Betul Erdogdu Sakar

Collection and Analysis of a Parkinson Speech Dataset With Multiple Types of Sound Recordings

There has been an increased interest in speech pattern analysis applications of Parkinsonism for building predictive telediagnosis and telemonitoring models. For this purpose, we have collected a wide variety of voice samples, including sustained vowels, words, and sentences compiled from a set of speaking exercises for people with Parkinsons disease. There are two main issues in learning from such a dataset that consists of multiple speech recordings per subject: 1) How predictive these various types, e.g., sustained vowels versus words, of voice samples are in Parkinsons disease (PD) diagnosis? 2) How well the central tendency and dispersion metrics serve as representatives of all sample recordings of a subject? In this paper, investigating our Parkinson dataset using well-known machine learning tools, as reported in the literature, sustained vowels are found to carry more PD-discriminative information. We have also found that rather than using each voice recording of each subject as an independent data sample, representing the samples of a subject with central tendency and dispersion metrics improves generalization of the predictive model.

An Emboli Detection System Based on Dual Tree Complex Wavelet Transform

Automated decision systems for emboli detection is a crucial need since it is being done by visual determination of experts which causes excess time consumption and subjectivity. This work presents an emboli detection system using various dimensionality reduction algorithms on Doppler ultrasound signals recorded from both forward and reverse flow of blood transformed via Fast Fourier Transform (FFT), Discrete Wavelet Transform (DWT), and Dual Tree Complex Wavelet Transform (DTCWT). The combined forward and reverse DTCWT based features produced the highest performance when fed to SVMs classifier. As to compare dimensionality reduction algorithms, although PCA and LDA gave comparable accuracies, LDA has accomplished these accuracies only with two components due to its less than the number of classes’ orthogonal projective directions limitation. SVMs yielded higher classification accuracies than k-NN with all considered dimensionality reduction methods since SVMs classifier is more robust to noise and irrelevant features. With the ability to localize well both in time and frequency, wavelet transform based extracted features gave higher overall classification accuracies than FFT with the more stable classifier SVMs. Additionally, DTCWT accuracies are higher with SVMs than those of DWT since it also has the ability of being shift-invariant.

Determination of the optimal threshold value that can be discriminated by dysphonia measurements for unified Parkinson's Disease rating scale

Recently, there is an increasing motivation to develop telemonitoring systems that enable cost-effective screening of Parkinsons Disease (PD) patients. These systems are generally based on measuring the motor system disorders seen in PD patients by the help of non-invasive data collection tools. Vocal impairments one of the most commonly seen PD symptoms in the early stages of the disease, and building such telemonitoring systems based on detecting the level of vocal impairments results in reliable motor UPDRS tracking systems. In this paper, we aim to determine the optimal UPDRS threshold value that can be discriminated by the vocal features extracted from the sustained vowel phonations of PD patients. For this purpose, we used an online available PD telemonitoring dataset consisting of speech recordings of 42 PD patients. We converted the UPDRS prediction problem into a binary classification problem for various motor UPDRS threshold values, and fed the features to k-Nearest Neighbor and Support Vector Machines classifiers to discriminate the PD patients whose UPDRS is less than or greater than the specified threshold value. The results indicate that speech disorders are more significantly seen in the patients whose UPDRS exceeds the experimentally determined threshold value (15). Besides, considering that the motor UPDRS ranges from 0 to 108, relatively low UPDRS threshold of 15 validates that vocal impairments can be used as early indicators of the disease.

An emboli detection system based on Dual Tree Complex Wavelet Transform and ensemble learning

Embolic signals are used for the identification of active embolic sources in stroke-prone individuals.Dual Tree Complex Wavelet Transform (DTCWT) is used as a new feature extractor from forward and reverse Doppler ultrasound signals.The features acquired from forward and reverse flow directions of the blood are fed into k-NN and SVMs.The individual predictions of classifiers are combined using ensemble stacking method considering that the forward and reverse blood flow coefficients carry different characteristics.The results show that the DTCWT is superior to the DWT and FFT. The traditional visual and acoustic embolic signal detection methods based on the expert analysis of individual spectral recordings and Doppler shift sounds are the gold standards. However, these types of detection methods are high-cost, subjective, and can only be applied by experts. In order to overcome these drawbacks, computer based automated embolic detection systems which employ spectral properties of emboli, speckle, and artifact using Fourier and Wavelet Transforms have been proposed. In this study, we propose a fast, accurate, and robust automated emboli detection system based on the Dual Tree Complex Wavelet Transform (DTCWT). Employing the DTCWT, which does not suffer from the lack of shift invariance property of ordinary Discrete Wavelet Transform (DWT), increases the robustness of the coefficients extracted from the Doppler ultrasound signals. In this study, a Doppler ultrasound dataset including 100 samples from each embolic, Doppler speckle, and artifact signal is used. Each sample obtained from forward and reverse blood flow directions is represented by 1024 points. In our method, we first extract the forward and reverse blood flow coefficients separately using DTCWT from the samples. Then dimensionality reduction is applied to each set of coefficients and both of the reduced set of coefficients are fed to classifiers individually. Subsequently, in the view that the forward and reverse blood flow coefficients carry different characteristics, the individual predictors of these classifiers are combined using ensemble stacking method. We compare the obtained results with Fast Fourier Transform and DWT based emboli detection systems, and show that the features extracted using DTCWT give the highest accuracy and emboli detection rate. It is also observed that combining forward and reverse coefficients using stacking ensemble method improves the emboli and artifact detection rates, and overall accuracy.

Analyzing the effectiveness of vocal features in early telediagnosis of Parkinson's disease

The recently proposed Parkinson’s Disease (PD) telediagnosis systems based on detecting dysphonia achieve very high classification rates in discriminating healthy subjects from PD patients. However, in these studies the data used to construct the classification model contain the speech recordings of both early and late PD patients with different severities of speech impairments resulting in unrealistic results. In a more realistic scenario, an early telediagnosis system is expected to be used in suspicious cases by healthy subjects or early PD patients with mild speech impairment. In this paper, considering the critical importance of early diagnosis in the treatment of the disease, we evaluate the ability of vocal features in early telediagnosis of Parkinsons Disease (PD) using machine learning techniques with a two-step approach. In the first step, using only patient data, we aim to determine the patient group with relatively greater severity of speech impairments using Unified Parkinson’s Disease Rating Scale (UPDRS) score as an index of disease progression. For this purpose, we use three supervised and two unsupervised learning techniques. In the second step, we exclude the samples of this group of patients from the dataset, create a new dataset consisting of the samples of PD patients having less severity of speech impairments and healthy subjects, and use three classifiers with various settings to address this binary classification problem. In this classification problem, the highest accuracy of 96.4% and Matthew’s Correlation Coefficient of 0.77 is obtained using support vector machines with third-degree polynomial kernel showing that vocal features can be used to build a decision support system for early telediagnosis of PD.

A micro emboli vs non-emboli classification system based on the directional dual tree rational dilation wavelet transform

Transcranial Doppler (TCD) is a widely used, non-invasive, rapid and reproducible monitoring method for observing the condition of middle cerebral artery. Micro embolic signals, which appear in various clinical scenarios such as; carotid stenosis, aortic arch plaques, atrial fibrillation, myocardial infarction, patent foramen ovale and valvular stenosis, can be detected by the analysis of TCD signals. Discrete wavelet transform based methods were frequently used in literature for micro embolic signal detection. However, in all the previously used complex/non-complex discrete wavelet transform based methods, low Q-factor wavelets were employed for feature extraction. Low Q-factor wavelets have been successfully used for processing piecewise smooth signals but for the embolic signals, a discrete wavelet transform with better frequency resolution is needed. Therefore in this study, a novel Directional Dual Tree Rational Dilation Wavelet Transform (DDT-RADWT), in which the Q-factor of the analysis and synthesis filters can be adjusted due to the properties of signal of interest, is used as the feature extractor. DDT-RADWT is applied to a dataset consisting of 130 micro embolic signals and 130 non-embolic signals (65 artifacts and 65 Doppler speckles) and the obtained coefficients are used as features. In the proposed method, in order to utilize from the different frequency characteristics of micro embolic, artifact and Doppler speckle signals, the DDT-RADWT is applied with high Q-factor filters. The extracted coefficients are given to k-NN and SVM classifiers with the aim of discriminating two classes of micro embolic signals and non-embolic signals. The results show that higher general accuracy and micro embolic signal detection accuracies are obtained with high Q-factor wavelet analysis.

Variable Importance Analysis in Default Prediction using Machine Learning Techniques.

In this study, different data mining techniques were applied to a finance credit data set from a financial institution to provide an automated and objective profitability measurement. Two-step methodology was used Determining the variables to be included in the model and deciding on the model to classify the potential credit application as “bad credit (default)” or “good credit (not default)”. The phrases “bad credit” and “good credit” are used as class labels since they are used like this in financial sector jargon in Turkey. For this twostep procedure, different variable selection algorithms like Random Forest, Boruta and machine learning algorithms like Logistic Regression, Random Forest, Artificial Neural Network were tried. At the end of the feature selection phase, CRA and III variables were determined as most important variables. Moreover, occupation and product number were also predictor variables. For the classification phase, Neural Network model was the best model with higher accuracy and low average square error also Random Forest model better resulted than Logistic Regression model.

A comparative analysis of speech signal processing algorithms for Parkinson’s disease classification and the use of the tunable Q-factor wavelet transform

Abstract In recent years, there has been increasing interest in the development of telediagnosis and telemonitoring systems for Parkinson’s disease (PD) based on measuring the motor system disorders caused by the disease. As approximately 90% percent of PD patients exhibit some form of vocal disorders in the earlier stages of the disease, the recent PD telediagnosis studies focus on the detection of the vocal impairments from sustained vowel phonations or running speech of the subjects. In these studies, various speech signal processing algorithms have been used to extract clinically useful information for PD assessment, and the calculated features were fed to learning algorithms to construct reliable decision support systems. In this study, we apply, to the best of our knowledge for the first time, the tunable Q-factor wavelet transform (TQWT) to the voice signals of PD patients for feature extraction, which has higher frequency resolution than the classical discrete wavelet transform. We compare the effectiveness of TQWT with the state-of-the-art feature extraction methods used in diagnosis of PD from vocal disorders. For this purpose, we have collected the voice recordings of 252 subjects in the context of this study and extracted multiple feature subsets from the voice recordings. The feature subsets are fed to multiple classifiers and the predictions of the classifiers are combined with ensemble learning approaches. The results show that TQWT performs better or comparable to the state-of-the-art speech signal processing techniques used in PD classification. We also find that Mel-frequency cepstral and the tunable-Q wavelet coefficients, which give the highest accuracies, contain complementary information in PD classification problem resulting in an improved system when combined using a filter feature selection technique.

Prediction of level and abrupt changes of ozon concentration

While, in stratosphere, high level ozone concentration protects the Earth against ultraviolet radiation, in lower troposphere it has negative effects on human health and environment. The goal of this study is to determine the feature groups that are related to abrupt changes in the level of ozone. Linear discriminant analysis and support vector machines methods are used to explore which combination of features are predictive of abrupt changes in ozone level on the simulation dataset collected in Ankara, Turkey, by an automatic air quality monitoring station operated by the ministry of environment and urban planning. The dataset consists of one year of measurements of air pollutants and the meteorological factors. The obtained results showed that particulate matters, nitric oxides and temperature are most effective parameters in the classification of absurt rise and fall in the level of ozone.

A validation method for comparing classifiers on imbalanced datasets

In this study, to compare the robustness and learning capability of the classifiers on imbalanced datasets, a cross validation method that generates class-imbalanced training sets is proposed. The method will also be used to evaluate the accuracies of methods developed for dealing with the class-imbalance problem. The proposed method is used to generate imbalanced datasets from three biomedical datasets. Then, k-Nearest Neighbor, Support Vector Machines and Multi Layer Perceptron classifiers are compared using various settings of their hyper-parameters that affect their complexities. The experimental results show that SVMs are simply the most robust of all when applied to imbalanced datasets.