle information are removed using dual-tree complex wavelet transform, which is a shift invariant transform with limited redundancy compared to the conventional discrete wavelet transform. The comparative results of individual feature sets and ensemble of sets, which are extracted using different window and wavelet types, for both pre-processed and non-pre-processed data with different machine learning algorithms, are extensively evaluated and compared.

Classification of respiratory signals by linear analysis

The aim of this study is the classification of wheeze and non-wheeze epochs within respiratory sound signals acquired from patients with asthma and COPD. Since a wheeze signal, having a sinusoidal waveform, has a different behavior in time and frequency domains from that of a non-wheeze signal, the features selected for classification are kurtosis, Renyi entropy, f50/ f90 ratio and mean-crossing irregularity. Upon calculation of these features for each wheeze and non-wheeze portion, the whole data scattered as two classes in four dimensional feature space is projected using Fisher Discriminant Analysis (FDA) onto the single dimensional space that separates the two classes best. Observing that the two classes are visually well separated in this new space, Neyman-Pearson hypothesis testing is applied. Finally, the correct classification rate is %95.1 for the training set, and leave-one-out approach pursuing the above methodology yields a success rate of %93.5 for the test set.

Multilevel Statistical Inference From Functional Near-Infrared Spectroscopy Data During Stroop Interference

Functional near-infrared spectroscopy (fNIRS) is an emerging technique for monitoring the concentration changes of oxy- and deoxy-hemoglobin (oxy-Hb and deoxy-Hb) in the brain. An important consideration in fNIRS-based neuroimaging modality is to conduct group-level analysis from a set of time series measured from a group of subjects. We investigate the feasibility of multilevel statistical inference for fNIRS. As a case study, we search for hemodynamic activations in the prefrontal cortex during Stroop interference. Hierarchical general linear model (GLM) is used for making this multilevel analysis. Activation patterns both at the subject and group level are investigated on a comparative basis using various classical and Bayesian inference methods. All methods showed consistent left lateral prefrontal cortex activation for oxy-Hb during interference condition, while the effects were much less pronounced for deoxy-Hb. Our analysis showed that mixed effects or Bayesian models are more convenient for faithful analysis of fNIRS data. We arrived at two important conclusions. First, fNIRS has the capability to identify activations at the group level, and second, the mixed effects or Bayesian model is the appropriate mechanism to pass from subject to group-level inference.

A wavelet-based instrument for detection of crackles in pulmonary sounds

Crackles are discontinuous adventitious respiratory sounds, which are considered as signs of various pulmonary disorders, therefore their detection is important in the analysis of lung sounds. In this work, an instrument for separating crackles from stationary lung sounds and quantifying their characteristics is realized with DSP. The detection algorithm is based on increasing transient to background ratio by adaptive filtering and implementing nonlinear operators to wavelet based decomposed lung sounds.

Classifying Respiratory Sounds with Different Feature Sets

In this study, different feature sets are used in conjunction with (k-nearest neighbors) k-NN and artificial neural network (ANN) classifiers to address the classification problem of respiratory sound signals. A comparison is made between the performances of k-NN and ANN classifiers with different feature sets derived from respiratory sound data acquired from one microphone placed on the posterior chest area. Each subject is represented by a single respiration cycle divided into sixty segments from which three different feature sets consisting of 6 th order AR model coefficients, wavelet coefficients and crackle parameters in addition to AR model coefficients are extracted. Classification experiments are carried out on inspiration and expiration phases separately. The two class recognition problem between healthy and pathological subjects is addressed

Feature extraction for pulmonary crackle representation via wavelet networks

In this study, wavelet networks have been used to parameterize and quantify pulmonary crackles with an aim to depict the waveform with a small set of meaningful parameters. Complex Morlet wavelets are used at the nodes of both single and double-node networks to model the waveforms with the double-node rendering smaller modeling error. The features extracted from the model parameters have been compared with the conventional time domain features in a two-class clustering experiment with nearly 90% matching between the clusters of different parameter sets and with the model parameters forming clusters more closely distributed around their means and better separated from each other. Moreover, using simulated crackles embedded on real respiratory sounds, features extracted from wavelet networks have been shown to be more robust to background vesicular sounds compared to conventional parameters which are very sensitive to noise.

Using lung sounds in classification of pulmonary diseases according to respiratory subphases

Auscultation-based diagnosis of pulmonary disorders relies heavily on the presence of adventitious sounds and on the altered transmission characteristics of the chest wall. The phase information of the respiratory cycle within which adventitious sounds occur is very helpful in diagnosing different diseases. In this study, respiratory sound data belonging to four pulmonary diseases, both restrictive and obstructive, along with healthy respiratory data are used in various classification experiments. The sound data are separated into six subphases, namely, early, mid, late inspiration and expiration and classification experiments using a neural classifier are carried out for each subphase. The AR parameters acquired from segmented sound signals, prediction error and the ratio of expiration to inspiration durations are used to construct the feature set to the neural classifier. Classification experiments are carried out between healthy and pathological sound segments, between restrictive and obstructive sound segments and between two different disease sound segments. The results indicate that the classifier performance demonstrates subphase dependence for different diseases. These results may shed light in eliminating redundant feature spaces in building an expert system using lung sounds for pulmonary diagnosis.