Biomedical Signal Processing / Medical Signal Processing / Biosignal Processing
Parastoo Sadeghinia; Hamed Danandeh Hesar
Volume 16, Issue 3 , December 2022, , Pages 271-287
Abstract
Phonocardiography (PCG) signals provide valuable information about the heart valves .These auditory signals can be useful in the early diagnosis of heart diseases. Automatic heart sound classification has a promising potential in the field of heart pathology. In this research, a new method based on machine ...
Read More
Phonocardiography (PCG) signals provide valuable information about the heart valves .These auditory signals can be useful in the early diagnosis of heart diseases. Automatic heart sound classification has a promising potential in the field of heart pathology. In this research, a new method based on machine learning techniques is proposed for discriminating normal and abnormal heart sounds. In this method, first, the heart sounds are segmented into 4 main parts: S1, S2, systole and diastole segments. From these segments, statistical and time-frequency features are extracted for classification. Before classification, the distinctive features are selected using two approaches. In the first approach, the feature selection is accomplished using particle swarm optimization algorithm (PSO). In the second approach, we use Sequential Forward Feature Selection (SFFS) method. The proposed method was evaluated on the Physionet 2016 Challenge database using 10-fold cross-validation method. In this database, the number of normal and abnormal PCG signals are not balanced. Therefore, in this paper, the synthetic minority over-sampling technique (SMOTE) is applied to produce balanced data. The evaluation results showed that the proposed method can distinguish the normal heart sounds from abnormal ones with accuracy of 98/03% and sensitivity and specificity of 97.64%, 98.43%respectively.
Biomedical Signal Processing / Medical Signal Processing / Biosignal Processing
Masoud Moradi; Sina Shamekhi
Volume 16, Issue 2 , September 2022, , Pages 167-182
Abstract
In recent years, the fabrication of devices that can facilitate the difficulty of communication between deaf people and the general public and translate sign language has attracted interest from researchers. But problems such as low accuracy and calculation speed and the high cost of tools have hindered ...
Read More
In recent years, the fabrication of devices that can facilitate the difficulty of communication between deaf people and the general public and translate sign language has attracted interest from researchers. But problems such as low accuracy and calculation speed and the high cost of tools have hindered the commercialization of research. Another challenge in making a practical tool is the necessity of good performance of the methods in the perspective of training by leave-one-subject-out or in other words classifying the data of a new person. Therefore, in this article, an efficient method for detecting hand gestures with the purpose of sign language translation has been presented, so that while using a method with lower dimensions, better performance can be obtained in all kinds of training methods. In the proposed method, the features consisting of the mean absolute value, variance, root mean square, waveform length, kurtosis, and skewness have been extracted from the empirical wavelet transformation of the electromyogram and inertial signals. Then, by the ReliefF method, effective features have been selected and for the classification of hand gestures, a support vector machine classifier has been used. The accuracy percentages of the proposed method on the PSL database and DB2, DB3, DB5, and DB7 datasets of the NinaPro database, have been respectively obtained as follows: 99.31%, 97.11%, 96.58%, 96.12%, and 97.32% in the word-subject training approach, 99.78%, 97.22%, 95.46%, 97.23%, and 97.72% in the word-all-subject training approach, and 97.43%, 94.68%, 89.66%, 91.55%, and 94.81% in the leave-one-subject-out method.
Abolfazl Tabatabaei; Vali Derhami; Razieh Sheikhpour; Mohammad-Reza Pajoohan
Volume 13, Issue 4 , December 2019, , Pages 337-348
Abstract
Feature selection is a well-known preprocessing technique in machine learning, data mining and especially bioinformatics microarray analysis with a high-dimension, low-sample-size (HDLSS) data. The diagnosis of genes responsible for disease using microarray data is an important issue to promoting knowledge ...
Read More
Feature selection is a well-known preprocessing technique in machine learning, data mining and especially bioinformatics microarray analysis with a high-dimension, low-sample-size (HDLSS) data. The diagnosis of genes responsible for disease using microarray data is an important issue to promoting knowledge about the mechanism of disease and improves the way of dealing with the disease. In feature selection methods based on information theory, which cover a wide range of feature selection methods, the concept of entropy is used to define criteria for relevance, redundancy and complementarity. In this paper, we propose a new relevancy criterion based on the concept of pure continuity rather than the concept of entropy. In the proposed method, to control and reduce redundancy, the relevancy between a feature and each class is separately examined, while in most of the filter methods the value of a feature is measured based on its relation to the entire class. This solution allows us to identify the most efficient features (genes) of each class separately, while identifying common features (genes) is also possible. Discretization is another challenge in some available techniques. Using a homomorphism transformation in proposed method avoids engaging with discretization complexities, while taking advantages of it. Seven types of cancer microarrays with three types of classification models (e.g. NB, KNN and SVM) are used to establish a comparison between the proposed method and other relevant methods. The results confirm the efficiency of the proposed method in the term of accuracy and number of selected genes as two parameters of classification.
Mahla Dehtaghi Zadeh; Farid Saberi-Movahed; Mahdi Eftekhari
Volume 13, Issue 3 , October 2019, , Pages 223-234
Abstract
DNA micro-array datasets play crucial role in machine learning and recognition of various kinds of cancer structures. Micro-array datasets are typically characterized by the high number of features and the small number of samples. Such problems may result in overfitting and low prediction accuracy of ...
Read More
DNA micro-array datasets play crucial role in machine learning and recognition of various kinds of cancer structures. Micro-array datasets are typically characterized by the high number of features and the small number of samples. Such problems may result in overfitting and low prediction accuracy of classifiers due to the irrelevant features, and therefore, they are considered as a challenging task in machine learning. The direct way to deal with such challenges is dimensionality reduction of data. In this regard, feature selection method acts as an effective solution for dimensinality reduction and increasing efficiency of learning algorithms. In this paper, by using the concept of “the basis for the DNA micro-array datasets”, a new feature selection method is introduced. To be more specific, rather than utilizing the entire micro-array dataset for tackling the problem of feature selection, a basis that is a muchmore smaller subset of the micro-array dataset is used. This method is based on subspace learning and matrix factorization. Finally, by making use of the DNA micro-array datasets, the effectiveness of the proposed method is evaluated, and the obtained results are compared with some state-of-the-art supervised feature selection methods.
Bioinformatics / Biomedical Informatics / Medical Informatics / Health Informatics
Hossein Bankikoshki; Seyed Ali Seyyedsalehi; Fatemeh Zare Mirakabad
Volume 11, Issue 3 , September 2017, , Pages 219-230
Abstract
The use of genomic nucleotide sequences as biochemical signals in machine learning methods is possible by converting these sequences into numerical codes. This conversion results in an unrealistic increase in the dimension of the data and encounters some data analysis operations such as visualization ...
Read More
The use of genomic nucleotide sequences as biochemical signals in machine learning methods is possible by converting these sequences into numerical codes. This conversion results in an unrealistic increase in the dimension of the data and encounters some data analysis operations such as visualization and feature extraction with constraints. Therefore, one should use the dimensionality reduction technics in order to return the data to its real dimension. In this study, a deep autoencoder neural network has been used to reduce the dimension of binding site sequence data on the human genome. In order to determine whether the information of real data is preserved in compressed data, we perform a two-class classification using a support vector machine. The results show that information is almost entirely preserved in compression. Then, compressed data is used for visualization as well as feature selection by analysis of variance. The results show that the first, the tenth and eighth positions in the sequences are the most informative positions. While the majority of the previous works deal with gene expression data of microarrays and compare a few dimension reduction algorithms, this paper for the first time uses an autoencoder on nucleotide sequence data and provides a comprehensive comparison between the performance of the dimension reduction technics and machine learning algorithms.
Biomedical Image Processing / Medical Image Processing
Amir Ehsan Lashkari; Fatemeh Pak; Mohammad Firouzmand
Volume 9, Issue 1 , April 2015, , Pages 71-84
Abstract
Breast cancer is the most common type of cancer among women. The important key to treat the breast cancer is early detection of it because according to many pathological studies more 80% of all abnormalities are still benign at primary stages; so in recent years, many studies and extensive research done ...
Read More
Breast cancer is the most common type of cancer among women. The important key to treat the breast cancer is early detection of it because according to many pathological studies more 80% of all abnormalities are still benign at primary stages; so in recent years, many studies and extensive research done to early detection of breast cancer with higher precision and accuracy. Infra-red breast thermography is an imaging technique based on recording temperature distribution patterns of breast tissue. Compared with breast mammography technique, thermography is more suitable technique because it is noninvasive, non-contact, passive and free ionizing radiation. In this paper, a full automatic high accuracy technique for classification of suspicious areas in thermogram images with the aim of assisting physicians in early detection of breast cancer has been presented. Proposed algorithm consists of four main steps: pre-processing & segmentation, feature extraction, feature selection and classification. At the first step, using full automatic operation, region of interest (ROI) determined and the quality of image improved. Using thresholding and edge detection techniques, both right and left breasts separated from each other. Then relative suspected areas become segmented and image matrix normalized due to the uniqueness of each person's body temperature. At feature extraction stage, 23 features, including statistical, morphological, frequency domain, histogram and Gray Level Co-occurrence Matrix (GLCM) based features are extracted from segmented right and left breast obtained from step 1. To achieve the best features, feature selection methods such as mRMR, SFS, SBS, SFFS, SFBS and GA have been used at step 3. Finally to classify and TH labeling procedures, different classifiers such as AdaBoost, SVM, kNN, NB and PNN are assessed to find the best suitable one. The results obtained on native database showed the best and significant performance of the proposed algorithm in comprise to the similar studies. According to experimental results, mRMR combined with AdaBoost with the maximum accuracy of 92%, and SFFS combined with AdaBoost with a maximum accuracy of 88%, are the best combination of feature selection and classifier for evaluation of the right and left breast images respectively.
Biomedical Signal Processing / Medical Signal Processing / Biosignal Processing
Mahdi Khezri; Seyed Mohammad Firoozabadi; Seyed Ahmad Reza Sharafat
Volume 8, Issue 4 , February 2015, , Pages 339-358
Abstract
In this study, we propose decision level fusion of multimodal physiological signals to design an affect identification system using the MIT database. Four types of physiological signals, including blood volume pressure (BVP), respiration rate (RSP), skin conductance and facial muscles activities (fEMG) ...
Read More
In this study, we propose decision level fusion of multimodal physiological signals to design an affect identification system using the MIT database. Four types of physiological signals, including blood volume pressure (BVP), respiration rate (RSP), skin conductance and facial muscles activities (fEMG) were utilized as affective modalities. To collect the above-mentioned database, researchers used personalized imagery to elicit the desired affective states from a single subject and recorded the corresponding physiological signals simultaneously. In this study, the best subset of features for each signal was determined using previously calculated time and frequency domain features. To this end, sequential floating forward selection (SFFS) and RELIEF feature selection algorithms were evaluated. A new feature set, formed by concatenating the selected features, was partitioned into three subsets. Each subset was then fed into a classifier to identify the desired affective states. The majority voting method was applied to fuse the results obtained by the subsystems. Three types of classification methods, namely SVM, LDA and KNN were evaluated to design an affect identification system. The results showed remarkable performance from the system in identifying the desired scenarios with an acceptable accuracy and speed of response. Using the RELIEF feature selection method, along with SVM as a classifier, an overall recognition accuracy of 93.8% was obtained, which is better than the results reported with the use of the above-mentioned database so far.
Biological Computer Modeling / Biological Computer Simulation
Mohammad Jazlaeiyan; Hadi Shahriar Shahhoseini
Volume 8, Issue 4 , February 2015, , Pages 371-383
Abstract
Human visual system operates superior than best machine vision systems in object recognition. So, researchers in machine vision and neuroscience try to model human visual system in order to employ it in machine. HMAX is one of the best operating models in this area. It is based on the function of brain ...
Read More
Human visual system operates superior than best machine vision systems in object recognition. So, researchers in machine vision and neuroscience try to model human visual system in order to employ it in machine. HMAX is one of the best operating models in this area. It is based on the function of brain cells in the ventral stream of visual cortex and contains four computational layers. In the learning stage, many image partitions called image patches are extracted randomly with different sizes from training images. This random selection of image patches is one of the drawbacks of HMAX which decreases the performance and increases the computational complexity of the algorithm. In this paper, a novel patch selection from the set of random patches is proposed. In this method, using a recursive approach, optimal patches are selected from optimal features of training images by mutual information maximization feature selection. The performance of proposed algorithm in binary classification (existence or non-existence of objects in the images) is compared with HMAX and the superiority is proved.
Biomedical Signal Processing / Medical Signal Processing / Biosignal Processing
Amin Janghorbani; Mohammad Hasan Moradi; Abdollah Arasteh
Volume 7, Issue 2 , June 2013, , Pages 163-174
Abstract
Acute hypotension episodes (AHEs) are one of the hemodynamic instabilities with high mortality rate that is frequent among many groups of patients. Prognosis of acute hypotension episodes can help clinicians to diagnose the cause of this physiological disorder and select proper treatment based on this ...
Read More
Acute hypotension episodes (AHEs) are one of the hemodynamic instabilities with high mortality rate that is frequent among many groups of patients. Prognosis of acute hypotension episodes can help clinicians to diagnose the cause of this physiological disorder and select proper treatment based on this diagnosis. In this study two groups of features, physiological and chaotic features, were extracted from the physiological time series to be applied for prediction of AHEs in the future 1 hour time interval. The best set of the features from the extracted features were selected using Genetic Algorithm (GA) and were classified by SVM. The prediction accuracy for physiological features was 87.5% and for chaotic features was 85%. In order to improve prediction accuracy, physiological and chaotic features were employed simultaneously in feature selection and the best combination of these features was selected by GA and classified by SVM. The best prognosis accuracy, which was achieved in this study by classification of the selected features, was 95% that was better than other previously studies on the same database.