Comparison of Feature Selection Techniques in Machine Learning for Anatomical Brain MRI in Dementia

We present a comparative split-half resampling analysis of various data driven feature selection and classification methods for the whole brain voxel-based classification analysis of anatomical magnetic resonance images. We compared support vector machines (SVMs), with or without filter based feature selection, several embedded feature selection methods and stability selection. While comparisons of the accuracy of various classification methods have been reported previously, the variability of the out-of-training sample classification accuracy and the set of selected features due to independent training and test sets have not been previously addressed in a brain imaging context. We studied two classification problems: 1) Alzheimer’s disease (AD) vs. normal control (NC) and 2) mild cognitive impairment (MCI) vs. NC classification. In AD vs. NC classification, the variability in the test accuracy due to the subject sample did not vary between different methods and exceeded the variability due to different classifiers. In MCI vs. NC classification, particularly with a large training set, embedded feature selection methods outperformed SVM-based ones with the difference in the test accuracy exceeding the test accuracy variability due to the subject sample. The filter and embedded methods produced divergent feature patterns for MCI vs. NC classification that suggests the utility of the embedded feature selection for this problem when linked with the good generalization performance. The stability of the feature sets was strongly correlated with the number of features selected, weakly correlated with the stability of classification accuracy, and uncorrelated with the average classification accuracy.     

A Twin Multi-Class Classification Support Vector Machine

Twin support vector machine (TSVM) is a novel machine learning algorithm, which aims at finding two nonparallel planes for each class. In order to do so, one needs to resolve a pair of smaller-sized quadratic programming problems rather than a single large one. Classical TSVM is proposed for the binary classification problem. However, multi-class classification problem is often met in our real world. For this problem, a new multi-class classification algorithm, called Twin-KSVC, is proposed in this paper. It takes the advantages of both TSVM and K-SVCR (support vector classification-regression machine for k-class classification) and evaluates all the training points into a “1-versus-1-versus-rest” structure, so it generates ternary outputs { ?1, 0, +1}. As all the samples are utilized in constructing the classification hyper-plane, our proposed algorithm yields higher classification accuracy in comparison with other two algorithms. Experimental results on eleven benchmark datasets demonstrate the feasibility and validity of our proposed algorithm.     

Enhanced active segment selection for single-trial EEG classification

In this study, an electroencephalogram (EEG) analysis system is proposed for single-trial classification of both motor imagery (MI) and finger-lifting EEG data. Applying event-related brain potential (ERP) data acquired from the sensorimotor cortices, the system mainly consists of three procedures; enhanced active segment selection, feature extraction, and classification. In addition to the original use of continuous wavelet transform (CWT) and Student 2-sample t statistics, the two-dimensional (2D) anisotropic Gaussian filter further refines the selection of active segments. The multiresolution fractal features are then extracted from wavelet data by using proposed modified fractal dimension. Finally, the support vector machine (SVM) is used for classification. Compared to original active segment selection, with several popular features and classifier on both the MI and finger-lifting data from 2 data sets, the results indicate that the proposed method is promising in EEG classification.     

Sparse deep neural networks on imaging genetics for schizophrenia case–control classification

Deep learning methods hold strong promise for identifying biomarkers for clinical application. However, current approaches for psychiatric classification or prediction do not allow direct interpretation of original features. In the present study, we introduce a sparse deep neural network (DNN) approach to identify sparse and interpretable features for schizophrenia (SZ) case–control classification. An L ₀‐norm regularization is implemented on the input layer of the network for sparse feature selection, which can later be interpreted based on importance weights. We applied the proposed approach on a large multi‐study cohort with gray matter volume (GMV) and single nucleotide polymorphism (SNP) data for SZ classification. A total of 634 individuals served as training samples, and the classification model was evaluated for generalizability on three independent datasets of different scanning protocols (N = 394, 255, and 160, respectively). We examined the classification power of pure GMV features, as well as combined GMV and SNP features. Empirical experiments demonstrated that sparse DNN slightly outperformed independent component analysis + support vector machine (ICA + SVM) framework, and more effectively fused GMV and SNP features for SZ discrimination, with an average error rate of 28.98% on external data. The importance weights suggested that the DNN model prioritized to select frontal and superior temporal gyrus for SZ classification with high sparsity, with parietal regions further included with lower sparsity, echoing previous literature. The results validate the application of the proposed approach to SZ classification, and promise extended utility on other data modalities and traits which ultimately may result in clinically useful tools.     

Automatic temporal lobe atrophy assessment in prodromal AD: Data from the DESCRIPA study

BackgroundIn the framework of the clinical validation of research tools, this investigation presents a validation study of an automatic medial temporal lobe atrophy measure that is applied to a naturalistic population sampled from memory clinic patients across Europe.MethodsThe procedure was developed on 1.5-T magnetic resonance images from the Alzheimer's Disease Neuroimaging Initiative database, and it was validated on an independent data set coming from the DESCRIPA study. All images underwent an automatic processing procedure to assess tissue atrophy that was targeted at the hippocampal region. For each subject, the procedure returns a classification index. Once provided with the clinical assessment at baseline and follow-up, subjects were grouped into cohorts to assess classification performance. Each cohort was divided into converters (co) and nonconverters (nc) depending on the clinical outcome at follow-up visit.ResultsWe found the area under the receiver operating characteristic curve (AUC) was 0.81 for all co versus nc subjects, and AUC was 0.90 for subjective memory complaint (SMC_nc) versus all co subjects. Furthermore, when training on mild cognitive impairment (MCI-nc/MCI-co), the classification performance generally exceeds that found when training on controls versus Alzheimer's disease (CTRL/AD).ConclusionsAutomatic magnetic resonance imaging analysis may assist clinical classification of subjects in a memory clinic setting even when images are not specifically acquired for automatic analysis.     

Genetic algorithm pruning of probabilistic neural networks in medical disease estimation

A hybrid model consisting of an Artificial Neural Network (ANN) and a Genetic Algorithm procedure for diagnostic risk factors selection in Medicine is proposed in this paper. A medical disease prediction may be viewed as a pattern classification problem based on a set of clinical and laboratory parameters. Probabilistic Neural Network models were assessed in terms of their classification accuracy concerning medical disease prediction. A Genetic Algorithm search was performed to examine potential redundancy in the diagnostic factors. This search led to a pruned ANN architecture, minimizing the number of diagnostic factors used during the training phase and therefore minimizing the number of nodes in the ANN input and hidden layer as well as the Mean Square Error of the trained ANN at the testing phase. As a conclusion, a number of diagnostic factors in a patient’s data record can be omitted without loss of fidelity in the diagnosis procedure.     

Smart sampling and incremental function learning for very large high dimensional data

Very large high dimensional data are common nowadays and they impose new challenges to data-driven and data-intensive algorithms. Computational Intelligence techniques have the potential to provide powerful tools for addressing these challenges, but the current literature focuses mainly on handling scalability issues related to data volume in terms of sample size for classification tasks.This work presents a systematic and comprehensive approach for optimally handling regression tasks with very large high dimensional data. The proposed approach is based on smart sampling techniques for minimizing the number of samples to be generated by using an iterative approach that creates new sample sets until the input and output space of the function to be approximated are optimally covered. Incremental function learning takes place in each sampling iteration, the new samples are used to fine tune the regression results of the function learning algorithm. The accuracy and confidence levels of the resulting approximation function are assessed using the probably approximately correct computation framework.The smart sampling and incremental function learning techniques can be easily used in practical applications and scale well in the case of extremely large data. The feasibility and good results of the proposed techniques are demonstrated using benchmark functions as well as functions from real-world problems.     

A Meta-Cognitive Learning Algorithm for an Extreme Learning Machine Classifier

This paper presents an efficient fast learning classifier based on the Nelson and Narens model of human meta-cognition, namely ‘Meta-cognitive Extreme Learning Machine (McELM).’ McELM has two components: a cognitive component and a meta-cognitive component. The cognitive component of McELM is a three-layered extreme learning machine (ELM) classifier. The neurons in the hidden layer of the cognitive component employ the q-Gaussian activation function, while the neurons in the input and output layers are linear. The meta-cognitive component of McELM has a self-regulatory learning mechanism that decides what-to-learn, when-to-learn, and how-to-learn in a meta-cognitive framework. As the samples in the training set are presented one-by-one, the meta-cognitive component receives the monitory signals from the cognitive component and chooses suitable learning strategies for the sample. Thus, it either deletes the sample, uses the sample to add a new neuron, or updates the output weights based on the sample, or reserves the sample for future use. Therefore, unlike the conventional ELM, the architecture of McELM is not fixed a priori, instead, the network is built during the training process. While adding a neuron, McELM chooses the centers based on the sample, and the width of the Gaussian function is chosen randomly. The output weights are estimated using the least square estimate based on the hinge-loss error function. The hinge-loss error function facilitates prediction of posterior probabilities better than the mean-square error and is hence preferred to develop the McELM classifier. While updating the network parameters, the output weights are updated using a recursive least square estimate. The performance of McELM is evaluated on a set of benchmark classification problems from the UCI machine learning repository. Performance study results highlight that meta-cognition in ELM framework enhances the decision-making ability of ELM significantly.     

Segmentation of magnetic resonance brain images through discriminant analysis

Segmentation (tissue classification) of medical images obtained from a magnetic resonance (MR) system is a primary step in most applications of medical image post-processing. This paper describes nonparametric discriminant analysis methods to segment multispectral MR images of the brain. Starting from routinely available spin-lattice relaxation time, spin-spin relaxation time, and proton density weighted images (T1w, T2w, PDw), the proposed family of statistical methods is based on: (i) a transform of the images into components that are statistically independent from each other; (ii) a nonparametric estimate of probability density functions of each tissue starting from a training set; (iii) a classic Bayes 0-1 classification rule. Experiments based on a computer built brain phantom (brainweb) and on eight real patient data sets are shown. A comparison with parametric discriminant analysis is also reported. The capability of nonparametric discriminant analysis in improving brain tissue classification of parametric methods is demonstrated. Finally, an assessment of the role of multispectrality in classifying brain tissues is discussed.     

Hybrid extreme rotation forest

This paper proposes the Hybrid Extreme Rotation Forest (HERF), an innovative ensemble learning algorithm for classification problems, combining classical Decision Trees with the recently proposed Extreme Learning Machines (ELM) training of Neural Networks. In the HERF algorithm, training of each individual classifier involves two steps: first computing a randomized data rotation transformation of the training data, second, training the individual classifier on the rotated data. The testing data is subjected to the same transformation as the training data, which is specific for each classifier in the ensemble. Experimental design in this paper involves (a) the comparison of factorization approaches to compute the randomized rotation matrix: the Principal Component Analysis (PCA) and the Quartimax, (b) assessing the effect of data normalization and bootstrapping training data selection, (c) all variants of single and combined ELM and decision trees, including Regularized ELM. This experimental design effectively includes other state-of-the-art ensemble approaches in the comparison, such as Voting ELM and Random Forest. We report extensive results over a collection of machine learning benchmark databases. Ranking the cross-validation results per experimental dataset and classifier tested concludes that HERF significantly improves over the other state-of-the-art ensemble classifier. Besides, we find some other results such as that the data rotation with Quartimax improves over PCA, and the relative insensitivity of the approach to regularization which may be attributable to the de facto regularization performed by the ensemble approach.     

Wavelet-based fractal features with active segment selection: application to single-trial EEG data

Feature extraction in brain-computer interface (BCI) work is one of the most important issues that significantly affect the success of brain signal classification. A new electroencephalogram (EEG) analysis system utilizing active segment selection and multiresolution fractal features is designed and tested for single-trial EEG classification. Applying event-related brain potential (ERP) data acquired from the sensorimotor cortices, the proposed system consists of three main procedures including active segment selection, feature extraction, and classification. The active segment selection is based on the continuous wavelet transform (CWT) and Student's two-sample t-statistics, and is used to obtain the optimal active time segment in the time-frequency domain. We then utilize a modified fractal dimension to extract multiresolution fractal feature vectors from the discrete wavelet transform (DWT) data for movement classification. By using a simple linear classifier, we find significant improvements in the rate of correct classification over the conventional approaches in all of our single-trial experiments for real finger movement. These results can be extended to see the good adaptability of the proposed method to imaginary movement data acquired from the public databases.     

Efficient exploration through active learning for value function approximation in reinforcement learning

Appropriately designing sampling policies is highly important for obtaining better control policies in reinforcement learning. In this paper, we first show that the least-squares policy iteration (LSPI) framework allows us to employ statistical active learning methods for linear regression. Then we propose a design method of good sampling policies for efficient exploration, which is particularly useful when the sampling cost of immediate rewards is high. The effectiveness of the proposed method, which we call active policy iteration (API), is demonstrated through simulations with a batting robot.     

Improving the replicability of neuroimaging findings by thresholding effect sizes instead of p‐values

The classical approach for testing statistical images using spatial extent inference (SEI) thresholds the statistical image based on the p‐value. This approach has an unfortunate consequence on the replicability of neuroimaging findings because the targeted brain regions are affected by the sample size—larger studies have more power to detect smaller effects. Here, we use simulations based on the preprocessed Autism Brain Imaging Data Exchange (ABIDE) to show that thresholding statistical images by effect sizes has more consistent estimates of activated regions across studies than thresholding by p‐values. Using a constant effect size threshold means that the p‐value threshold naturally scales with the sample size to ensure that the target set is similar across repetitions of the study that use different sample sizes. As a consequence of thresholding by the effect size, the type 1 and type 2 error rates go to zero as the sample size gets larger. We use a newly proposed robust effect size index that is defined for an arbitrary statistical image so that effect size thresholding can be used regardless of the test statistic or model.     

Neural maps in remote sensing image analysis.

We study the application of self-organizing maps (SOMs) for the analyses of remote sensing spectral images. Advanced airborne and satellite-based imaging spectrometers produce very high-dimensional spectral signatures that provide key information to many scientific investigations about the surface and atmosphere of Earth and other planets. These new, sophisticated data demand new and advanced approaches to cluster detection, visualization, and supervised classification. In this article we concentrate on the issue of faithful topological mapping in order to avoid false interpretations of cluster maps created by an SOM. We describe several new extensions of the standard SOM, developed in the past few years: the growing SOM, magnification control, and generalized relevance learning vector quantization, and demonstrate their effect on both low-dimensional traditional multi-spectral imagery and approximately 200-dimensional hyperspectral imagery.     

Diagnostic Specificity of Neurophysiological Endophenotypes in Schizophrenia and Bipolar Disorder

Background: The utility of an endophenotype depends on its ability to reduce complex disorders into stable, genetically linked phenotypes. P50 and P300 event-related potential (ERP) measures are endophenotype candidates for schizophrenia; however, their abnormalities are broadly observed across neuropsychiatric disorders. This study examined the diagnostic efficiency of P50 and P300 in schizophrenia as compared with healthy and bipolar disorder samples. Supplemental ERP measures and a multivariate classification approach were evaluated as methods to improve specificity. Methods: Diagnostic classification was first modeled in schizophrenia (SZ = 50) and healthy normal (HN = 50) samples using hierarchical logistic regression with predictors blocked by 4 levels of analysis: (1) P50 suppression, P300 amplitude, and P300 latency; (2) N100 amplitude; (3) evoked spectral power; and (4) P50 and P300 hemispheric asymmetry. The optimal model was cross-validated in a holdout sample (SZ = 34, HN = 31) and tested against a bipolar (BP = 50) sample. Results: P50 and P300 endophenotypes classified SZ from HN with 71% accuracy (sensitivity = .70, specificity = .72) but did not differentiate SZ from BP above chance level. N100 and spectral power measures improved classification accuracy of SZ vs HN to 79% (sensitivity = .78, specificity = .80) and SZ vs BP to 72% (sensitivity = .74, specificity = .70). Cross validation analyses supported the stability of these models. Conclusions: Although traditional P50 and P300 measures failed to differentiate schizophrenia from bipolar participants, N100 and evoked spectral power measures added unique variance to classification models and improved accuracy to nearly the same level achieved in comparison of schizophrenia to healthy individuals.Key words: event-related potential, P50, P300, N100, gamma frequency     

Advanced imaging of multiple mRNAs in brain tissue using a custom hyperspectral imager and multivariate curve resolution

Simultaneous imaging of multiple cellular components is of tremendous importance in the study of complex biological systems, but the inability to use probes with similar emission spectra and the time consuming nature of collecting images on a confocal microscope are prohibitive. Hyperspectral imaging technology, originally developed for remote sensing applications, has been adapted to measure multiple genes in complex biological tissues. A spectral imaging microscope was used to acquire overlapping fluorescence emissions from specific mRNAs in brain tissue by scanning the samples using a single fluorescence excitation wavelength. The underlying component spectra obtained from the samples are then separated into their respective spectral signatures using multivariate analyses, enabling the simultaneous quantitative measurement of multiple genes either at regional or cellular levels.     

Discriminative multi-task feature selection for multi-modality classification of Alzheimer’s disease

Recently, multi-task based feature selection methods have been used in multi-modality based classification of Alzheimer’s disease (AD) and its prodromal stage, i.e., mild cognitive impairment (MCI). However, in traditional multi-task feature selection methods, some useful discriminative information among subjects is usually not well mined for further improving the subsequent classification performance. Accordingly, in this paper, we propose a discriminative multi-task feature selection method to select the most discriminative features for multi-modality based classification of AD/MCI. Specifically, for each modality, we train a linear regression model using the corresponding modality of data, and further enforce the group-sparsity regularization on weights of those regression models for joint selection of common features across multiple modalities. Furthermore, we propose a discriminative regularization term based on the intra-class and inter-class Laplacian matrices to better use the discriminative information among subjects. To evaluate our proposed method, we perform extensive experiments on 202 subjects, including 51 AD patients, 99 MCI patients, and 52 healthy controls (HC), from the baseline MRI and FDG-PET image data of the Alzheimer’s Disease Neuroimaging Initiative (ADNI). The experimental results show that our proposed method not only improves the classification performance, but also has potential to discover the disease-related biomarkers useful for diagnosis of disease, along with the comparison to several state-of-the-art methods for multi-modality based AD/MCI classification.     

Extreme Learning Machine for Huge Hypotheses Re-ranking in Statistical Machine Translation

In statistical machine translation (SMT), a possibly infinite number of translation hypotheses can be decoded from a source sentence, among which re-ranking is applied to sort out the best translation result. Undoubtedly, re-ranking is an essential component of SMT for effective and efficient translation. A novel re-ranking method called Scaled Sorted Classification Re-ranking (SSCR) based on extreme learning machine (ELM) classification and minimum error rate training (MERT) is proposed. SSCR contains four steps: (1) the input features are normalized to the range of 0 to 1; (2) an ELM classification model is constructed for hypothesis ranking; (3) each translation hypothesis is ranked using the ELM classification model; and (4) the highest ranked subset of hypotheses are selected, in which the hypothesis with best predicted score based on MERT (system score) is returned as the final translation result. Compared with the baseline score (lower bound), SSCR with ELM classification can raise the translation quality up to 6.7% in IWSLT 2014 Chinese to English corpus. Compared with the state-of-the-art rank boosting, SSCR has a relatively 7.8% of improvement on BLEU in a larger WMT 2015 English-to-French corpus. Moreover, the training time of the proposed method is about 160 times faster than traditional regression-based re-ranking.