Fisher Linear Classifier Research Articles

Preictal state detection using prodromal symptoms: A machine learning approach.

A reliable identification of a high-risk state for upcoming seizures may allow for preemptive treatment and improve the quality of patients' lives. We evaluated the ability of prodromal symptoms to predict preictal states using a machine learning (ML) approach. Twenty-four patients with drug-resistant epilepsy were admitted for continuous video-electroencephalographic monitoring and filled out a daily four-point questionnaire on prodromal symptoms. Data were then classified into (1) a preictal group for questionnaires completed in a 24-h period prior to at least one seizure (n1 =58) and (2) an interictal group for questionnaires completed in a 24-h period without seizures (n2 =190). Our prediction model was based on a support vector machine classifier and compared to a Fisher's linear classifier. The combination of all the prodromal symptoms yielded a good prediction performance (area under the curve [AUC]=.72, 95% confidence interval [CI] = .61-.81). This performance was significantly enhanced by selecting a subset of the most relevant symptoms (AUC=.80, 95% CI = .69-.88). In comparison, the linear classifier systematically failed (AUCs<.6). Our findings indicate that the ML analysis of prodromal symptoms is a promising approach to identifying preictal states prior to seizures. This could pave the way for development of clinical strategies in seizure prevention and even a noninvasive alarm system.

PREDICTING FINANCIAL DISTRESS FOR ROMANIAN COMPANIES

Using a moderately large number of financial ratios, we tried to build models for classifying the companies listed on the Bucharest Stock Exchange into low and high risk classes of financial distress. We considered four classification techniques: Support Vector Machines, Decision Trees, Bayesian logistic regression and Fisher linear classifier, out of which the first two proved to have the highest prediction accuracy. Classifiers were trained and tested on randomly drown samples and on four different databases built starting from the initial financial indicators. As the literature related to the topic on Romanian data is very scarce, our study, by using a variety of methods and combining feature selection and principal components analysis, brings a new approach to predicting financial distress for Romanian companies.

Single‐signal entity approach for sung word recognition with artificial neural network and time–frequency audio features

Singing voice recognition is very different from speech recognition or automatic speech recognition because there are distinct differences between speaking and singing voices. The problem is complex because music audio signals with their background instrumental accompaniments are regarded as noise sources that degrade the performance of the recognition system. This study proposes a statistical learning method to recognise words in a vocal audio signal with background music and to classify the region of a singing voice in a polyphonic audio signal. The goal of this study is to solve the problem of recognising words from sung input without using any method to separate instrumental from the background. This study also applies a concept from image recognition by using a spectrogram feature as an image to solve the problem. An audio signal with accompanying music was analysed and transformed into a spectrogram feature. To recognise it, the entire spectrogram feature was sliced, forming a feature vector for a feed-forward neural network classifier. Several classification functions were compared, including K-Nearest Neighbour, Fisher Linear Classifier, Linear Bayes Normal Classifier, Naive Bayes Classifier, Parzen Classifier and Decision Tree. The results show that using a feed-forward neural network can effectively recognise sung words at an accuracy rate of more than 93.0%. In particular, this system can recognise cross-language music data.

Combined classification error rate estimator for the Fisher linear classifier

Classification error rate estimation is one of the most important issues in machine learning and pattern recognition. This problem has been studied by many researchers and a number of error estimators have been proposed. However, theoretical analysis and empirical experiments show that most of these error estimation techniques are biased. One way to correct this biasis to use a linear combination of two different error rate estimators. In this paper we propose a new combined classification error rate estimator designed specially for the Fisher linear classifier. Experiments with real world and synthetic data sets show that resubstitution, leave-one-out, repeated10-fold cross-validation, repeated 2-foldcross-validation, basic bootstrap, 0.632 bootstrap, zero bootstrap, D-method, DS-method and M-method are outperformed by the proposed combined error rate estimator (in terms of root-mean-square error). DOI: http://dx.doi.org/10.5755/j01.itc.45.4.14268

Classification of bee pollen grains using hyperspectral microscopy imaging and Fisher linear classifier

The rapid and accurate classification of bee pollen grains is still a challenge. The purpose of this paper is to develop a method which could directly classify bee pollen grains based on fluorescence spectra. Bee pollen grain samples of six species were excited by a 409-nm laser diode source, and their fluorescence images were acquired by a hyperspectral microscopy imaging (HMI) system. One hundred pixels in the region of interest were randomly selected from each single bee pollen species. The fluorescence spectral information in all the selected pixels was stored in an n-dimensional hyperspectral data set, where n=37 for a total of 37 hyperspectral bands (465 to 645 nm). The hyperspectral data set was classified using a Fisher linear classifier. The performance of the Fisher linear classifier was measured by the leave-one-out cross-validation method, which yielded an overall accuracy of 89.2%. Finally, additional blinded samples were used to evaluate the established classification model, which demonstrated that bee pollen mixtures could be classified efficiently with the HMI system.

Novel, Objective, Multivariate Biomarkers Composed of Plasma Amino Acid Profiles for the Diagnosis and Assessment of Inflammatory Bowel Disease

BackgroundInflammatory bowel disease (IBD) is a chronic intestinal disorder that is associated with a limited number of clinical biomarkers. In order to facilitate the diagnosis of IBD and assess its disease activity, we investigated the potential of novel multivariate indexes using statistical modeling of plasma amino acid concentrations (aminogram).Methodology and Principal FindingsWe measured fasting plasma aminograms in 387 IBD patients (Crohn's disease (CD), n = 165; ulcerative colitis (UC), n = 222) and 210 healthy controls. Based on Fisher linear classifiers, multivariate indexes were developed from the aminogram in discovery samples (CD, n = 102; UC, n = 102; age and sex-matched healthy controls, n = 102) and internally validated. The indexes were used to discriminate between CD or UC patients and healthy controls, as well as between patients with active disease and those in remission. We assessed index performances using the area under the curve of the receiver operating characteristic (ROC AUC). We observed significant alterations to the plasma aminogram, including histidine and tryptophan. The multivariate indexes established from plasma aminograms were able to distinguish CD or UC patients from healthy controls with ROC AUCs of 0.940 (95% confidence interval (CI): 0.898–0.983) and 0.894 (95%CI: 0.853–0.935), respectively in validation samples (CD, n = 63; UC, n = 120; healthy controls, n = 108). In addition, other indexes appeared to be a measure of disease activity. These indexes distinguished active CD or UC patients from each remission patients with ROC AUCs of 0.894 (95%CI: 0.853–0.935) and 0.849 (95%CI: 0.770–0.928), and correlated with clinical disease activity indexes for CD (rs = 0.592, 95%CI: 0.385–0.742, p<0.001) or UC (rs = 0.598, 95%CI: 0.452–0.713, p<0.001), respectively.Conclusions and SignificanceIn this study, we demonstrated that established multivariate indexes composed of plasma amino acid profiles can serve as novel, non-invasive, objective biomarkers for the diagnosis and monitoring of IBD, providing us with new insights into the pathophysiology of the disease.

Asymboost-based fisher linear classifier for face recognition

When using AdaBoost to select discriminant features from some feature space (e.g. Gabor feature space) for face recognition, cascade structure is usually adopted to leverage the asymmetry in the distribution of positive and negative samples. Each node in the cascade structure is a classifier trained by AdaBoost with an asymmetric learning goal of high recognition rate but only moderate low false positive rate. One limitation of AdaBoost arises in the context of skewed example distribution and cascade classifiers: AdaBoost minimizes the classification error, which is not guaranteed to achieve the asymmetric node learning goal. In this paper, we propose to use the asymmetric AdaBoost (AsymBoost) as a mechanism to address the asymmetric node learning goal. Moreover, the two parts of the selecting features and forming ensemble classifiers are decoupled, both of which occur simultaneously in AsymBoost and AdaBoost. Fisher Linear Discriminant Analysis (FLDA) is used on the selected features to learn a linear discriminant function that maximizes the separability of data among the different classes, which we think can improve the recognition performance. The proposed algorithm is demonstrated with face recognition using a Gabor based representation on the FERET database. Experimental results show that the proposed algorithm yields better recognition performance than AdaBoost itself.

Segmentation and supervised classification of image objects in Epo doping-control

A software system Gel Analysis System for Epo (GASepo) has been developed within an international WADA project. As recent WADA criteria of rEpo positivity are based on identification of each relevant object (band) in Epo images, development of suitable methods of image segmentation and object classification were needed for the GASepo system. In the paper we address two particular problems: segmentation of disrupted bands and classification of the segmented objects into three or two classes. A novel band projection operator is based on convenient object merging measures and their discrimination analysis using specifically generated training set of segmented objects. A weighted ranks classification method is proposed, which is new in the field of image classification. It is based on ranks of the values of a specific criterial function. The weighted ranks classifiers proposed in our paper have been evaluated on real samples of segmented objects of Epo images and compared to three selected well-known classifiers: Fisher linear classifier, Support Vector Machine, and Multilayer Perceptron.

A three-gene predictor for early intrahepatic recurrence of hepatocellular carcinoma after curative hepatectomy

We previously developed a DNA microarray-based system that out-performs traditionally used clinical parameters for prediction of early intrahepatic recurrence (IHR) of hepatocellular carcinoma (HCC). Because DNA microarray is too expensive for daily clinical use, we used a quantitative real-time reverse transcription-polymerase chain reaction (QRT-PCR) to develop a lower-cost predictor for early IHR. From the 12 early IHR-related genes integrated in the previous predictor, we selected 6 genes whose levels showed the strongest association between data from the 2 distinct DNA microarray platforms with the same sample set. Expression of these 6 genes relative to that of GAPDH was measured by QRT-PCR in 82 HCCs. Of the 82 HCCs, 39 and 43 were assigned to training and independent test sets, respectively. By searching all combinations (n=2-6) of the 6 genes, we found an optimal combination of 3 genes (HLADRA, DDX17 and LAPTM5) that minimized the leave-one-out error for prediction of early IHR in the training set. The 3-gene predictor constructed with the Fisher linear classifier correctly predicted early IHR or non-recurrence in 35 (81.4%) of 43 HCCs in the independent test set and had a high positive predictive value of 72.7% and a high negative predictive value of 84.4%. Multivariate analysis with the stepwise logistic regression showed that the 3-gene predictor [F(x)<0] was an independent risk factor for early IHR (risk ratio, 13.6; p=0.006), indicating its potential as an easy-to-use predictor for accurate prediction of early IHR of HCC.

Improvement of the classification accuracy in discriminating diabetic retinopathy by multifocal electroretinogram analysis

Abstract The multifocal electroretinogram (mfERG) is a newly developed electrophysiological technique. In this paper, a classification method is proposed for early diagnosis of the diabetic retinopathy using mfERG data. MfERG records were obtained from eyes of healthy individuals and patients with diabetes at different stages. For each mfERG record, 103 local responses were extracted. Amplitude value of each point on all the mfERG local responses was looked as one potential feature to classify the experimental subjects. Feature subsets were selected from the feature space by comparing the inter-intra distance. Based on the selected feature subset, Fisher's linear classifiers were trained. And the final classification decision of the record was made by voting all the classifiers' outputs. Applying the method to classify all experimental subjects, very low error rates were achieved. Some crucial properties of the diabetic retinopathy classification method are also discussed. *Supported by National Natural S...

Using Multifocal ERG Responses to Discriminate Diabetic Retinopathy

Increase the accuracy in classifying subjects with or without early diabetic retinopathy, by analyzing the multifocal electroretinogram (mfERG) responses. The mfERG was recorded for 14 control subjects (Normal) and 26 non-insulin dependent diabetic patients, 16 of the patients had no apparent retinopathy (NDR), the other 10 had mild to moderate non-proliferative retinopathy (NPDR). The first order kernel (K1) and the first slice of second order kernel (K21) of the mfERG were summed across the field and exported as trace arrays. The feature subset was selected from the kernel arrays by criterion of inter-intra distance and sequential forward searching strategy. Based on the selected features, Fisher's linear classifiers were trained and the classification error rates were given. With one feature selected, the error rates of the classifiers were bigger than 23% between the NDR and NPDR subjects. When six features were selected to train the classifiers, the classification error rate decreased to 6.5% between Normal and NDR, 0 between Normal and NPDR, and 9.6% between NDR and NPDR. The criterion function of inter-intra distance was calculated for each feature in the K1 and K21 arrays. According to rank of the function value, during early DR, deformations should be easier to find at the front limb of peaks N2 and P1 on the K1 trace, and at the front limb of peak N2 on the K21 trace. As a conclusion, the multifocal ERG responses have great potentials in exploring diabetic retinopathy, assisted with the pattern classification methods.

Liver fibrosis identification based on ultrasound images captured under varied imaging protocols

Diagnostic ultrasound is a useful and noninvasive method in clinical medicine. Although due to its qualitative, subjective and experience-based nature, ultrasound image interpretation can be influenced by image conditions such as scanning frequency and machine settings. In this paper, a novel method is proposed to extract the liver features using the joint features of fractal dimension and the entropies of texture edge co-occurrence matrix based on ultrasound images, which is not sensitive to changes in emission frequency and gain. Then, Fisher linear classifier and support vector machine are employed to test a group of 99 in-vivo liver fibrosis images from 18 patients, as well as other 273 liver images from 18 normal human volunteers.

Fusion of classifiers for protein fold recognition

Predicting the three-dimensional structure of a protein from its amino acid sequence is an important problem in bioinformatics and a challenging task for machine learning algorithms. Given (numerical) features, one of the existing machine learning techniques can be then applied to learn and classify proteins represented by these features. We show that combining Fisher's linear classifier and K-Local Hyperplane Distance Nearest Neighbor we obtain an error rate lower than previously published in the literature.

Oligonucleotide microarray for prediction of early intrahepatic recurrence of hepatocellular carcinoma after curative resection

Hepatocellular carcinoma has a poor prognosis because of the high intrahepatic recurrence rate. There are technological limitations to traditional methods such as TNM staging for accurate prediction of recurrence, suggesting that new techniques are needed. We investigated mRNA expression profiles in tissue specimens from a training set, comprising 33 patients with hepatocellular carcinoma, with high-density oligonucleotide microarrays representing about 6000 genes. We used this training set in a supervised learning manner to construct a predictive system, consisting of 12 genes, with the Fisher linear classifier. We then compared the predictive performance of our system with that of a predictive system with a support vector machine (SVM-based system) on a blinded set of samples from 27 newly enrolled patients. Early intrahepatic recurrence within 1 year after curative surgery occurred in 12 (36%) and eight (30%) patients in the training and blinded sets, respectively. Our system correctly predicted early intrahepatic recurrence or non-recurrence in 25 (93%) of 27 samples in the blinded set and had a positive predictive value of 88% and a negative predictive value of 95%. By contrast, the SVM-based system predicted early intrahepatic recurrence or non-recurrence correctly in only 16 (60%) individuals in the blinded set, and the result yielded a positive predictive value of only 38% and a negative predictive value of 79%. Our system predicted early intrahepatic recurrence or non-recurrence for patients with hepatocellular carcinoma much more accurately than the SVM-based system, suggesting that our system could serve as a new method for characterising the metastatic potential of hepatocellular carcinoma.

Instrumental measurement of skin colour and skin type as risk factors for melanoma: a statistical classification procedure.

A statistical procedure to evaluate melanoma risk in Caucasian subjects on the basis of colorimetric measurement of skin colour and Fitzpatrick phototype is described. One hundred and sixty melanoma patients and 546 randomized healthy subjects of similar age, sex and place of origin were examined in the same period for skin colour using a tristimulus colorimeter and for Fitzpatrick phototype. A clinical score for classification purposes was obtained by statistical discriminant analysis with multivariate data transformation and dimension reduction techniques. A Fisher linear classifier was chosen for its simplicity and robustness in correctly predicting melanoma risk in new subjects. The classification rule was designed to avoid classifying subjects at high risk for melanoma as low risk, i.e. to give a negligible number of false negatives at the expense of more false positives. The procedure is objective and readily adapted to different clinical requirements. This is only a preliminary study but it is hoped that by performing more complex statistical analyses, e.g. neural networks, and adding other parameters (proven risk factors such as number of naevi) the performance will be further improved.

Expected classification error of the Fisher linear classifier with pseudo-inverse covariance matrix

The pseudo-Fisher linear classifier is considered as the “diagonal” Fisher linear classifier applied to the principal components corresponding to non-zero eigenvalues of the sample covariance matrix. An asymptotic formula for the expected (generalization) error of the Fisher classifier with the pseudo-inversion is derived which explains the peaking behaviour: with an increasing number of learning observations from one up to the number of features, the generalization error first decreases, and then starts to increase.

Variable selection for the classification of postoperative cardiac patients

A set of 13 extensively used hemodynamic, ventilatory and gas analysis variables are measured (on-line or off-line) on 200 patients in an intensive care unit (ICU) during the 6 h immediately following cardiac surgery. Since the existence of two well-separated classes of low- and high-risk patients has been previously shown the divergence criterion is then used to identify those variables which, at three equidistant observation times, possess the greatest separation power. Such variables always include the cardiac index (CI). representative of cardiac performance, and two indices related to respiratory efficiency and metabolic rate, i.e. the carbon dioxide production index (VCO 2I) and the arterio-venous oxygen difference (avO 2D). The Fisher linear classifier, utilizing these three features, is then tested by using the rotation method. The results show good performance of the linear classifier, which exhibits a probability of correct recognition always greater than 87%, thus suggesting the possibility of obtaining interesting improvements by means of more sophisticated classifiers.

Generalized chord transformation for distortion-invariant optical pattern recognition

An optical processor that realizes a generalized chord transformation is described. The wedge-ring detector samples of an autocorrelation are shown to be the histograms of the chord distributions. This dimensionality reduced set of features is used as the feature vector inputs for a Fisher linear classifier to determine the class of the input object independent of geometrical distortions. Initial discussions on the use of different classifiers, the polarity of the classifier’s output, and selection of the image training set are also advanced.

Automatic Classification of Spirometric Data

Pattern recognition principles have been applied to 200 sets of spirometric data obtained from pulmonary function laboratory patients. Each patient was classified by a pulmonary specialist as normal, restricted, or mildly, moderately, severely, or very severely obstructed. Each patient was represented by a five-element pattern vector consisting of forced vital capacity (FVC), forced expiratory volume in one second (FEV1), midmaximum flow rate (MMFR), and flow rates with 50 and 25 percent of the vital capacity remaining (V?50 and V?25) normalized by predicted values. By Karhunen-Loeve expansion techniques, this vector was reduced to a two-feature pattern vector with only a 6 percent residual mean square representation error. The more important feature essentially represented the average of the three flow rates, while the second feature depended on FVC and FEV1. Data were divided into training and testing sets, and using the former, a parametric Bayes classifier and one-and two-layer pair-wise Fisher linear classifiers, were designed to assign patterns described by the two derived features to one of the six categories. With the testing set, overall recognition rates were 81 to 82 percent, with most errors representing misclassifications within the four obstructive categories. If the four obstructive classes were considered as a single class, the recognition rate increased to about 94 percent.